How-to
How to measure your AI search visibility
A simple, honest method for measuring AI search visibility: which questions to ask, how often, what to record, and which free reports and tags to use.
By Reece Chown
· 8 min read
The short answer
There is no rank to look up in an AI answer, so you measure by sampling. Write the questions your customers ask, put them to each assistant several times, and record how often you are named and which sites are cited. Add three free sources: visits tagged utm_source=chatgpt.com, Bing's AI Performance report and Google Search Console's generative AI report. Then repeat monthly and compare.
Why this is harder than checking a ranking
With ordinary search, you can look up where you rank for a phrase. It is roughly the same for everyone and it changes slowly. AI answers break all three assumptions.
- Answers vary. Ask the same question twice and you may get different businesses named.
- Answers are personal. Location, wording, the model in use and the user's history can all change the reply.
- There is no position. You are named first, named in passing, cited as a source without being named, or absent.
- Nobody publishes the data. OpenAI gives site owners no report of impressions or mentions. Microsoft and Google now offer partial ones.
So the honest unit of measurement is a rate from a sample: "named in 4 of 6 runs", not "ranked third".
Step 1: write the question set
Everything depends on asking the right questions. Aim for 10 to 20, in four groups.
| Group | Example | What it tells you |
|---|---|---|
| Category | "Who is a good accountant for a small design studio in Leeds?" | Whether you are in the running at all |
| Problem | "How do I see file copy progress on a Mac?" | Whether you are found by people who do not know your kind of product exists |
| Comparison | "Alternatives to [best-known competitor]?" | Whether you are considered an option |
| Brand | "What is [your business] and is it any good?" | Whether what is said about you is accurate |
Rules for writing them:
- Use your customers' words. Look at your enquiry emails, search terms and support questions.
- Leave your name out of every group except the last.
- Keep them fixed. The value is in comparing the same questions over time. Change one and you lose its history.
Step 2: ask, and ask again
For each question and each assistant:
- Start a fresh conversation. Sign out or use a private window where you can.
- Make sure web search is on, if the assistant has a setting for it.
- Ask the question exactly as written.
- Save the whole answer and its list of sources.
- Repeat at least three times.
Do this for ChatGPT, Claude, Gemini, Perplexity and Copilot. Twenty questions, five assistants and three runs is 300 answers, which is a long afternoon by hand. That is the case for automating it.
Automating it
Most of these companies offer a way for software to send questions and receive answers, with web search switched on. A short script can run your question set on a schedule and save the results.
Two cautions from having done it:
- The developer route is not identical to the app. The model, the search behaviour and any personalisation can differ from what a customer sees in the consumer product. It is a consistent proxy, which is what you need for a trend, not a replay of anyone's actual session.
- It is cheap. My baseline of 39 questions through OpenAI's search-enabled model cost about $1.29 by a deliberately high estimate. If you are quoted a large monthly fee for "AI visibility tracking", ask what the fee is for. The asking is the cheap part.
Step 3: record the same things every time
For each answer, note:
| What | How to record it |
|---|---|
| Named? | Yes or no: does your business or product name appear in the answer? |
| Cited? | Yes or no: is your website among the sources? Which page? |
| Position | First named, named among others, or mentioned in passing |
| Accurate? | Is what it says about you true? Note any wrong facts |
| Who else | Which competitors were named |
| Sources | Which websites were cited |
"Cited" and "named" are different, and both matter. In my own tests I counted a result as a yes if the answer or its sources named my product or domain.
Then work out, for each assistant:
- Mention rate: the share of runs in which you were named.
- Citation rate: the share in which your site was a source.
- Share of voice: your mentions as a share of all business mentions across the set.
- Top competing sources: the websites cited most often when you were not.
Step 4: add the free data
Sampling tells you what assistants say. These tell you what actually happened.
Visits from assistants
OpenAI says ChatGPT adds utm_source=chatgpt.com to the links it sends people to. In your analytics, filter for that source. Other assistants show up as referrers, such as perplexity.ai, gemini.google.com and copilot.microsoft.com, though not every visit carries one.
If your site has no analytics, your server logs hold the same information. This site uses no tracking scripts at all, and I can still see these visits in the logs.
Expect small numbers. Many people read the answer and never click. A handful of well-matched visits can be worth more than a hundred idle ones: the customer who found my app through ChatGPT arrived ready to buy.
Bing Webmaster Tools: AI Performance
Since February 2026, Bing has reported how often your pages are cited in Copilot and Bing's AI summaries, which pages, and a sample of the "grounding queries" used to find them. It is free once your site is verified. How to set it up.
Google Search Console: generative AI report
Since June 2026, Search Console has shown impressions from AI Overviews and AI Mode, by page, country, device and date. Google's post says the report reached all sites by the end of August 2026. It counts impressions and does not separate the two features.
Crawler visits
Your server logs show which AI crawlers came, how often and what response they got. A crawler that stops visiting, or starts receiving errors, is an early warning that something has changed. Which crawlers to look for.
Ask your customers
The cheapest method of all. Add "How did you hear about us?" to your enquiry form, or ask in your first reply. That single question is how I learned a customer had found me through ChatGPT. No dashboard would have told me.
Step 5: read the results without fooling yourself
- One run proves nothing. In my own baseline each question was asked once. I reported it as a sample and said so. Three or more runs per question is the minimum for a rate you can compare.
- Small changes are noise. Going from 3 of 6 to 4 of 6 is not a trend. Going from 1 of 6 to 5 of 6 across several questions probably is.
- Give changes time. After changing your site, wait for crawlers to return. I allow two to three weeks before measuring again.
- Watch accuracy as well as presence. Being named with the wrong price is not a win.
- Models change. A drop may have nothing to do with you. Note the date and the model if you can, and look at whether competitors moved too.
What to do with what you find
| You see | Do this |
|---|---|
| Never named, never cited | Check access and indexing first. Start here |
| Cited but not named | Your page is useful but your brand is not clear on it. State who you are and what you offer near the answer |
| Named with wrong facts | Find the source of the error and correct it everywhere |
| Named for brand questions only | You are known but not considered. Write pages for category and comparison questions |
| The same competitor source wins repeatedly | Study that page. It shows what the assistant wants |
| Named by one assistant, absent from another | Look at that assistant's crawler and index: it may not be able to reach you |
A monthly routine
- Run the question set across the assistants, three runs each.
- Update mention and citation rates.
- Check the three free reports and your referral visits.
- Read every answer that got a fact wrong, and fix the source.
- Pick one question you are losing, open the pages being cited, and improve or write your own.
That is all monitoring is. It can be done by hand in a couple of hours a month for a small question set.
Frequently asked questions
Can I see how often ChatGPT mentions my business?
Not directly. OpenAI does not give site owners a report of mentions or impressions. You can sample by asking ChatGPT your customers' questions repeatedly, and you can count visits it sends, which carry utm_source=chatgpt.com.
What is a good AI visibility score?
There is no standard scale, and scores from different tools cannot be compared. A useful measure is your own mention rate on a fixed set of questions, tracked over time and set against the competitors named in the same answers.
How often should I measure?
Monthly suits most businesses. Measure before and after any significant change to your site, allowing two to three weeks for crawlers to revisit before the "after" measurement.
Do I need a paid tool to track AI visibility?
No. You can do it by hand or with a simple script, and the cost of asking the questions is small. Paid tools save time on large question sets and add reporting. Check exactly what they measure and how many times they ask each question.
Why do I get a different answer from my colleague?
Assistants vary their answers, and they may take account of location, wording, the model and account history. That is why a single answer is not evidence, and why measurement has to be a rate across repeated runs.
Sources
- OpenAI, Publishers and developers FAQ. Checked 10 November 2026.
- Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools, 10 February 2026.
- Google Search Central Blog, Introducing Search generative AI performance reports in Search Console, 3 June 2026.
- My own baseline of 39 questions across 11 websites, 10 November 2026: the case study.