How to Track Your Brand in AI Answers

Why one check proves nothing

Ask an assistant the same question twice and you may get two different answers with different sources. Ahrefs describes AI Overviews as probabilistic, with citations and related queries shifting from one query to the next, and Microsoft warns that model changes, competing content and demand shifts can all move citation counts. A screenshot is an anecdote. A rate across many runs is evidence.

So the job is sampling: a fixed list of prompts, run the same way, repeatedly, with the results counted. Everything below is built around that.

Step 1: build the prompt set

Write about 20 prompts in four groups. The examples are illustrative, for an invented Limassol physiotherapy clinic.

Write prompts the way customers talk, not the way you describe your service. Pull wording from enquiry emails, reviews and Search Console queries. Do not change the wording between months, or you are measuring the wording.

Step 2: set the run rules

• Same assistants every time. For example Google AI Overviews, ChatGPT with search on, Perplexity and Copilot. Add Claude if your buyers use it.
• Same settings. Search or web mode on or off, location, language, and a clean browser profile where you can.
• Three runs per prompt. In a fresh conversation each time.
• Same day each month. And record the date.
• Do not coach the assistant. No "pretend you are a local expert" preambles. Use the plain question.

Step 3: log the right columns

Step 4: calculate and read the numbers

The worked example below is illustrative. With 20 prompts, 3 assistants and 3 runs each, you collect 180 answers a month. If you are named in 27 of them, your mention rate is 15%. If you are linked in 9, your link rate is 5%.

One answer is 0.56 percentage points of the rate, so small swings are mostly chance. As a working rule, treat a change under about 5 points as noise at this sample size, and act only on changes that hold for two months running. This is a rule of thumb based on the arithmetic of the sample, not a statistical guarantee, because repeated runs of the same prompt are not independent.

Report three numbers each month: mention rate, link rate and accuracy rate. Add the names of the competitors who appear most. That list is often more useful than your own rate.

Step 5: add first-party data

Your sampling shows what assistants say. First-party reports show what reached you.

• Google Search Console. The generative AI performance report counts impressions in AI Overviews and AI Mode by page, country, device and date. Its help page does not describe clicks or average position for the report, and it can be missing when a site has too few impressions. The numbers also sit inside the Web search type in the main Performance report.
• Bing Webmaster Tools. The AI Performance report, in preview, shows total citations, average cited pages, grounding queries and page-level citations across Copilot and Bing summaries. A June 2026 update added Intents, Topics, Citation Share and Compare. It does not show clicks or which competitors were cited.
• GA4 referrals. In June 2025 SEO consultant Glenn Gabe reported that links in ChatGPT's "More" section had started to carry utm_source=chatgpt; OpenAI had not documented it at that time. Check your own data. Build a custom channel group for AI sources above the default Referral group.
• Server logs. Look for OAI-SearchBot, Claude-SearchBot, PerplexityBot and the user-triggered agents. OpenAI publishes IP lists for its agents. A bot visit shows access, not that you were cited.

Step 6: check accuracy, then fix the source

If an assistant gets your hours, prices or services wrong, trace where the wrong fact might come from: an old page, a directory listing, a profile. Fix the source, then recheck next month. Google's AI Overviews page says the feature can and will make mistakes, and offers a feedback route on results.

Mistakes that spoil the data

• Changing prompts mid-stream. A rate for new wording cannot be compared with a rate for old wording. If you must change a prompt, retire the old one and note the break.
• Checking only when you feel good. Run the log on a fixed date, not after a good week.
• Logged-in personalization. Your own history shapes what an assistant shows you. Use a clean profile where you can, and note that real customers see something different.
• Counting mentions without reading them. Being named in a list of "companies to avoid" is not a win. The accuracy and tone columns exist for this reason.
• Reporting one number. A single rate hides whether the gain came from one assistant or one prompt group. Break it down.
• Treating tool scores as facts. A vendor's visibility score is the vendor's sample, with its own weighting.

A one-page monthly report

Keep the report short enough that someone will read it. Four blocks are enough:

1. Mention rate, link rate and accuracy rate this month, with last month's beside them.
2. The three prompts where you gained or lost most, with the assistant named.
3. Any wrong statement found, with the source you traced it to and its fix.
4. The competitors who appeared most, and the one change you will make next month.

When to automate

A spreadsheet handles about 20 prompts for one market. Past that, or with several markets and competitors, a tracker saves hours. Our AI visibility tools comparison lists published prices and what is included. Google-specific measurement has its own short routine in how to track AI Overviews.

What a visibility audit from us looks like

Our AI search optimization service begins with this method: a prompt set built with you, a baseline across the assistants you choose, a list of access and content fixes, and a repeat check. The report shows the data and the limits, and we quote in writing. To start on your own, write the first five prompts today.