How ClarAI measures AI visibility: the questions, the engines, the sample and what we cannot see

Which questions does ClarAI ask?

Your tracked set of buyer questions. ClarAI writes them on your first analysis from your personas, your product and your competitors. You can add your own and remove any, and a removed question stays removed. Since 24 September 2026 each analysis re-measures the same set rather than writing a new one, so one week compares with the next. New questions are written only on your first analysis, when you press Refresh prompts, or when you change your brand details.

Since 1 October 2026 at least half of the questions behind your score leave out your brand name. Before that, the sampled questions almost always named the brand, so the score mostly showed whether AI repeats a name it was given. Now it also shows whether AI brings you up on its own.

Which engines answer?

ClarAI reads three engines today: ChatGPT, Gemini and Claude, through the vendors' official APIs. Google AI Overviews is built but not switched on yet. Perplexity and Google AI Mode are not monitored.

Each analysis records which engines actually answered. When an engine does not answer on a run, because of a timeout, a quota limit or an API error, that result says so: the call is left out of your score instead of counting as not mentioned, and the dashboard names the engine that did not answer, with Gemini shown as Google AI.

How is each engine asked, and which ones search the web?

These are the default settings in our code as of 14 October 2026. Web search for ChatGPT and grounding for Gemini are switches we can turn on; when one is turned on, the first analysis after it is marked as a new starting point in your trends, and this table changes with it.

How does ClarAI ask ChatGPT?

Through OpenAI's official API, with the buyer's question only. ClarAI asks the model it names, gpt-4o-mini by default, and records the model with each answer. Without web search the answer comes from the model alone. OpenAI's web search tool is built in and is a switch: where it is switched on, ChatGPT can search the web before it answers and return the pages it used, and the first analysis after the switch is marked as a new starting point in your trends.

Either way, this is the API, not the ChatGPT app. A person signed in to ChatGPT gets the model and search behaviour OpenAI chooses for their plan, plus their own memory and personalisation, so the answers they see can differ from the ones ClarAI reads.

What does ClarAI send to the engines?

The buyer's question, as a buyer would ask it, and nothing else. Your brand name is never in the request. We changed this on 28 September 2026 after testing it on real projects: with the brand named in the request, ChatGPT mentioned it in 16 of 20 answers; with the question only, in 12 of 20. Claude showed the same pattern. It was a small test (20 questions on ChatGPT, 6 on Claude) and not statistically significant on its own; we changed the method because naming the brand can only push the count up.

Then ClarAI reads every answer itself. Your brand counts as mentioned when its name or website appears as a whole word, so a short brand name does not match inside a longer ordinary word. A small reader model settles names that are also ordinary words or places (a trip to Patagonia is not Patagonia the brand) and notes position, tone and the competitors named. The engine never grades its own answer.

How many answers sit behind a number?

Fewer than you might assume, and we say so. By default a full analysis asks each engine 4 of your tracked questions, so 12 answers across the three engines today. A quick first look asks fewer, and free accounts use a leaner sample. These are our code defaults as of 14 October 2026; whatever the setting, every reading prints the sample it actually rests on.

That is why every headline carries a coverage caption and a data-quality label. The label follows the IAB's August 2026 guidance, which "treats fewer than 50 queries per measurement program as exploratory rather than directional". A reading from fewer than 50 distinct questions is labelled Exploratory; from 50 or more, Directional. ClarAI's default sample sits well below that line, so most readings say Exploratory, and that is the honest label.

For a sense of scale, Mike King of iPullRank estimates that at a 50% appearance rate, 20 runs of one prompt give roughly plus or minus 22 points, and 50 runs roughly plus or minus 14. One answer is an anecdote; a rate needs repetition.

When is a change real?

• A range, not a point. Once there is enough history, the score and mention rate show a likely range and the number of answers behind it.
• A normal run-to-run range. It needs at least 3 analyses; with fewer, no band is shown rather than an invented one. It is measured on the analyses before your latest one, and it is never narrower than the latest run's own sampling margin, so a small sample cannot turn an ordinary wobble into an alarm.
• Like-for-like trends. Each analysis records which questions it asked, which engines answered and how it measured. When two analyses differ, the trend shows a marked break and "What changed" says why, so a change of method never reads as a change in visibility.
• Failed calls left out. An engine that did not answer shows as not measured, never as zero.
• Pooled readings. For each question and engine, answers are pooled over the last 28 days with a likely range, and no engine gets a verdict before 7 answers.

Which citations count?

Only real sources. In the default setup Claude answers with a live web search, and ClarAI keeps only the pages its answer actually cites, not every page it looked at. ChatGPT and Gemini, as we ask them by default, answer from the model without a search, so ClarAI shows no sources for them: what looked like sources there was the model's memory, and some of those pages did not exist. When an opportunity checks whether AI cited your page, it counts only engines that return real citations.

What can ClarAI not see?

A logged-in session. ClarAI asks the engines through their APIs, which can differ from what one person sees inside a signed-in ChatGPT, Gemini or Claude account: the model version, personalisation, memory and live experiments all move an answer. Read your score as the API view of your visibility, not a guarantee of any one person's screen. LLM Pulse, a vendor in the same category, makes the same point about logged-in users.

Why does every number show its sample size?

Because the number underneath it is not stable. Ask the same engine the same buying question twice and you can get two different answers, so one run is a sample, not a measurement. A tool that reports one precise score with no sample size is reporting one roll of the dice as if it were the house edge.

A coverage caption says exactly what it means. As an example of the format, a caption reading "4 of 72 prompts measured" means the project tracks 72 buyer questions and that run measured 4 of them. A partial scan should never read as a full one.

A confident single number hides three things: the sample size, so you cannot tell a thorough measurement from a single question; the variance, so ordinary noise looks like a gain or a loss; and the raw answers, so you cannot check whether the engine actually named you. ClarAI keeps the answer text and its citations beside every score so you can read them yourself.

What should you ask any vendor, us included?

Mike King's list of questions for any AI visibility vendor is a good test. Here are our answers.

How has the method changed?

Method changes only, dated, each with its release on the changelog.

• 24 September 2026 (v5.46.0). Each analysis re-measures the same set of questions, so weeks compare like with like, and ranges appear beside the score once there is enough history.
• 24 September 2026 (v5.48.0). Every reading carries a data-quality label, Exploratory or Directional, following the IAB's guidance.
• 25 September 2026 (v5.49.0). An engine that does not answer is left out of the score instead of counting as not mentioned, and only real cited sources are shown.
• 25 September 2026 (v5.50.0). Discovery appears beside the score, trends mark analyses that are not comparable, and the normal range is measured on earlier runs.
• 28 September 2026 (v5.52.0). The engines receive the buyer's question only, and ClarAI reads the answers itself.
• 29 September 2026 (v6.0.0). Readings per question and engine are pooled over 28 days, with no verdict before 7 answers.
• 1 October 2026 (v6.8.0). At least half of the questions behind the score leave out your brand name.
• 6 October 2026 (v6.15.0). Claude searches the web from the buyer's country for single-country markets.

To see what an engine can read on your own site today, the free grader checks your crawler rules, snippet controls and page structure with no signup. It measures readiness, not visibility.