About this series. Verify Before You Buy reads AI-visibility marketing claims the way a procurement desk should: one reporting pattern per part, taken apart and converted into questions you can ask in writing before you sign. The subject is the reporting format, not any particular vendor — every sales line quoted in this series is a labeled format example, and every measured figure we cite carries a source and a date you can check.
What a lift claim leaves out
A lift is a ratio of two rates, and every rate is a fraction. Before the "3×" means anything, four pieces of the fraction have to exist in writing:
- The denominator. A share of what — all tracked prompts? Only the prompts where an AI answer appeared? Only valid runs after failures are excluded? Each choice produces a different number from the same data, and a report that never names its denominator can switch between them without ever being wrong.
- The sample. How many prompts, and how many runs of each. A "lift" computed over a handful of prompts read once is a different object from one computed over a pinned panel with repeated runs — even when the headline percentage looks identical.
- The query-set version. Were the questions frozen between the "before" and the "after"? If the prompt list changed in the middle, the two measurements are of different things, and their ratio is not a lift — it is a coincidence.
- The window and the baseline protocol. Same engines, same locale, same repetition count, stated dates on both sides. Without this, "before versus after" quietly becomes "one setup versus another."
None of these questions accuses anyone of anything. They are the questions any rate has to answer before it goes into a budget — which is exactly why asking them in writing is such an efficient filter.
The denominator itself moves
There is a structural reason to insist on the denominator in writing: on AI surfaces, the denominator is not stable ground. It is a dial the platform turns.
The clearest published example: on Semrush's 10M+ keyword panel, 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 — vendor-published panel figures, data January through November 2025; accessed July 17, 2026). The surface roughly quadrupled and then shed a third of its footprint within one year, with no action required from any brand or agency.
The same denominator, three readings in one year
Share of panel queries that triggered an AI Overview (Semrush, 10M+ keywords, 2025)
Now put the format example next to that curve. A mention rate measured against "answers that appeared" can triple because your content earned it — or because the platform turned the dial and the denominator collapsed under the fraction. From the headline number alone, the two cases are indistinguishable. A lift measured against a moving denominator is not something you can procure against, unless the report states which denominator it used and holds it fixed across the comparison.
And even with a frozen denominator, a single-pass reading moves on its own: published measurement research reports 10–34% output variance from sampling alone (arXiv 2601.21339, 2026-01, preprint; accessed 2026-07-16). Rate movement smaller than that band is not evidence of anything — we collected the published record on this in a separate evidence review, which this series uses as its measurement-reliability anchor.
Five asks that fit in one procurement email
1. Ask for the denominator definition in writing. "Share of what, exactly — all tracked prompts, appeared answers, or valid runs?" One sentence to answer; the answer becomes part of the contract's vocabulary.
2. Ask for n, k, and the interval. How many prompts (n), how many runs each (k), and the confidence interval next to the rate. A stated interval turns a marketing number into a falsifiable one.
3. Ask whether the query set is version-pinned. The before and the after must run the same frozen question list, and the version identifier should appear in the report.
4. Ask for the window and the baseline protocol. Engines, locale, repetition count, and dates — identical on both sides of the comparison, in writing.
5. Ask whether the number survives a re-run. Under the same protocol, within what band is the number expected to reproduce? A vendor who publishes intervals can answer this directly; the published evidence above tells you why a single-pass number cannot.
Where we stand (held to the same asks)
We sell both measurement and execution — which is why every measurement ships with a receipt. The five asks above are the format our own reports are built to pass: every paid CiteAngle report states the denominator per metric, prints n and the Wilson 95% confidence interval next to every applicable rate, pins the query-set version, separates measured cells from unmeasured ones, and seals each run with a hash you can verify. The external figures we cite publicly — including every figure in this article — are logged in our public claims registry with source and access date, and the full protocol is on our methodology page. If you want to see the format on your own brand before any contract, start with the free snapshot.
Verify Before You Buy — reading AI-visibility claims
This is Part 1. The series takes one reporting pattern per part and converts it into questions a buyer can ask in writing — practices, never named vendors, with every quoted sales line labeled as a format example. Its measurement-reliability anchor is our published evidence review, A Single Run Is a Point Estimate, Not a Number. Next: Part 2 — coverage counts versus measured counts (forthcoming). All research: CiteAngle Research — series index.