How we measure AI visibility

Your visibility score is the share of real buyer questions where an AI engine mentions your brand: visibility = mentions / total samples. If we ask 25 questions 3 times each and your brand shows up in 30 of those 75 answers, your visibility is 40%. We report this per engine and per week. We deliberately do not report a “rank” or “position” — see below for why that number would be noise.

Each engine is checked on the surface your customers actually use, not a proxy:

Surface disclosure: Perplexity is measured via the official API — API answers can differ from what a logged-in consumer sees in the Perplexity app (no personalization, different model routing). The DataForSEO-collected surfaces (ChatGPT, Google AI Overviews / AI Mode, Gemini) are scraped from clean logged-out sessions — they can likewise differ from a logged-in user’s personalized UI. We label each engine with the surface actually measured and never present one surface as another.

Every report and dashboard footer names the exact engine set behind its numbers (“Measured AI services: …”), generated from the engines actually used — never a hardcoded list. We report what those measured engines recommend; we never claim to know “what AI recommends in your country” unqualified. Where a market has significant AI services we do not measure yet (for example Yandex Alice and GigaChat in Russia), the footer says so explicitly until they are added.

The same honesty applies to competitors: not every brand named in an AI answer competes in your market. Detected competitors default to unclassified; only competitors confirmed as direct (by you, or by a classifier above a strict confidence floor) drive Competitor Gap and the money estimate, and market dominance is only ever computed from our own measured answers, never asserted from world knowledge.

AI answers are not deterministic — the same question can produce different shortlists on consecutive runs. So we never rely on a single answer. Every monitored prompt runs as N prompts × 3 samples, every week, per engine (Claude via the official API with web search is sampled 1×/week). Google AI Overview / AI Mode runs where no surface renders are recorded as surface-absent and excluded from the visibility denominator, not counted as an absence. Around every visibility number we compute a Wilson score interval at 95% confidence (z = 1.96):

That band is drawn on every chart. If the band is wide, we tell you the sample is small — we never pretend to more precision than the data supports.

Being named in an answer is not the same as being recommended. Our headline metric, Recommendation Share, counts only the answers where the AI puts you on an explicit shortlist (a numbered/bulleted “best options” list) or endorses you directly (“we recommend…”) — the answers that actually shift a buyer. It uses the same Wilson CI as visibility. If an engine mentions you in 40% of answers but only recommends you in 15%, that gap is the most actionable thing on your dashboard: the model knows you exist but doesn’t pick you.

The order in which AI engines list brands is close to random between runs. Independent research (see SparkToro’s study) shows list order in LLM answers is unstable run-to-run, so a tool telling you that you “moved from #4 to #2” is selling noise. We show the median mention order only as a secondary signal, clearly labeled as close to random, and never treat it as a metric to optimize.

When your visibility moves, we test whether the move is statistically real before we alert you. The current week’s rate is compared against your rolling 4-week baseline using a two-proportion z-test:

We fire a drop or gain alert only when p < 0.05 two weeks in a row, or when the change is at least 15 percentage points on a sample of 30+ answers. We also mark known AI model releases on your charts — when an engine ships a new model, everyone’s numbers jump, and you should know the difference between “the model changed” and “your visibility changed”.

This is the part everyone else fakes. When you implement a fix, we freeze a baseline window and tag the prompts the change is meant to move (the target group). We also hold out a set of untouched prompts in the same niche (the control group). After a 3-day lag — long enough for the change to settle — we measure both groups over an equal-length window and compute a difference-in-differences:

Subtracting the control’s change removes everything that moved for the whole market — a model update, a seasonality swing, a competitor’s campaign. What’s left is the part plausibly caused by your change. We report it with a confidence interval and one of five verdicts:

For a single small brand there may be no clean control group. In that case we still show the before/after with its CI, but we cap the verdict at “probable” and say plainly that part of the move could be market drift. We would rather under-claim than sell you a false certainty.

We generate buyer questions from your niche across the whole journey — problem, solution, comparison, shopping — with at least 70% of prompts never mentioning your brand (that’s where discovery actually happens). You can edit any prompt. Edits create a new frozen version of the prompt set, and every chart shows a visible trend-break marker at the week the set changed — so you never mistake a questionnaire change for a market change.

Every number in your dashboard links back to the full answer text it came from, with the engine, timestamp, and the sources the engine cited. No black boxes: if we say you were mentioned in 12 of 75 answers, you can read all 75.

AI engines and data providers time out. When that happens, failed runs are excluded from the denominator, the affected report or weekly snapshot is explicitly flagged as partial, and the gap is visible on your charts. We never fill a gap with an estimate, and we never silently pretend a missing answer was a “not mentioned”.

Nobody can guarantee placement in an AI answer — and you should be wary of anyone who does. AmIVisible shows you where the odds are, what changed, and whether a change is distinguishable from noise. We measure the best available estimate with honest bounds; we don’t sell certainty that can’t exist. When the honest answer is “it didn’t work” or “we can’t tell yet”, that’s exactly what you’ll see.