Research · Action guide Action

Before you buy an AI visibility tool, stand up the official rails Google and Bing already give you

Two official publisher reports for AI search already exist, and both are free: Search Console's generative AI performance report and Bing Webmaster Tools' AI Performance preview. Add GA4's AI-referred sessions and you have three rails of first-party evidence before a single vendor invoice. Here is exactly what each rail proves, what it structurally cannot, and the operating structure that keeps the evidence honest.

Two official reports are already measuring your AI search presence

Most teams shopping for an "AI visibility" dashboard have not yet opened the reports the platforms themselves publish. Google Search Console now ships a dedicated generative AI performance report, rolled out progressively to properties, covering impressions from AI Overviews and AI Mode — viewable by page, country, device, and date, and exportable from the UI. Official primary support.google.com

On the Microsoft side, Bing Webmaster Tools announced AI Performance as a public preview on February 10, 2026. It reports citation activity across Copilot, Bing's AI-generated summaries, and select partner integrations: total citations, average cited pages per day, citation trends, sampled grounding query phrases, and citation counts per URL. Official primary blogs.bing.com

Both come from the operator of the surface, both are tied to your verified property, and both cost nothing. If AI search matters to your pipeline, these rails should be live before any tool evaluation starts — partly because they are the baseline any vendor's numbers must reconcile against, and partly because knowing what they cannot show you is how you decide what to buy.

What Google's generative AI report proves — and what it structurally cannot

The report answers one question with platform authority: how often links to your site appeared inside Google's AI experiences, and where. Its measurement rules are published: an external-link click from AI Mode counts as a click, a follow-up question is treated as a new query, and an AI Overview link impression is counted when the part of the answer holding the link is expanded or scrolled into view — with links in the same container sharing one position. Official primary support.google.com

What it structurally does not carry: the prompts people actually typed, the answer text, which exact URL was cited, what the answer said about your brand, and whether a citation caused a click. As of our July 14, 2026 review, we found no public read API for this dedicated report — plan around UI exports rather than assuming a feed. And one discipline matters from day one: if the report is not visible on your property yet, that is a rollout state, not evidence of zero AI impressions. A blank and a zero are different facts. Official primary support.google.com

What Bing's AI Performance preview proves — and where it stops

Bing's preview is the closest thing to official citation evidence on the market today: the platform itself telling you how often your pages were used as grounding, with sampled query phrases and per-URL counts. That is a real signal you can watch move. Official primary blogs.bing.com

Read it for what it is. A citation count is proof your pages are being pulled into answers — it is not proof of clicks, and the preview does not show the full answer a user saw, the position or role of your citation inside it, or how it converted. It is a UI surface: as of our July 14, 2026 review we found no public read API, so treat exports and captures as the record. Rising citations mean your content is in the grounding pool. Which questions, said how, against whom — that lives outside this report.

GA4 tells you the floor, and submission receipts are not performance

The third rail is your own analytics. GA4's session source and medium dimensions surface sessions referred by AI services — an honest, observable lower bound of AI-driven visits. It is a floor, not the total: an AI referral is visible only when the service sends a referrer, and apps, browsers, and redirect policies routinely collapse such visits into direct or unknown. Use it as the ground-truth of arrivals, not as the size of AI's influence. Official primary developers.google.com

One trap to clear away while you are wiring this up: indexing and submission APIs measure nothing. Google's Indexing API is scoped by policy to job postings and livestream markup, and a success response from any submission channel — including IndexNow — means the request was received, not that anything was crawled, indexed, ranked, or cited. Keep delivery receipts in a different drawer from performance evidence. Official primary developers.google.com · indexnow.org

Three rails, kept separate — the structure that keeps the numbers honest

Everything above sorts into three rails, and the operating rule is simple: report them side by side, never blended into one score.

The three-rail operating structure RAIL A Platform telemetry Search Console · Bing Webmaster Tools · GA4 — APIs and exports Proves: impressions, clicks, sessions on your property Cannot: see inside AI answers RAIL B Official UI receipts Google generative AI report · Bing AI Performance — exports Proves: AI impressions, citation activity Cannot: prompts, roles, causality RAIL C Controlled observation Fixed question panel, repeated runs of the answers themselves Proves: presence, mention, citation and role per question Never labeled as platform data Reported side by side — each number keeps its source and its limits Never blended into a single "AI visibility score" that hides which rail said what

The habit that makes this work is stating, for every cell in a report, whether it was observed, reported as zero, available only in a UI, or simply not offered by any official surface. Zero and unobserved are different answers to different questions, and a structure that refuses to confuse them is what lets you compare a Google number, a Bing number and a panel number without fooling yourself. This is the same rail-and-state frame our own reporting runs on — the full framework is public on our methodology page.

The gap the rails leave open — and how to close it

Line the rails up against the chain you actually care about and the gap is visible.

What each rail observes along the visibility chain Indexeddiscoverable Seen in searchimpressions Clickedsessions Cited in AIgrounding use Role in answerwhat it says Outcomeskey events RAIL A · platform telemetry GA4 key events RAIL B · official UI receipts AI impressions Bing citations RAIL C · controlled observation the stage no official surface reports — the answer itself

No official rail tells you which buyer questions your brand shows up in, what the answer says when it does, or whether the citation next to your category goes to you or to a competitor. Those are the facts that decide where your next content and PR dollar goes — and they are exactly what a controlled observation panel exists to capture: a fixed set of real buyer questions, asked repeatedly across the AI surfaces that matter, with every appearance, mention and citation recorded per question.

That panel is what CiteAngle operates as Rail C — engineered to the same evidence discipline as the official rails it sits beside. Our Panorama diagnostic runs a 4,900-cell measurement grid in full and shows the status of every cell, and applicable citation rates are reported with Wilson score 95% confidence intervals, so you can tell a real shift from noise before you reallocate budget. You get the official rails' story and the inside of the answers, in one report that never mixes the two.

Turn the official rails on this week — they are free, and they are the baseline any vendor should be measured against. Then, when you want to see the answers themselves — which questions you win, who gets cited instead, and what to fix first — see how our measurement tiers work or request a scoped proposal with your market and target buyer.

Ready to turn US visibility evidence into a market plan? Tell us the US buyer, target state and decision date.

Request a working scope