How to Measure AI Search Visibility: A Sampling Sheet

Ask each engine the same buyer prompts on a fixed schedule, save every answer, and calculate the share of answers that mention, recommend or cite your brand. Report each rate as a count out of a total, per engine and per country, never as one blended score.

An engine is an AI answer product: ChatGPT, Perplexity, Gemini, Google AI Mode (Google's conversational search) or Google AI Overviews (the summary above regular results). An answer is one generated response to one prompt, run once. Count these five things:

A mention means the answer names the brand. A recommendation means it frames the brand as an option for the user's need. A citation means the interface visibly presents a source link to a specific page. We keep all three separate, because collapsing them into one score hides the reason a brand appears or disappears. A brand can be named without being recommended, recommended without receiving a source link, or cited without receiving favorable language.

Position needs the most restraint. Record it only when an answer presents a true list of comparable options, then report the distribution or median across repeated answers. An ordinal place in one generated answer is not a stable ranking.

A ranking report also misses much of what AI answers draw on. In a study of 55,393 trending queries (13 March to 21 April 2026), accepted to the 2026 Association for Computing Machinery (ACM) Internet Measurement Conference, about 30% of the sources Google AI Overviews cited did not appear in the first page of Google's regular results (arXiv 2605.14021).

Wider than most reports admit. Thirty mentions in 100 answers means the true rate is plausibly anywhere from 22% to 40%, and 400 answers only narrows it to about 26% to 35%. Because generated answers vary, each rate is an estimate from a sample.

Sampling the same 200 queries per topic once a day for nine days in February 2026 showed that many apparent differences between domains on Perplexity Search, OpenAI's SearchGPT and Gemini fall within normal measurement noise (Sielinski, arXiv 2603.08924). A rate needs a range beside it.

We calculate that range with the Wilson score interval, a standard way to put a margin of error on a percentage from a sample. NIST's Dataplot reference lists it among the recommended methods for proportions. With x answers that mention you out of n collected, observed rate p = x ÷ n, and z = 1.96 for 95% confidence:

range = [ p + z²/2n ± z·√( p(1−p)/n + z²/4n² ) ] ÷ ( 1 + z²/n )

Our calculation for the same observed rate at different sample sizes:

In a spreadsheet, put mentions in A2 and answers in B2, then:

• Low: =((A2/B2+1.96^2/(2*B2))-1.96*SQRT((A2/B2)*(1-A2/B2)/B2+1.96^2/(4*B2^2)))/(1+1.96^2/B2)
• High: the same formula with +1.96*SQRT(...) in place of -1.96*SQRT(...)

This assumes every answer is an independent draw. Repeating one prompt many times is not fully independent, so real ranges are wider than these. A ten-prompt snapshot is useful for reading what engines say about you, but it is not a rate.

About 323 answers for a range of plus or minus 5 points on a rate near 30%, and about 356 answers in each period to detect a move from 30% to 40% with 80% power, meaning an 80% chance of catching a real change of that size. Both are our calculations, and both assume independent answers.

The working:

• Precision: n = z² × p(1−p) ÷ margin² = 1.96² × 0.3 × 0.7 ÷ 0.05² = 322.7, so 323.
• Detecting a change: for 30% versus 40% at 95% confidence and 80% power, n per period = (1.96·√(2·0.35·0.65) + 0.8416·√(0.3·0.7 + 0.4·0.6))² ÷ 0.1² = 355.9, so 356.

What happens when you fall short: with 100 answers in each of two periods, a rise from 30% to 40% has a 95% range of −3.1 to +23.1 points on the difference. That range includes zero, so you cannot call it a change. With 400 answers in each period, the same rise gives +3.4 to +16.6 points.

To test any two periods, compute 1.96 × √( p₁(1−p₁)/n₁ + p₂(1−p₂)/n₂ ) and check whether the observed difference is larger than that margin. If not, report "no detectable change."

Our recommendation is to buy breadth before depth. Cover many prompts across your buyer's intents, run each three to five times, and pool weeks into a monthly report. Rerunning the same 20 prompts 50 times shrinks run-to-run noise but leaves you blind to every question you did not ask. Print the smallest change your sample can detect on every report.

Write 20 to 100 buyer prompts tagged by intent, fix the run conditions, and log one row per answer. The tagged prompt set is the measuring instrument, and the log is the evidence behind every rate.

Prompt set. Include category discovery ("best X for Y"), comparison ("A vs B"), use case, objection ("is X worth it") and branded prompts. Take the wording from sales calls, support tickets and your Search Console queries, not from your keyword list. Give each prompt an ID, an intent tag, an owner and a version number. Never edit prompt text mid-series; a new wording is a new prompt.

Run conditions. Use a signed-out session (no logged-in account, so earlier chats do not shape the answer), a fixed country, the same engine surface every time, and record the date. Run each prompt three to five times.

Response log. One row per answer:

Five checks before you trust the sheet:

1. Denominator check. Every rate in the report reads "x of n." Fail if any rate has no n.
2. Replay check. Rerun 10 prompts on one engine the same day. If your mention call flips on any prompt, raise that prompt's runs before reading anything into it.
3. Coding check. Two people code the same 20 saved answers. Pass if they agree on every mention and citation call; any disagreement on "recommended" means the definition needs rewriting.
4. Alias check. Search each saved answer for brand-name variants, misspellings and your domain. Pass if the count matches your logged mentions.
5. Breakpoint note. When an engine changes models, a prompt is reworded, or the collection method changes, start a new series.

For a quick pass or fail screen before building all this, our 20-prompt content gap test runs each prompt twice in three tools, which is 120 answers.

Because engines cite different kinds of sources and some change their sources by country, a blended rate hides where you win and where you lose. Keep one row per engine, per country, per period.

In Study 007, we asked four engines 150 decision questions, phrased as Indian users would ask them, once each on October 5, 2026, and asked 50 of those questions from the US as a comparison, 800 answers in all. The counts below are the 150 India-asked answers per engine: community platforms such as Reddit, YouTube and Quora were cited in 117 answers from Google AI Mode, 73 from Perplexity, 29 from Gemini and 3 from ChatGPT (Citation Research Study 007).

In Study 008, we asked 60 buying questions from India, the UK and the US, giving 720 answers collected on October 5, 2026. Perplexity returned identical source lists in all three countries, while ChatGPT and Google AI Mode changed sources by country (Citation Research Study 008).

Both studies were single runs collected signed-out through a third-party service, and the country evidence rests on India-asked answers. They show that engines differ, not what your US rates are. Before relying on any engine difference for your market, run your own US sample on your own prompts.

They show what happened on your site or in their own AI features: impressions in Search Console, Copilot citations in Bing, and assistant visits in Google Analytics 4. None records whether an answer recommended you over a competitor, so they complement sampling and do not replace it.

Google states that AI Overviews and AI Mode traffic is included in the Performance report under the "Web" search type, and that there are no additional requirements to appear in them (Google Search Central). For the regular expression (regex), a text pattern that matches several referrer names at once, and a test to catch what the native channel misses, use our GA4 guide.

Compare the two periods' difference to its margin, then look at the saved answers before you act. A change is a question to investigate, not a verdict. Begin with the response record: which prompts changed, on which engine, in which country, after which model or collection change.

When sentiment changes, keep the exact language that drove the label. A favorable score without the answer text is hard to trust, and a negative score without the objection is hard to act on. Access can support discoverability but does not guarantee a citation, recommendation or top placement.

A spreadsheet is enough for a first read of about 20 prompts. It stops working when you need several hundred answers per period across several engines, because hand-transcribing that many answers is where errors creep in. At that point, automate the runs but keep the same log fields and ranges.

This is how we do it at PageLens.ai: we run your tracked prompts on a schedule across the engines on your plan, record each answer with the pages it cites, and report visibility, share of voice and the adjectives engines use to describe you. As of October 2026, our plans in US dollars are:

Agency pricing is custom. Annual billing saves 8% (PageLens.ai pricing).

Here is what those numbers mean for precision, by our arithmetic. Launch's 300 weekly answers across three engines is about 100 per engine per week if they split evenly. At a 30% rate that is a range of about 22% to 40% for one engine in one week, narrowing to about 26% to 35% when four weeks are pooled to 400. Growth's 500 daily answers across five engines is about 100 per engine per day. Daily reruns of the same prompts cut run-to-run noise but not the risk of having asked the wrong questions, so widen the prompt library as well as the cadence.

For a first baseline, our free audit takes your website URL and returns a one-time report on how often AI names your brand, which competitors it names instead, and which sources the answers draw on. Treat the audit as a read of wording and sources, not a rate; any rate from a handful of answers has a very wide range (3 of 10 is 10.8% to 60.3%). If you are weighing paid tools, our 14-day trial checklist lists what to test.

If you want your prompt set, baseline and ranges built for you, Book a demo.