Measure your own page on the same standard We measure every figure on this page to one fixed standard. Put that standard on your own address now. Check one page free →
Where every registered claim stands right now
Registered claims
The wording below is the registered wording, unedited. Registry terms stay in Korean, the language of the source ledger.
-
We judge the forecast four weeks in: the same questions at the same repeat count, measured again, and we call it a change only once the Wilson 95% intervals stop overlapping. The floor is a return of 2.0, read over two consecutive weeks.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
On searches where an AI Overview appears, the page at position 1 loses close to half its clicks — down 58%, worse than the 34.5% drop measured a year earlier (Ahrefs, 300,000 keywords).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
62% of AI citations come from outside Google's top 10 (Ahrefs, 863,000 keywords and 4 million citations; 38% sit inside the top 10, against 76% a year ago).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Brands that reach Google's first page get mentioned more often in AI answers too. The study pulled about 600,000 PAA questions from more than 300,000 keywords, narrowed them to 10,000 questions, and found a correlation of roughly 0.65 (finance and SaaS only).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A public correlation study covering 75,000 brands (backlinks 0.218 against mentions 0.664); the body text puts referring-domain correlation in the 0.2 to 0.3 range.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Worldwide, 1 adult in 10 talks to an AI. A published analysis of 1.5 million real conversations found that more than 70% of them happen outside work; the figure 2 years earlier was 53%. Of the whole set, 3 categories cover 80%, and information seeking makes up roughly 1 in 4.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
June 2026 StatCounter share puts Google at 86.67% and Bing at 8.73%; Yahoo Search at 2.55% and DuckDuckGo at 1.53%
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
June 2026 StatCounter share for Japan puts Google at 59.04%, Bing at 33.10% and Yahoo at 6.23% (DuckDuckGo 0.72%).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
Panorama — 50 questions · 7 repeat asks · 23 engines (20 surfaces observed). We run the whole measurement grid and publish every status; that comes to 350 observations per surface.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Once a read starts, the same question goes out 7 times (k7) — to catch how the answer to each query moves.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
A Wilson 95% confidence interval goes on the share of runs we can score. Citation presence stays undetermined until the lower bound clears 0%; once it does, presence is confirmed. An interval no wider than ±10pp counts as precise enough to compare. (Worked example: cited in 5 of 7 valid runs.)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
We report it as "Cited in 33% of runs (95% CI 12–65%, n=9)" — a formatting example for the report. The 33%/17%/0% in the hero mockup carries a 'format example' label alongside it.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Growth Operations runs one Panorama a month · Category Leadership sends 10 weekly signals · Competitive Signal goes out once a month, 3 in total (raised in v5.1 on 2026-07-18) · the quarterly program runs on a 3-month cycle · the free snapshot comes once a month · the demo runs once on each of the 6 AI answer surfaces.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The search-and-AI Refit Sprint runs 4 weeks · the Category Sprint runs 8 weeks, spread over 2 execution waves.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Once a month, three times a quarter: Compass 14-grade signal measurement. Month one fixes the baseline. From month two the report carries the change against the month before.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Growth Operations — two priority moves a month, six a quarter. Category Leadership — four priority moves a month, twelve a quarter. Authority-source campaign pack — one a month, three a quarter. One move is scoped to one asset: a page, a profile or an evidence asset.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Pay for a Radar, Compass 14 or Panorama audit, then start an implementation program — Refit Sprint, Category Sprint, Growth Operations or Category Leadership — within 30 days, and the full amount you paid for that audit comes off. It applies to the single most recent audit, and Competitive Signal sits outside it. One audit never credits against another audit.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Free-snapshot raw data is held for 90 days and then deleted. The sprint deliverable carries the priorities for the next 90 days.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Free-snapshot measurement raw data is held for 90 days and then deleted. Paid-audit raw data, together with the delivery and transaction records, is held separately for up to five years — for contract verification, disputes and required records. This follows the terms of service and privacy policy that took effect on 2026-07-14.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Free snapshot — a preliminary read that puts one to three questions to six places where AI answers show up, one run each. You can request it once a month. Asking again and confidence intervals are the range the paid audit widens into.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Free snapshot — results usually inside ten minutes, and that is a measured figure. The six places where AI answers show up run in parallel, so the time follows whichever one answers slowest.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Report any defect within 10 business days of delivery. We fix it at no charge, and if it cannot be fixed we refund the amount allocated to that deliverable.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
When the fault is on our side — the service was not provided properly, or the deliverable is defective — you can withdraw your order and get a refund within 7 days of the date it was provided.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
A report share link stays valid for 7 days after it is issued (a signed link that expires).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Compass runs the full grid: 20 queries, 7 times each, across 20 engines. Panorama runs 50 queries 7 times each across 20 engines. Observations per surface come to 350.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Wilson 95% confidence intervals on repetition ratios wherever they apply. Reporting format example: citation rate 33% (95% CI 12–65%, n=9).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Refit Sprint 4 weeks · Category Sprint 8 weeks (2 execution waves)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Refit Sprint · 4 weeks · Category Sprint · 8 weeks
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
US list prices, billed in USD: Radar $399, Compass14 $1,199, Panorama $2,599, Wide Pack $599, Focus 28 $1,799, Focus 40 $2,499, Search·AI Foundation Refit Sprint $15,900, Category Sprint $29,900, Competitive Signal Program $3,490/mo, Growth Operations $8,900/mo, Category Leadership $15,900/mo. These are fixed list prices set for the US market. The figure you see is the figure you pay, and it does not move with the exchange rate.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Panorama — 50 questions, each asked 7 times, across 21 engines. The grid runs in full and every reading carries its status · 350 observations per surface
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Applicable citation rates are reported with Wilson score 95% confidence intervals and the effective sample size
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
An NBER working paper analyzing 1.5 million real ChatGPT conversations found that more than 70% of usage now happens outside of work, up from 53% two years earlier.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
HubSpot Japan State of Sales 2026 (designed by HubSpot Japan, fielded by Macromill on 2025-10-30 to 31, buyers n=515): among the sources consulted while considering a purchase, internet search fell from 57.7% in 2024 to 51.3% in 2025, while generative AI rose from 4.3% to 11.7%. A separate question puts generative-AI usage experience at 36.7% (33.0% within the past year); the denominator differs, so that figure cannot be added to the ones above. Among those users, 52.4% say it led them to add a candidate and 55.3% say it shaped the final decision. Expectation of value that only a human salesperson brings stands at 81.2%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
IT Communications × B2B Marketing purchase behavior survey 2025 (fielded 2025-05-27 to 28, n=517 employees involved in evaluating or deciding on adoption): at the evaluation stage, 49.3% turn to various web media and 35.4% to the vendor's own website. On how many vendors get compared, 16.8% compare one, 31.9% compare two and 32.7% compare three — 81.4% compare three or fewer, up from 68.1% in the 2022 survey. Purchases under ¥3 million run mostly through the web, while those of ¥10 million and above tend to combine several channels such as trade shows and seminars.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Reiwa 7 Communications Usage Trend Survey (Japan's Ministry of Internal Affairs and Communications; individual survey of people who used the internet within the past year): 74.8% used a search service nationwide (n=35,794), against 81.3% in Tokyo (n=724). No significance test was run, and the article states that alongside the figures.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
"What Gets Cited" controlled study: 252,000 trials across 6 LLMs, two injected candidate documents at a time, 18 systematically varied factors with brand anonymization and order counterbalancing (arXiv:2605.25517).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Citation Selection to Citation Absorption observational analysis: 602 prompts, 21,143 citations, 18,151 cited pages; citation frequency and actual answer influence are decoupled (arXiv:2604.25707).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Evaluating Verifiability in Generative Search Engines (Findings of EMNLP 2023, human evaluation of four commercial systems): on average 51.5% of generated sentences fully supported by their citations, 74.5% of citations support their attached sentence — 2023 snapshot of earlier products, stated as such in the article.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Google officially announced that AI Overviews would expand on 2024-08-16 into regions including Japan, in the local language.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Subscription dashboards stay wide and shallow at $29–$800/month — published prices for global AI visibility subscription dashboards: free to $29 at entry, $189–$599 for growth tiers, $699–$800 for premium (monthly subscription; frozen sample of 30 public pricing pages, verified 2026-07). The surface wording $29–$800/month is that ladder in short form.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
One publicly documented program runs $15,000 a month for six months — that is the only public case for a global deep AI search strategy program, and the figure is stated as that single case only.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
Engine calls in the panel fix explicit location parameters — city, state, country and timezone — and the customer's region selection is sealed with a sha256 receipt, verified fail-closed against the stored hash; a regional score publishes only from panels that pass the location-receipt gate.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Rebuild one listed Korean company's filings on a TTM basis (the trailing four quarters) and advertising expense as a share of revenue falls from 29.16% in 4Q20 to 17.01% in 3Q25, then rebounds for two straight quarters to 18.12% in 1Q26 (TTM peak 30.26%, in 2Q22). Across the same 2020–2025 stretch, advertising spend in absolute terms grew 4.16x, from KRW 64.1 billion to KRW 266.7 billion, while revenue grew 6.94x — the ratio came down not because the company spent less on advertising but because the denominator grew faster. Even through the 13 quarters in which the ratio was falling (4Q22→3Q25), absolute advertising spend rose 1.80x. Index TTM advertising spend itself to 4Q19 = 1 and 1Q26 stands at 8.0x. Among 9 peers, this is the only one whose ratio came down; the companies that raised advertising by more than 3.4x (3.46x and 3.43x) saw their ratio go up instead. What set this one apart was not the size of the advertising increase but a revenue multiple of 3.84x.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
Rebuilding one listed Korean company's filings quarter by quarter, of the 29 quarters only 1Q22 — the quarter where advertising expense as a share of revenue peaked at 35.08% — posted an operating loss (KRW -1,025 million).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
The publicly documented case for a one-off deep audit in Korea is KRW 8.9 million — measured on 2026-07-07 from the public site of a vendor claiming 5 engines including Naver, and stated as that anonymous public case only.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
NAVER AI Briefing t0 measurement (2026-07-16) — a 50-query panel (32 commercial / 13 informational / 5 branded), 7 repetitions per query, 350 runs in total, 343 valid observations (the 7 provider errors were concentrated on a single query and are counted as not measured). AI Briefing appeared in all 343 valid observations — a display rate of 100%, 343/343.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
NAVER AI Briefing t0 cited-domain distribution (against the 343 valid observations) — blog.naver.com 151 times, 44.0% (the widest share within this round); i-boss.co.kr 78 times, 22.7%; ranketai.com 42 times, 12.2%; developers.google.com 38 times, 11.1%; openads.co.kr 35 times, 10.2%; namu.wiki 28 times, 8.2%; ai.idearabbit.co.kr 25 times, 7.3%; aeocheck.co.kr 21 times, 6.1%; mepix.ai 21 times, 6.1%; zestcompany.co.kr 19 times, 5.5%. There were 91 unique domains in all, with the long-tail share sitting at around 5%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
NAVER AI Briefing t0 brand mentions (9-company measurement panel, against the 343 valid observations) — mentions in the answer body occurred for 3 of the 9 companies: Zest Company 26 times, 7.6% (4 queries); GeoRank 14 times, 4.1% (2 queries); ARTIFEX, our own company, 7 times, 2.0% (1 query). The remaining 6 companies had 0 mentions, 0%. Our own CiteAngle brand drew 0 citations and 1 mention. These are measured facts reported as they stand, with no disparaging characterization.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
No ad elements were detected in any of the 343 valid NAVER AI Briefing t0 observations, by our own detection criteria — the per-run verdict fields in the ledger read ad_slots_detected=0, sponsored_citations=0 and advertisers=[] across 343/343 runs, with status=ok on every one.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Google announced at I/O 2026 that AI Mode reached 1 billion monthly users within its first year, with query volume more than doubling each quarter — platform self-reported figure, attributed to Google's own announcement (2026-05-19).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Bing isn't one surface. Its index supplies Yahoo Search results under a long-standing partnership, underpins Microsoft Copilot, and — by multiple industry analyses — still influences which sources ChatGPT search surfaces. Our US panel measures Bing and Bing Copilot directly.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Japan's Ministry of Internal Affairs and Communications, White Paper on Information and Communications (Reiwa 7 edition, July 2025) — 26.7% of individuals in Japan have used generative AI, against 9.1% the year before, and 55.2% of companies use generative AI in their work, against 42.7% the year before.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
LLMO adoption on the Japanese marketer side: from 8.8% running it in earnest — measured on a base of marketers who know the term (LANY, n=100, surveyed August 2025) — to "41.6% already doing it plus 42.8% considering it, 84.4% together" (Nyle, n=420, surveyed March 2026). In that same Nyle survey the largest obstacle was "we don't know how to measure the effect," at 30.9%. The surveying body and the sample size must be stated alongside these figures.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Yahoo! Search began serving generative-AI answers on its PC version on 2025-08-05; the smartphone version has had them since October 2024 — per LY Corporation's official announcement.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Profound, a global AI visibility tool, set out its measurement methodology of 1 run per day ("once a day") on its official blog (2026-07-08), and in the same post reported that in its own experiment, repeating the measurement 10 times cut day-to-day noise in citation share by roughly 40%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
One academic analysis holds that single-run AI visibility measurement is "fundamentally unreliable" because of non-determinism, and the same paper estimates that 40 to 150 queries per platform are needed to obtain a 95% confidence interval 5 percentage points wide on citation share. The unit of that estimate is the number of queries placed in one sample; repetition count is handled in the same paper as a separate axis, 9 days of daily collection. We make no claim that our repetition count is sufficient — we publish the width of the measured confidence interval in the report exactly as it came out.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
One academic measurement finds that LLM output varies by 10 to 34% from sampling alone.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A search-industry trade contribution (Search Engine Land, 2026-06-10) treats a single observation as a point estimate and recommends repeated measurement with confidence intervals reported alongside.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Published prices for AI-search work in Korea (each firm's official public pricing page, verified 2026-07-16): one-off diagnostics KRW 880,000–3,300,000; execution retainers running from the KRW 2,000,000-a-month range to over KRW 5,000,000 a month. Global subscription tools publish a $29–699/mo band, and on their public pages the global subscription tools in this grid do not list Naver AI Briefing among the surfaces they measure.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
Source link 1 · Source link 2 · Source link 3 · Source link 4
-
Naver AI Briefing measurement: within what we checked on 2026-07-16 (the roster of leading global tools plus Korean public pages), we found no full replication of our measurement spec — one Korean firm does carry a Naver AI listing. This is an observation bounded by public pages and by that day's checking range, not a verdict on the whole market.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
Across the public pages of Korean AI-search visibility services that we checked (as of 2026-07-16), we did not find published mention of measurement audit apparatus — confidence intervals, configuration hashes, sealed ledgers. This is an observation bounded by public pages and by that day's checking range, not a verdict on the whole market.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
On the Korean AI-search market in 2026 we cite 3 verification posts published on public pages (2026-07-04, 2026-05-18, 2026-06-08), by original link and date only — we do not restate them with names attached, and the judgment is left to the source. The figure of 6 reports is the community notice's own count.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Profound, a global AI-visibility tool, explains their once-a-day reading methodology on their official blog (2026-07-08); their own experiment found that ten repeated runs cut day-to-day noise in citation share by about 40%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
An academic analysis finds single-run AI-visibility measurement 'fundamentally unreliable' due to nondeterminism, and estimates that a five-percentage-point-wide 95% confidence interval on citation share needs roughly 40–150 queries per platform. The unit in that estimate is how many queries go into one sample; repetition is a separate axis the same paper handles as nine daily collections. We claim no sufficiency for a repeat count — we publish the measured confidence-interval widths as they are.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Academic measurement finds LLM outputs shift by 10–34% from sampling alone.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A search-industry trade contribution (Search Engine Land, 2026-06-10) treats a single observation as a point estimate and recommends repeated measurement with confidence intervals reported alongside.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Published list prices of subscription AI-visibility dashboards (official pricing pages, accessed 2026-07-16): Otterly $29–489/mo, Profound $99/$399/mo (Enterprise custom), Semrush AI Visibility Toolkit $99/mo per domain billed annually, Ahrefs Brand Radar $699/mo — a $29–699/mo band, $348 to $8,388 a year.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
Source link 1 · Source link 2 · Source link 3 · Source link 4
-
Pew Research Center measured US Google visits: users who encountered an AI summary clicked a traditional result link on 8% of visits versus 15% without a summary, and clicked a link inside the AI summary on just 1% of such visits.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
G2's The Answer Economy report (March 2026 survey, n=1,076): 51% of surveyed B2B software buyers now start research with an AI chatbot more often than with Google. G2 itself has a stake in AI-search visibility, which we state alongside the figure.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
One of the major overseas AI-visibility vendors documents a once-a-day measurement method on its official blog (2026-07-08). The same post reports that in the vendor's own experiment, 10 repeated measurements cut the daily noise in citation share by about 40%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
An academic analysis calls single-run AI-visibility measurement fundamentally unreliable, because the systems are nondeterministic. The same paper estimates that 40 to 150 queries per platform are needed to reach a 95% confidence interval 5 percentage points wide on citation share. That estimate counts the queries inside one sample; in the same paper, repetition sits on a separate axis, collected daily over 9 days. We do not claim our repeat count is sufficient. We print the measured confidence-interval width in the report exactly as it comes out.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
One academic measurement puts the swing in LLM output at 10 to 34% from sampling alone.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A contributed piece in the search industry trade press (Search Engine Land, 2026-06-10) treats a single observation as a point estimate, and recommends reporting repeated measurements together with confidence intervals.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Published prices for the major overseas AI-visibility tools fall in a band of $29 to $699 per month (each vendor's official pricing page, checked 2026-07-16).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
Source link 1 · Source link 2 · Source link 3 · Source link 4
-
A comparison grid of publicly stated information on AI-visibility measurement products in Korea and abroad (based on each vendor's own public pages, accessed 2026-07-16, company names shown as initials) — published price, measured surfaces, collection cadence, repeats and confidence intervals, raw data and verification artifacts, rearranged as the facts those pages state. It is not an evaluation, and it is not a ranking.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
CiteAngle Panorama — 23 engines, 20 surfaces observed, 7 repeat asks, Wilson 95% confidence-interval width printed alongside, every reading published with its status, sealed hash ledger, public /claims registry (facts about our own product, set on the same axes in the anonymized comparison article)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
A measurement product that pairs printed measured confidence-interval widths with a sealed hash ledger and a public /claims registry — within the range of public pages we checked (11 vendors in Korea and abroad, accessed 2026-07-16), we did not find a provider in Korea disclosing the same combination. This is not an assertion about the market as a whole.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Public-page comparison grid of global AI-visibility vendors (each vendor's own public pages, accessed 2026-07-16, names shown as initials) — published pricing, engines measured, cadence, repeats and confidence intervals, raw-data access and verification artifacts, restated as facts, not rankings.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
CiteAngle Panorama on the same axes (EN comparison article) — 21 engines, 7 repeated runs per prompt-engine pair, measured Wilson 95% CI width printed in the report, every reading published with its status, sealed hash ledger plus public /claims registry
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Public information on Japanese AI-visibility measurement services, laid out side by side (based on each vendor's official public pages, accessed 2026-07-16, company names shown as initials): published price, what is measured, measurement cadence, repeats and confidence intervals, raw data, verification artifacts. It rearranges what those pages state, and it is neither a ranking nor an evaluation.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
CiteAngle Panorama on the same axes (JA comparison article) — 20 engines, 7 repeats per prompt-engine pair, measured Wilson 95% confidence-interval width printed in the report, every reading published in full down to its status, sealed hash ledger plus public /claims registry, Compass 14 at ¥198,000 / Panorama at ¥428,000 (excluding tax, published yen list price), execution programs from ¥580,000 per month
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The AI answer that shows up in LINE's in-app web search is the same feature as the "AI answer" in Yahoo! Search. LY Corporation's official announcement names "web search in the LINE app" inside the scope it covers (release no. 009460). So a spec change on that answer surface lands on Yahoo! Search and LINE at the same time. We track this primary announcement and plan our Japan-market work on that basis.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
LY Corporation has officially announced "Agent i in chat", which calls AI from a LINE chat room, for release during 2026 (announced 2 July 2026, release no. 020594). Once the launch is officially confirmed, we start verifying it as a measurement target and check the citation UI and the access path.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
We found no source-citing AI answer surface in the AI services Kakao owns and operates (as surveyed on 2026-07-16). The moment a launch is confirmed, we run a pre-commercialization baseline (t0) measurement.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
In our own research measurement of one sector — 9 law firm brands read across 6 AI answer surfaces (2026-07-16, 393 valid observations) — a leading brand already held the answer seat on all 6. Per-surface leader mention rates ran 27–69%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Within this round, the same law firm was mentioned in 68.6% of one engine's answers and in only 2.9% of another's (Wilson 95% CIs separate). The same brand was cited as a source in 21.4% of one engine's answers, and turned up in just 2.9% of another engine's answers.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Of the 9 law firms, only 1 made the top band on all 6 AI answer surfaces — and 1 firm from outside the traditional top tier reached the top band on 4 surfaces, standing level with the big names.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
In the same measurement, a firm with a 44.3% mention rate still drew zero source citations on that engine: 0 out of 70 valid observations. Across the 54 brand-and-engine pairings, 28 came back with 0 citations.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Naver AI ad launch schedule — advertiser center opens Wednesday 15 July 2026; ad serving opens Tuesday 21 July 2026. The surface carrying the first AI ad is Integrated Search > AI Briefing.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Naver AI Briefing measurement-roster check (accessed 2026-07-17): across the official engine-roster documents of 8 global vendors, we could not confirm any entry for a Naver AI surface (some support Naver for traditional search only). Among 3 Korean vendors, 1 shows it on the home page, 1 lists it as an optimization target (not a list of measured surfaces), and 1 names 4 LLMs. This is an observation bounded by the public documents we read and by the access date, not a verdict on the whole market. Examples of how those documents are worded: "35+ languages, 100+ countries", "240+ countries" — these describe prompt routing and localization, not a roster of measured engines.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
Source link 1 · Source link 2 · Source link 3 · Source link 4
-
"Your brand appears in the answers for 22% of tracked prompts" — an example of hypothetical report phrasing, not a factual claim. The body uses this figure explicitly as an example of a single-run point estimate (same shape as ci-format-example).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Worked out with the same production Wilson function our reports use (95%, z=1.96): at k=1, a query cited on every run (1/1: [20.7%, 100%]) and one never cited (0/1: [0%, 79.3%]) give intervals that overlap, so statistics cannot tell them apart. At k=7 they split — the 7/7 lower bound of 64.6% clears the 0/7 upper bound of 35.4%. Hold the same observed rate fixed and the interval narrows: 79.3pp at k=1, 56.2pp at k=3, 35.4pp at k=7, 21.5pp at k=14, 8.8pp at k=40. In this always-cited boundary case, moving from k=7 to k=40 narrows it 4.0×.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Run the same aggregator over our own k=1 run of 2026-07-17 (clinic vertical — 10 queries, 6 surfaces, 9 brands) and the confidence-interval half-width widens to as much as ±24.5 percentage points: on all six surfaces the 9 brands collapse into one top band and the surface leadership counts all tie. Rank discrimination is zero. The k=7 measured run of the same design does produce rank information. In the law-firm vertical the top band separates into 2 to 6 brands on every one of the six surfaces and leadership counts spread from 0 to 6; in the clinic vertical the leadership counts spread from 4 to 6.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Raising repetitions from 7 to 40 multiplies the measurement workload — queries times surfaces times repetitions — by 5.7. The interval does not narrow by nearly that much: 4.0x on the worked calculation for the boundary case cited in every repetition, and √5.7 ≈ 2.4x under the √k asymptotic law for mid-range observed rates.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Our own query-set quality diagnostic (law-firm vertical, 50 questions, 7 repetitions each, 14 engines; measurement window 2026-07-12): in one comparison-type query that named two brands by name, the mention rate for those two named brands measured 1.00 and 0.98 across 48 judged observations. We read that as an echo of the names already sitting in the question — an artifact of how the query was built — so we classified it as leading-bias and swapped it out. After the swap, the v4 query set is machine-verified to detect zero roster brand aliases in any question. Brand names stay anonymous outside the company, and the original queries and the query set itself stay closed as sales know-how.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Overviews were triggered for 6.49% of queries in January 2025, 24.61% in July, and 15.69% in November (Semrush AI Overviews study, 10M+ keyword panel, data January through November 2025). Those are the vendor's own panel figures, cited together with their publication date, and we use them for one directional point only: the denominator that decides when an AI answer fires moves when the platform moves it.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
"3× AI mentions in 60 days" is an invented example of how a vendor deck phrases things, not a factual claim. The article carries it under the label 'A format example — not a quote from any vendor.' and uses it to open the question of what the denominator was.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Bixby on Galaxy answers web questions with Perplexity-backed search summaries and source cards (One UI 8.5), Apple has been reported to have signed a deal to put Google's Gemini inside Siri (CNBC and TechCrunch, January 2026), and the answers in a ChatGPT voice conversation are written by GPT itself
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
Source link 1 · Source link 2 · Source link 3 · Source link 4 · Source link 5
-
A meta-analysis of 54 studies published in May 2026 (Zyppy — 54 pieces selected from experiments, patents and case reports) gave llms.txt the lowest score of any AI-citation factor it rated, 2.0 out of 10, and concluded that it had found no reliable evidence supporting an impact
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A quasi-experiment tracking 1,885 treated pages against 4,000 control pages (Ahrefs, published 2026-05-11) found no rise in AI citations for the pages that added schema — AI Overviews came in at −4.6%, significant, while AI Mode and ChatGPT showed no significant difference
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
We checked the official public pages of 12 overseas AI visibility measurement tools ourselves (accessed 2026-07-28). Two of them publish llms.txt generation and management as a feature of their own. Korean-surface measurement, Naver AI Briefing included, is what CiteAngle provides
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
C-SEO Bench, a benchmark accepted at NeurIPS 2025 that evaluated 54 style and content edit conditions across two tasks and six domains. Only 3 of the 54 conditions produced a significant positive improvement in citation ranking.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Google made three official announcements: on 16 August 2024, the expansion of AI Overviews to Japan and other markets with local-language support; on 9 September 2025, the Japanese rollout of AI Mode; and on 27 March 2026, the language and region expansion of Search Live, Japan included.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
LY Corporation made three official announcements: on 20 April 2026, its conversational AI agent Agent i; on 5 June 2026, per-user personalization and memory; and on 30 June 2026, the expansion of its coverage to 22 genres, its use of the OpenAI API, and a Yahoo! JAPAN ID login requirement.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Tokyo is not only the 23 wards. It has 23 special wards, 26 cities, 5 towns and 8 villages — 62 municipalities in all.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Microsoft announced AI Performance in Bing Webmaster Tools as a public preview on February 10, 2026. A second official announcement, on June 16, 2026, added the Intents, Topics, Citation Share and Compare views.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
AI Visibility Index, August 2026 issue (No. 1), law firms (KR) — surface leadership counts, our own measurement, window 2026-07-16 to 2026-07-17 KST (Seoul). Anonymous A reached the top band on 6 of the 6 AI answer surfaces measured. The top band is the set of brands whose own Wilson 95% CI overlaps that of the surface's leading brand by mention rate. Anonymous B, Anonymous C, Anonymous D and Anonymous E each reached it on 4 surfaces — leadership counts 6 · 4 · 4 · 4 · 4, a roster of 9 brands, every brand shown anonymized on public pages. Band membership states a tie by CI overlap, and we do not produce a single composite ranking.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 1, law firms (KR) — grid and distribution, our own measurement, 2026-07-16 round. Every pairing of the 9-brand roster with the 6 AI answer surfaces was published, mention rate and citation rate each carrying a Wilson 95% CI. The denominator behind each value is that surface's valid runs in which an AI answer was displayed: 70 each for ChatGPT, Claude, Perplexity, Grok and Naver AI Briefing, and 43 for Google AI Mode, where 27 runs with no answer displayed were excluded from the denominator and that count published. Representative values: Anonymous A on ChatGPT, mention rate 47% [36–59]; Anonymous E on Grok, 69% [57–78]; Anonymous B on ChatGPT, 51% [40–63]. Median roster mention rate by surface ran from 0.0% to 47.1%. Denominators differ from surface to surface, so we neither mix nor compare across them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 1, law firms (KR) — measurement receipt, our own measurement. 420 jobs (10 sealed queries · 6 surfaces · 7 repeats, with a brand mention judged up to 70 times per surface) closed 420/420, 0 unmeasured and 0 unjudgeable, and measurement status was recorded in 3 states: unmeasured, unjudgeable, absent. Measurement window 2026-07-16 21:51 to 2026-07-17 01:16 KST (Seoul). 6 output artifacts were published on the page under sha256 seals.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index, August 2026 issue (No. 1), dermatology and plastic surgery clinics (KR) — surface leadership counts, our own measurement, window 2026-07-16 to 2026-07-17 KST (Seoul). Anonymous A, Anonymous B and Anonymous C each reached the top band on 6 of the 6 AI answer surfaces measured — 3 brands tied at that count — and Anonymous D reached it on 5 surfaces (leadership counts 6 · 6 · 6 · 5, a roster of 9 brands, every brand shown anonymized on public pages). Median mention rates run low in this vertical, so the top band overlaps widely; membership in it is not a statement of rank order. We do not produce a single composite ranking.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 1, dermatology and plastic surgery clinics (KR) — grid and distribution, our own measurement, 2026-07-16 round. Every pairing of the 9-brand roster with the 6 AI answer surfaces was published, mention rate and citation rate each carrying a Wilson 95% CI. The denominator behind each value is that surface's valid runs in which an AI answer was displayed: 70 each for ChatGPT, Claude, Perplexity, Grok and Naver AI Briefing, and 53 for Google AI Mode, where 17 runs with no answer displayed were excluded from the denominator and that count published. Representative values: Anonymous A on Grok, mention rate 21% [13–32]; Anonymous B on ChatGPT, 9% [4–17]; Anonymous C on Perplexity, 17% [10–28]. Median roster mention rate by surface ran from 0.0% to 1.4%. On a surface where the mention rate is 0%, the top band can form across the entire roster, so the page states that band membership alone is not to be cited as evidence of presence. Denominators differ from surface to surface, so we neither mix nor compare across them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 1, dermatology and plastic surgery clinics (KR) — measurement receipt, our own measurement. 420 jobs (10 sealed queries · 6 surfaces · 7 repeats, with a brand mention judged up to 70 times per surface) closed 420/420, 0 unmeasured and 0 unjudgeable, and measurement status was recorded in 3 states: unmeasured, unjudgeable, absent. Measurement window 2026-07-16 21:51 to 2026-07-17 00:55 KST (Seoul). 6 output artifacts were published on the page under sha256 seals.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
In an independent academic convergence analysis covering 4 AI search engines (Schulte et al. 2026, arXiv:2604.07585), the standard error on a brand's detection rate from 1 measurement is ±0.37. The error falls as repeats accumulate, and the number of repeats that first reaches the statistical stability line — a standard error below 0.10 — is 7. Every paid diagnostic re-measures 7 times per query on each surface. That repeat count is not an option; it is a default sealed into the measurement-condition hash.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Compass — 20 questions, each asked 7 times, across 23 engines with 20 surfaces observed. The measurement grid is run in full and every status is published.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Compass — 20 questions, each asked seven times across every surface. Nothing is sampled and every reading keeps its status
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Radar — 5 questions, each asked 7 times, the full locale roster run with every status published (KR: 23 engines, 20 surfaces · US: 21 engines, 20 surfaces · JP: 20 engines, 19 surfaces)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Wide Pack — 10 more questions added inside the same measurement window, each one asked 7 times across that locale's roster of places · Focus 28 — the core 10 questions asked 21 more times each, for 28 cumulative runs · Focus 40 — asked 33 more times each, for 40 cumulative runs · Full-Focus — all 50 questions taken to 28 or 40 cumulative runs · either way, that market's places stay fixed at full width (20 in Korea, 20 in the US, 19 in Japan)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
A single measurement is a point estimate. To narrow the 95% confidence interval on citation share to a 5-percentage-point width, one research estimate puts the requirement at 40–150 queries per platform (§5.7 — distinct questions, not repeats of one).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
Price ladder v6, confirmed prices (KRW, VAT excluded): Radar 390,000 · Compass14 1,150,000 · Panorama 2,600,000 · Wide Pack 600,000 · Focus 28 1,750,000 · Focus 40 2,450,000 · Full-Focus quoted (anchor = floor: from 9,300,000 for k28 and 12,900,000 for k40) — deduction v6: when an execution program starts within 30 days of an audit payment, the full amount of that immediately preceding audit payment is deducted automatically (deduction between one audit and another was discontinued on 2026-07-19)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The ranking of the top 10 domains cited in AI search, from Ahrefs Brand Radar's Japan data (June 2026 edition) — the source URL is written into the body text, as Ahrefs' citation terms require.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
Observed 2026-07-17 — an analysis of the domain-type breakdown published by Optyino.ai (株式会社Wallabee). The source URL and the company name run alongside it, as the citation terms require, with a note that the figures come from API responses.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
A row of published-price facts for consulting-style spot audits sold inside Japan, taken from each firm's own public page, accessed 2026-07-18, with firm names anonymized to initials — it rearranges what those pages state, and it is not a rating or a ranking.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Public information gathered for 5 more Japanese vendors of AI-visibility measurement and SEO/LLMO tools, added as rows (each vendor's own public page, accessed 2026-07-18, names shown as initials) — published fee, what is measured, how often, repetition, confidence interval, raw data and verification mechanism, as those pages state them. It rearranges those facts, and it is not a ranking or a rating.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
In June 2026 Google Search Console opened a generative AI report and began showing a site's own generative AI exposure inside Google Search surfaces (the official documentation centers it on exposure, and it reaches some sites first, then widens).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
According to Naver's own announcement, AI Briefing is a surface used by 30 million monthly unique users.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
On searches where an AI summary appeared, a link was clicked on 8% of them — about half the 15% seen on searches without one (900 adults, 68,879 searches).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Nearly half of US adults (49%) now use AI chatbots, and 60% say they read the AI summaries at the top of search results
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
2 out of 3 US consumers (63%) said they would trust AI search results less once ads appear in them.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
In eBay's large-scale field experiment, 99.5% of the traffic credited to brand-keyword ads would have arrived through organic links anyway, with the ads switched off.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Google's meta-analysis of ad-pause experiments found 89% of ad clicks to be incremental - organic search did not recover them.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Worldwide Google search: organic clicks 33.6% against paid clicks 1.6%, measured from panel data.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
The US search-advertising market grew roughly 20-fold, from $5.1 billion in 2005 to $102.9 billion in 2024 (IAB/PwC).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Search Console opened its generative-AI report in June 2026: impression-centric exposure inside Google surfaces for a site you own, rolling out progressively.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
Search Console began offering its generative-AI report in June 2026.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
19% of AI Mode citations came from the top 20 search results for the same query - seoClarity, US desktop transactional-query sample, data as of 2025-09 (R1-C6, bound to the in-body 'rank tracking tool' FAQ item in KR/EN/JA).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
In a product-recommendation study, turning search on replaced 80.2% of the recommended products; only 15.8% of products that appeared without search survived — across 20,000 responses (Visibility Labs / Search Engine Land, 2026-06).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
In a Semrush survey of 622 US B2B buyers (519 AI users), 92% said AI influences the vendor shortlist and 83% said it influences the final decision; 71% then visit the vendor site and 63% re-verify on Google. Use-case fit (53%) outweighs brand awareness (7%) in why a vendor is trusted from an AI answer (Semrush, 2026-07).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Japanese-language AI citations sit in the long tail: the top 10 domains hold only 9.5% of the total (Principle Inc. / 株式会社プリンシプル, 'What Do LLMs Cite?' / 「LLMは何を引用するのか」, as of 2026-05; underlying data from Ahrefs Brand Radar).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Nielsen Digital Content Ratings (monthly total, March 2025, via MarkeZine's coverage) puts ChatGPT first among Japan's generative-AI assistant users, at 9.31 million people.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
LY's own survey (January 2026, earnings presentation deck p.25) puts daily generative-AI use in Japan at 16.0% — a separate metric, defined differently from experience-rate surveys.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
A tidy-up of what three Japanese B2B vendors and BOXIL publish about “portal bundle” AI visibility products (listing type, review-contract bundle type, free monitor type). We took the published fees, the format and the stated facts from each company's own product pages and official releases (accessed 2026-07-18) and laid them on one axis. It is not a ranking and not a rating. Vendor self-reported figures that arrive without their measurement conditions are not quoted.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-10
-
Measurement-design facts for the case-study-format page: the full grid is 50 prompts, each asked 7 times, across 21 engines and 20 surfaces, with Wilson 95% confidence intervals. These are facts about our own product and they are published in the methodology.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Fixed JPY list prices for the Japanese market, before tax (price v7, effective 2026-07-28: the USD price multiplied by 163.8, then snapped to the price points customary in Japan): Radar ¥64,800, Compass ¥198,000, Panorama ¥428,000, Wide Pack ¥98,000, Focus 28 ¥298,000, Focus 40 ¥408,000, Refit Sprint ¥2,600,000 for four weeks, Category Sprint ¥4,900,000 for eight weeks. We bill the price as shown, with no FX indexation. Prices are shown before tax: online payment, once it is enabled, adds consumption tax (10%) on the screen before you pay, and for invoice payment or a quote the tax treatment is stated on the quote and on the invoice.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
US subscription programs bill monthly or annually (annual is priced at ten months — '2 months free'): Competitive Signal $3,490/mo · $34,900/yr, Growth Operations $8,900/mo · $89,000/yr, Category Leadership $15,900/mo · $159,000/yr. Quarterly billing is retained for contracted invoicing only and is not surfaced (price v7 canon §4, 2026-07-28; toggle split confirmed 2026-07-19).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Japanese-market subscriptions, before tax, on a two-way monthly/annual toggle with no quarterly segment: Competitive Signal ¥580,000 monthly · ¥5,800,000 annually, Growth Operations ¥1,480,000 monthly · ¥14,800,000 annually, Category Leadership ¥2,600,000 monthly · ¥26,000,000 annually. The annual price is ten months' worth — two months free. The monthly view states a recommended minimum term of one quarter (3 months): one month gives you a baseline and the first movement, a quarter gives you the trend. Quarterly prices exist only for contracts handled by invoice or quote and are never shown on a web surface (price v7 §4, effective 2026-07-28; the two-way split was settled on 07-19).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Seventeen places where an answer shows up. What we measure: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overview, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, DuckDuckGo search assist, Bing News, Yahoo! JAPAN Search and the Yahoo! JAPAN AI answer — seventeen in all. The same question goes out seven times each. One Panorama run puts 50 questions to 15 services seven times apiece, 5,250 calls in all, and reads them at the 17 places where answers appear. How far the answers swing from one day to the next also becomes a number. Seventeen places to count from, fifteen services to ask. Spec and measurement are written separately. The spec above is what runs on the run you order. The Japanese measurement we have sealed on our own brand is the run of July 17, 2026. That roster is narrower than the contract spec above, so whenever we quote a Japanese measured figure we state the measurement date and the scope of the observation alongside it.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Japan is measured to the same spec. The contract spec is 50 questions, asked 7 times each, across 20 engines — 7,000 cells — and there are 19 places where a brand name can appear. We measure ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overviews, Google News, YouTube, Google short-form, Bing Search, Bing Copilot, DuckDuckGo search assist, Bing News, Yahoo! JAPAN Search, the Yahoo! JAPAN AI answer, Meta AI model answers and X Search — 19 surfaces. Spec and measurement are written separately: the Japanese measurement we have sealed on our own brand is the run of July 17, 2026, and that roster was narrower than today's contract spec. So whenever we quote a Japanese measured figure, we write the measurement date and the scope of the observation next to it.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Japan runs on the same spec. Seventeen places across fifteen AI and search engines, with 50 questions asked 7 times each. Where we read: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overviews, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, DuckDuckGo search assist, Bing News, Yahoo! JAPAN Search and the Yahoo! JAPAN AI answer — a Japan-only roster, distinct from the US panel above. Spec and measurement stay on separate lines in Japan too. The Japanese reading we have sealed on our own brand is dated July 17, 2026, on a roster narrower than the spec above, so every Japanese figure we quote carries its measurement date and its scope.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
For customers with a store or a facility, we take on what comes after as well. We submit the facility information to Yahoo! Place and keep what shows in Yahoo! Search and Yahoo! Map current. The next single thing to fix is not picked on a hunch. The analytics tab of the management tool is specified to accumulate impression counts, the action counts for the official-site, phone-number, directions and reservation buttons, and the split between direct and indirect search — so those are the numbers we work from.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
We write the press release in Japanese from the start rather than translating one, and put it on a Japanese distribution network that carries it to the outlets. Whether it runs is each outlet's call. A release that does run stays online as Japanese source material.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
If you run stores or facilities in Japan, we take the listing work. Yahoo! JAPAN gives owners a free tool for filing facility details, and we do that filing and keep the entry current. What people see on Yahoo! JAPAN Search and Yahoo! JAPAN Maps stays accurate. The same tool counts how often the listing was shown, and it counts the taps on your website, phone number, directions and booking button. It also separates the people who searched your name from the people who found you another way. Those counts pick the next thing we fix.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Press work runs on the same footing. We write the release in Japanese from the start, not as a translation of your English one. It goes out through a Japanese distribution network to Japanese outlets. Each outlet decides what it runs, and a release that gets picked up stays online as Japanese source material that search results and AI answers can both draw on.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
For a Japanese round we fix the pass criteria and register them before the measurement, not after. When a round produces numbers, the report carries them with the sample size and the moment we took the reading.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
If you have a store or a facility in Japan, we register and maintain its details in Yahoo! Places, so what goes out to Yahoo! JAPAN Search and Yahoo! JAPAN Maps stays current. The Places console collects the number of times the listing was shown, along with the clicks on the official site, the phone number, directions and the booking button. Those counts are the grounds on which we pick the one place to work on next.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
We do not translate the press release. We write it fresh in Japanese and send it out over a Japanese distribution network. Each outlet decides whether to run it, and an article that runs stays as Japanese source material.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
We fix the criteria for a Japanese round and register them before we measure. The question set, and the line that separates a pass from a fail, both go into a document before we observe anything, so the verdict in the report was not decided after seeing the result.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
We measure 20 surfaces: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Naver AI Briefing, Google Search, Google AI Overviews, Naver Search, Naver News, Google News, Bing News, YouTube, Google short-form video, Bing Search, Bing Copilot, DuckDuckGo search assist, DuckDuckGo search, and Meta AI model answers. Every run puts 50 questions through 7 re-asks across 23 engines, in full — 8,050 cells.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
The Korean market runs on the same grid. We take the 20 main surfaces, Naver AI Briefing among them, and run them in full: 50 queries, 7 re-asks each, 23 engines. Surfaces measured — ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Naver AI Briefing, Google Search, Google AI Overview, Naver Search, Naver News, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, Bing News, DuckDuckGo search assist, DuckDuckGo search, and Meta AI model answers. Twenty in all.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Surfaces: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overviews, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, Yahoo Search, DuckDuckGo search assist, DuckDuckGo organic, Bing News, Meta AI model answers, Google Shopping, and X search. The contract spec for each paid audit is 50 questions asked 7 times each across 21 engine captures — a 7,350-cell grid — reported across 20 surfaces with a status on every reading. Spec and measurement are separate lines here, and we print both. The US reading we have sealed on our own brand is dated July 27, 2026 — 50 questions across 15 answer surfaces, one pass.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
The US runs on the same spec. The contract spec is 50 questions, each asked 7 separate times, across 21 engines — 7,350 cells — and 20 places can carry your brand name. The roster is not Korea's: Yahoo Search, DuckDuckGo search assist and Bing News take the three slots Naver holds there. Spec and measurement stay on separate lines. The US reading we have sealed on our own brand is the round of July 27, 2026, a single pass of 50 questions across 15 places. Where we read: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overviews, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, Yahoo Search, DuckDuckGo search assist, DuckDuckGo organic, Bing News, Meta AI model answers, Google Shopping and X search — 20 places.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
The US market runs on the same spec. We rebuild the roster to match how Americans search, and it takes in Yahoo Search, DuckDuckGo search assist and Bing News — 20 major places in all. Where we read: ChatGPT, Perplexity, Claude, Grok, Gemini, Google AI Mode, Google Search, Google AI Overviews, Google News, YouTube, Google short-form video, Bing Search, Bing Copilot, Yahoo Search, DuckDuckGo search assist, DuckDuckGo organic, Bing News, Meta AI model answers, Google Shopping and X search — 20 places. Spec and measurement stay on separate lines. The US reading we have sealed on our own brand is the round of July 27, 2026, a single pass of 50 questions across 15 places. Rounds that stack a question up to seven times run at the next measurement, as the spec says.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-10
-
Opensurvey, AI Search Trend Report 2026, first-half edition (fielded March 2025 and December 2025; the question was which search services the respondent had used in the past three months, multiple response; base at each wave was 990 and 989 people who had used a search service at least once) — use for search: ChatGPT 39.6% to 54.5%, Gemini 9.5% to 28.9%, Naver 85.3% to 81.6%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
In the same report, trust in ChatGPT's search results — rated by people who had used it, counted as the top 2 points on the positive side of the scale, with 345 and 614 raters at the two waves — fell from 57.6% to 49.1%. Use for search rose over the same period. Trust and everyday use moved in opposite directions, and both readings stand.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
What people did when the answer they wanted did not come back on the first try (people who had used ChatGPT, multiple response, March 2025 N=345, December 2025 N=614) — typing the question again ranked first, 74.3% to 77.2%; searching on a different generative AI, 17.7% to 30.0%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Opensurvey, AI Search Trend Report 2026, second-half edition, public release (fielded July 2026) — 56.9% of generative-AI users ask the AI the moment something comes to mind instead of holding the question, and 42.3% keep digging with follow-up questions.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
DoubleVerify, the digital ad verification firm, bills by the Measured Transaction — its 10-K says so in as many words — and revenue went from $182.66 million in 2019 to $748.29 million in 2025, a factor of 4.10 (recomputed from the audited XBRL figures in its SEC filings, checked July 22, 2026).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Some 30 checks run automatically on every deploy, and a failure blocks the deploy — our own site ships only after it clears the same gate.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Deterministic gates audit crawl, indexing, rendering and structured data — structured data goes through a deep check of the required fields for each type, with 8+ types implemented.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
We do not measure SEO surfaces apart from AI surfaces. They sit on one grid, with the same queries and the same repeat spec, and organic surfaces are part of the grid we sell.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Whether what shows up in search carries over into AI citations is measured in our own data and shown as a cross-tab. Any figure quoted from it carries its scope label.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
How we run IndexNow: URL notification is automated after each deploy. An accepted notification (HTTP 200) is not indexing. We make no claim of indexing success or of a shorter lead time, and the receipt carries that note.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Outside benchmarks put it in numbers: models follow a user's plainly stated preferences less than 10% of the time within ten turns of chat (PrefEval), and the best model in a long-running conversation simulation scores 52.8% (HorizonBench).
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Public AI chat corpora lean hard toward English — 78% of turns in LMSYS-Chat-1M and 92.35% of ShareGPT conversations. Korea does not appear on WildChat's list of top collection countries, and Japanese IPs account for 1.94%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
Korea, measured — 50 questions, each asked 7 times. Captured 2026-07-15.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Mean number of source links attached to one answer, from the same Korean measurement (50 questions, 7 repeats, 14 engines): organic search 11.22 (n = 1,050), video 11.14 (n = 700), AI search summaries 6.82 (n = 1,050), news 6.58 (n = 700), AI assistants 6.19 (n = 1,400). Captured 2026-07-15.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
The same run collected 40,016 citation links pointing at 2,013 distinct domains — youtube.com 13.9%, blog.naver.com 7.1%, instagram.com 6.1%, with the remaining 2,010 domains taking 72.9%. Three domains take 27.1% between them. Captured 2026-07-15.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Re-running the identical prompt — no rewording at all — returns recommendation sets that overlap at a Jaccard of only 0.50–0.61 in a production study of OpenAI and Anthropic models.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-07
-
Google runs ads above and below AI Overviews in more than 200 markets. Ads placed inside AI Overviews themselves were, as of the date we checked, English-only and limited to 12 countries, which do not include Korea.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
In a single-run pilot of 30 Korean law-firm queries (observed 2026-07-11), 3,243 of the 12,345 citations gathered across 7 AI answer surfaces — 26.3%, 95% CI 25.5–27.1% — pointed at a domain that also showed up on the first page of web search for the same query; the ranges quoted here are the spread observed inside this one run. Co-visibility by surface: OpenAI web search API 5.1% (59/1,155), Naver AI Briefing 10.4% (91/875), Anthropic web search API 25.0% (1,262/5,043), Google AI Overviews 27.7% (57/206), Google AI Mode 31.5% (532/1,687), xAI web search API 34.7% (492/1,416), Perplexity API 38.2% (750/1,963) — a 7.5x spread between the highest and the lowest, with the 95% CIs at the two ends not overlapping. Naver organic search in the same run came to 6.5% (203/3,108). The citations covered 1,036 unique domains, and the largest counts outside a search first page were chambers.com with 594, blog.naver.com with 517 and youtube.com with 478.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
US citation composition, own observation (measured 2026-07-27, recounted from the measurement ledger 2026-07-28): one full pass over 50 US category and brand questions across 15 answer surfaces produced 8,662 citations resolving to 1,987 distinct domains. The ten most-cited domains carry 2,353 citations (27.2%); 1,088 domains (54.8%) are cited exactly once; 1,390 of the 1,987 domains appear on only one surface and two (semrush.com, searchengineland.com) reach eleven. Top ten by citations: youtube.com 1,152 (13.3% of all citations, of which 1,040 come from the two video answer surfaces) · semrush.com 318 · linkedin.com 189 · instagram.com 154 · searchenginejournal.com 117 · reddit.com 102 · facebook.com 92 · forbes.com 80 · explodingtopics.com 75 · searchengineland.com 74. Citations by surface (sum 8,662): Google News 2,675 (30.9% of the run, drawn from 635 distinct domains) · YouTube 836 (one domain) · Perplexity 759 · Google AI Mode 744 · Google short-form video 588 · Bing Copilot 533 · Bing Search 488 · Google Search 427 · Google AI Overviews 359 · Yahoo Search 350 · Grok 289 · ChatGPT 286 · Bing News 227 · DuckDuckGo search assist 87 · Claude 14.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
Open-platform share of US AI citations, own observation (measured 2026-07-27, recounted 2026-07-28): across the 45 category questions (the five brand-name questions in the set held out), 1,631 of 8,228 citations — 19.8% — pointed at nine platforms anyone can publish to (youtube.com, linkedin.com, instagram.com, facebook.com, reddit.com, tiktok.com, x.com, quora.com, medium.com). 1,308 of those 1,631 came from the two video answer surfaces; on the other thirteen surfaces the same nine platforms account for 323 of 6,908 citations, 4.7%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-01
-
Domain composition of Google AI Overviews citations, own observation (measured 2026-07-12 · 50 questions in the Korean large-law-firm category, 7 repeats per question over 2 measurement rounds for 14 repeats, 700 observations in total) — 657 citations after folding source links down to domains, 173 distinct domains, 48 questions where a source link was captured. The 21 domains inside the top 20: youtube.com 144 (21.9% · first place · across 32 questions), blog.naver.com 31, Firm F's own domain 27, m.blog.naver.com 26, lawtalk.co.kr 21, v.daum.net 16, Firm ① from outside the measured list 16, Firm B 14, Firm C 13, lawtimes.co.kr 13, legaltimes.co.kr 12, Firm A 11, namu.wiki 9, Firm D 8, magazine.hankyung.com 7, investchosun.com 7, hankyung.com 7, Firm E's main domain 7, Firm E's practice-area domain 7, instagram.com 6, dealsite.co.kr 6 — 408 of 657 citations together, 62.1% (by character: 1 video domain 144, 8 law-firm-owned domains 103, 4 UGC and social domains 72, 7 news and trade-press domains 68, 1 legal consultation platform 21). The remaining 152 domains hold 249 citations, 37.9%. 89 domains, 13.5%, were cited exactly once, and one firm's practice-area and regional domains number 21 and total 45 citations. Control observations — the Korean AI-search-optimization category on 2026-07-15 (50 questions, 7 repeats) drew 330 citations across 178 distinct domains, of which youtube.com took 44, 13.3%; the same category in Japan on 2026-07-17 (50 questions, 7 repeats) drew 283 citations, of which youtube.com took 21, 7.4%. In all three observations the most-cited domain is youtube.com. The Naver AI Briefing t0 comparison (2026-07-16) is bound to b1-naver-t0-citation-domains and rests on a different denominator — Naver's 44.0% is measured against valid observed repeats, while the shares here are measured against the count of citation links.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Raw measurement data from a free check: 90 days. Measurement records from a contracted diagnosis: up to 5 years. Copies of the email bodies we send: 30 days. Account information: until you close your account.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The health signal turns to caution when the longest-waiting measurement job in the queue passes 30 minutes (1800 seconds), or when record collection falls behind. What we judge on is waiting time, not the number of jobs — 1 job stuck for 1 hour weighs heavier than 20 jobs flowing through in 3 minutes.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The public MCP endpoints are open without a key, up to 30 calls a minute per address. Only the inspection tool that actually fetches an external page runs on a separate allowance of 5 a minute.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Which angle we read in which month is fixed in advance on the 12-month rotation table attached to the contract, and we take one out each month. At 6 months you have seen half the table. A full turn takes 12 months. A 1-year contract is billed at 10 months of the monthly price — 12 months of service for the price of 10.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
The scope of the Refit Sprint includes up to 10 technical gates — crawl, indexing, rendering, sitemap, canonical and the like.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
We recommend 3 months or more. Month 1 sets the baseline, and month 2 is where the first change shows.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Twelve months at the monthly rate: ¥6,960,000 / ¥17,760,000 / ¥31,200,000 — shown struck through beside the annual term price (¥5,800,000 / ¥14,800,000 / ¥26,000,000)
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
On our numbers table, the total row of the share-by-status column reads 100.00% — the denominator is the full count of observations and the statuses split every one of them, so the shares add up to 100%
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Prices confirmed by the CEO on 2026-08-03, in three currencies: Wide-Angle Panorama KRW 3,900,000 / $3,899 / ¥638,000 · Wide Pack KRW 800,000 / $799 / ¥128,000 (raised from KRW 600,000 / $599 / ¥98,000) · Angle Pack KRW 800,000 / $799 / ¥128,000 (newly added)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
In the disclosures of one listed Korean company, advertising expense as a share of revenue fell from 29.16% to 17.46%, down 11.7 percentage points. Divide the same two years' absolute advertising expense (64.1 billion won and 266.7 billion won — 641 and 2,667 in hundred-million-won units) by those shares (0.2916 and 0.1746) and revenue back-solves to roughly 219.8 billion won and roughly 1.5275 trillion won, a ratio of 6.94 between the two. The index chart on the page sets 2020 to 100 and redraws revenue at 694, advertising spend at 416 and the share at 60; the version that starts the advertising line at 1.00 in the first year carries the same values. A single line of annual advertising expense in the sixth year is about 1.2 times the whole of first-year revenue. The page uses four numbers — advertising expense 641 and 2,667, shares 0.2916 and 0.1746 — and everything else is division on those four.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
In the same 29-quarter reconstruction, the highest advertising-to-revenue share on a standalone-quarter basis (three months on their own) is 35.08% in 1Q22, and the two ends of the harvesting phase read off the TTM curve — the trailing four quarters summed — are 29.07% in 4Q22 and 17.01% in 3Q25. Across the 13 quarters in between, revenue growth never once came in below advertising growth. If each quarter's sign were a coin flip, a run like that is one case out of 8,192, or 0.0122%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
In the peer comparison drawn from the same ledger, over the very five quarters in which the subject company's TTM share fell 7.79 percentage points, one company in the same market rose 2.62 points and another stayed flat. Two companies that raised their absolute advertising spend around 3.4-fold from 2022 to 2025 saw their share rise as well; what set the subject company apart was a revenue multiple of 3.84 over the same period. In the negative control group, two contract manufacturers, advertising over revenue stayed flat — one below 0.5%, the other around 2%. The lagged correlations are +0.819 at the same point in time and +0.552 one quarter later, and the reverse direction (revenue to advertising) is larger at +0.668 one quarter later. With 25 observations the standard error on a correlation runs 0.21 to 0.23, so correlations sitting within 0.2 of each other cannot be told apart from this sample. Quarters were returned to standalone three-month values and the fourth quarter was taken as the annual figure minus the nine-month cumulative; all 14 self-checks of "four quarters sum to the annual figure" (7 years × 2 accounts) closed with a residual under 1 million won. That check caught one real case: a formatting error where using the "prior three months" revenue column of one report threw that year off by about 5.8 billion won.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
The monthly deliverables of the growth operations program (q-growth) include 2 priority actions
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
This article records 4 entry points as "not found" (ChatGPT, Perplexity, Gemini, Agent i) — all of them results of a search restricted to the official domain, run on 2026-07-15
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-11
-
"言及率が18%と出たとして、その業種で普通なのか、低いのか" ("If your mention rate comes out at 18%, is that normal for the industry, or low?") — a format example of a hypothetical phrasing, not a claim of fact. The 18% is an illustrative figure, there to make the point that without a comparison baseline a single in-house number cannot be judged high or low.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Advertising expense in absolute terms: 64.1 billion won (FY2020) → 266.7 billion won (FY2025) — 641 and 2,667 in hundred-million-won units. In the back-calculation and source-figures paragraphs the unit is dropped and the numbers are written bare as "641" and "2,667" (line 48: "dividing 641 by 0.2916 gives 219.8 billion won"; line 105: "four figures were used, advertising expense 641 and 2,667")
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2027-02
-
Question count is the only thing that separates the diagnostic tiers, so one question's weight in the whole result changes with the tier — in a 5-question run (Radar) a single question is 20% of the result, and in a 50-question run (Panorama) that same one question is 2%.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
The phase figure's vertical axis runs from 15% to 30% on 5-point ticks — 15%, 20%, 25%, 30%. That range brackets the series drawn on it, the TTM advertising-to-revenue share, whose trough is 17.01% and whose peak is 30.26%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
The density meter in the citability check places its value on a scale that runs from 0% to 100%. It reads the floor and the target straight from the thresholds set in code and draws both marks on that scale.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Free page check, daily limit — 10 a day on a signed-in account, 3 a day without signing in. Re-opening a saved result falls outside that limit.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
A period with 0 observations reads “not read yet”, not 0, and while the observation count stands at 0 the impression and performance charts stay empty — the screen keeps “measured, and it came out 0” apart from “not read yet”.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Operational backups rotate, so a deletion takes about 4 weeks to reach them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
Operational backups rotate, so a deletion takes about 4 weeks to reach them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
We act on an account or data deletion request within 5 business days, and you hear from us the moment it is deleted.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
VAT (10%)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
No product runs longer than 1 year.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2027-01
-
One answer cites more than one domain, so the column does not add up to 100%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
1 shoot, many pieces — a single shoot yields the vertical short-form, the horizontal version and the product cuts together (how many pieces is set contract by contract, and this sentence promises no number).
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
We are a Korean corporation — a foreign business for Japanese tax purposes — and the service we provide to businesses in Japan falls under "provision of telecommunications-based services to business customers" in the Consumption Tax Act. Under the reverse charge mechanism, the business receiving the service files and pays the consumption tax on it. This is the advance notice required by Article 62 of the Consumption Tax Act. For the time being, however, you need not include this purchase in your filing if the taxable period falls under any of the following: general taxation with a taxable sales ratio of 95% or more, the simplified taxation scheme, or the 20% special rule (2割特例). See National Tax Agency, Tax Answer No. 6118.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down
-
Of the 206 places, we read 64 right now — 24 in Korea, 16 in Japan, 24 in the US. Another 13 are real but not yet folded into that count.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-11
-
Search share comes from StatCounter — a dataset built from visits that reached sites through search. We take the median of the last 12 months.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-11
-
10,000+ influencers and creators — from nano up to 1 million followers.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-11
-
Sign after receiving the proposal and the discount runs up to 30% — the annual-contract discount and the full credit of the diagnostic fee stack together, and the contract terms leave room on top of that.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down · Recheck due 2026-11
-
We do not count a citation in an AI answer as a first-place search result. Some of the market's standard tools do: when an AI answer names a brand as a source, they record that brand at rank 1 in search.
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-11
-
Share of searches showing an AI summary where a plain link was clicked — Pew Research Center, observed 2025-07-22
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-11
-
Annual revenue of the ad verification firm DoubleVerify — recalculated from the audited figures in its US SEC filings (XBRL)
How to prove this wrong — open the source; if the value differs from this sentence, we take the line down · Recheck due 2026-11
-
AI Visibility Index, September 2026 issue (No. 2), law firms (KR) — surface leadership counts, our own measurement, window 2026-09-03 KST (Seoul). Anonymous A, Anonymous B, Anonymous C and Anonymous D each reached the top band on 5 of the 6 AI answer surfaces measured, a four-way tie. The top band is the set of brands whose own Wilson 95% CI overlaps that of the surface's leading brand by mention rate. Anonymous F reached it on four surfaces, and Anonymous G and Anonymous E on three each — leadership counts 5 · 5 · 5 · 5 · 4 · 3 · 3 · 0 · 0, a roster of 9 brands, every brand shown anonymized on public pages. Band membership states a tie by CI overlap, and we do not produce a single composite ranking. The change against the previous issue is an observed delta from re-measuring the same query set, surfaces and repeat count; this measurement does not separate out what moved those values.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 2, law firms (KR) — grid and distribution, our own measurement, 2026-09-03 round. Every pairing of the 9-brand roster with the 6 AI answer surfaces was published, mention rate and citation rate each carrying a Wilson 95% CI (108 cells). The denominator behind each value is that surface's valid runs in which an AI answer was displayed: 70 on all six surfaces, with 0 runs showing no answer. In the previous issue Google AI Mode stood at 43. Representative values: Anonymous A on ChatGPT, mention rate 46% [35–57]; Anonymous B on ChatGPT, 47% [36–59]. Median roster mention rate by surface ran from 10.0% to 51.4%. The change against the previous issue carries a 95% confidence interval on the difference of two proportions (derived from the Wilson intervals, Newcombe), and only the 19 cells whose interval excludes zero are listed — for the other 35 this sample could not tell the difference apart, which is not the same as there being none. Denominators differ from surface to surface, so we neither mix nor compare across them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 2, law firms (KR) — measurement receipt, our own measurement. 420 jobs (10 sealed queries · 6 surfaces · 7 repeats, with a brand mention judged up to 70 times per surface) closed 420/420, 0 unmeasured and 0 unjudgeable, and measurement status was recorded in 3 states: unmeasured, unjudgeable, absent. Protocol k7, measurement window 2026-09-03 16:12 to 20:22 KST (Seoul), confidence intervals by the Wilson 95% method. Outputs are sealed with sha256 so they can be checked against ours.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
In our own research measurement of one sector — 9 law firm brands read across 6 AI answer surfaces (2026-09-03, 70 valid observations per surface, 420 in total) — a leading brand already held the answer seat on all 6. Per-surface leader mention rates ran 14–83%.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index, September 2026 issue (No. 2), skin and plastic surgery clinics (KR) — surface leadership counts, our own measurement, window 2026-09-03 KST (Seoul). Anonymous B reached the top band on 6 of the 6 AI answer surfaces measured. The top band is the set of brands whose own Wilson 95% CI overlaps that of the surface's leading brand by mention rate. Anonymous A, Anonymous C, Anonymous E and Anonymous D each reached it on five surfaces — leadership counts 6 · 5 · 5 · 5 · 5 · 4 · 4 · 4 · 4, a roster of 9 brands, every brand shown anonymized on public pages. Mention rates in this sector sit low, so the top band overlaps widely and band membership should not be read as a ranking. The change against the previous issue is an observed delta from re-measuring the same query set, surfaces and repeat count; this measurement does not separate out what moved those values.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 2, skin and plastic surgery clinics (KR) — grid and distribution, our own measurement, 2026-09-03 round. Every pairing of the 9-brand roster with the 6 AI answer surfaces was published, mention rate and citation rate each carrying a Wilson 95% CI (108 cells). The denominator behind each value is that surface's valid runs in which an AI answer was displayed: 70 on all six surfaces, with 0 runs showing no answer. In the previous issue Google AI Mode stood at 53. Representative values: Anonymous B on ChatGPT, mention rate 7% [3–16]; Anonymous A on ChatGPT, 6% [2–14]. Median roster mention rate by surface ran from 0.0% to 1.4%. Where a surface's mention rate is 0%, the top band can form across the whole roster, and the page says so, to keep band membership from being cited on its own as evidence of presence. The change in mention rate against the previous issue carries a 95% confidence interval on the difference of two proportions (derived from the Wilson intervals, Newcombe), and only the 2 cells whose interval excludes zero are listed. Denominators differ from surface to surface, so we neither mix nor compare across them.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
AI Visibility Index No. 2, skin and plastic surgery clinics (KR) — measurement receipt, our own measurement. 420 jobs (10 sealed queries · 6 surfaces · 7 repeats, with a brand mention judged up to 70 times per surface) closed 420/420, 0 unmeasured and 0 unjudgeable, and measurement status was recorded in 3 states: unmeasured, unjudgeable, absent. Protocol k7, measurement window 2026-09-03 17:35 to 19:39 KST (Seoul), confidence intervals by the Wilson 95% method. Outputs are sealed with sha256 so they can be checked against ours.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
-
In our own research measurement of one sector — 9 skin and plastic surgery clinic brands read across 6 AI answer surfaces (2026-09-03, 70 valid observations per surface, 420 in total) — a leading brand held the answer seat on 5 of the 6. Leader mention rates by surface came to 7% on ChatGPT, 6% on Claude, 11% on Perplexity, 17% on Grok and 6% on Google AI Mode; on Naver AI Briefing every brand on the roster sat at 0%, so that surface has no leader.
How to prove this wrong — if the value differs from the copy we froze on the day we published, we take the line down
Entries in verification
Held as a count until the sourcing is ready — currently 23 entries.
Pre-publication check record
Since 2026-07-16, this registry has logged 6 statements the check stopped before they reached a page. Everything above cleared it.
Withdrawn claims
The addresses below were registered claims we have since withdrawn. If you cited one, that sentence is no longer something we claim. We do not republish the wording here.
#claim-ce2f268fee0#claim-c178a04c01c#claim-c82e2e97f56#claim-cf2694c5d1b#claim-c49f2ba34b5#claim-c04412ef80d#claim-c18f7a6430f#claim-c1e8fcafc7d#claim-cc6df07b141#claim-c60552f5e6a#claim-cf9fa90d951#claim-c1464ae5475#claim-cf18f4f300d#claim-c0a4c1876f2#claim-cab53df69e9#claim-ca04c236de9#claim-c970a2202ac#claim-c0dae4b4299#claim-c13edf05dec#claim-c983f61a3e1#claim-ce6d021ff01#claim-c04b001a7dc
The other ledger
This page collects what we claim. The readings we missed and the times our own instruments were wrong go on the page next door, along with how often AI answers cite us. Our own numbers
Every item a buyer checks
Before you buy a measurement you want to know where the questions went and how we built the set, and whether anyone checked the readings afterward. Our answer to each sits in one table, next to the page holding the value. See what we disclose · What we fix
Recompute it yourself
Each published rate is the number of times a brand appeared divided by the valid observations behind it, and the bracket beside it is a Wilson interval — the range the true value plausibly sits in once sampling noise is accounted for. One line below reproduces the 95% interval.
- Formula
- rate = appearances / valid observations · interval = Wilson 95% (z = 1.959963985)
- Command
python -c 'import sys,math;k,n=int(sys.argv[1]),int(sys.argv[2]);z=1.959963985;p=k/n;d=1+z*z/n;c=(p+z*z/(2*n))/d;h=z*math.sqrt(p*(1-p)/n+z*z/(4*n*n))/d;print(f"{p:.4f} {c-h:.4f} {c+h:.4f}")' 17 43The two trailing numbers are appearances and valid observations, in that order. Swap in your own and it computes them. Python is the only thing you need — there is nothing to install. It prints three numbers: the rate, the low end, the high end.
What opens here is the arithmetic. Execution order — which queries go in and in what sequence — travels inside the contract.
Registry reference date 2026-09-05. A statement withdrawn from a published page leaves the registry, and this page with it. The deployment checklist requires every registry version to be anchored to an external web-archive snapshot and to an independent timestamp we cannot alter.