How do I check what ChatGPT and other AI assistants say about my brand?

Write a fixed set of buyer questions and run each one three to five times in every assistant you care about, each time in a fresh chat with web search on. For every run, record whether your business is named, whether one of your pages is cited, which competitors appear and which facts are wrong. Note the date, the assistant and its settings. Report the rate across runs, never a single answer.

By Margareta Petrovic, founder of Visibility Mesh. We checked every setting and button name on this page against each vendor's help pages on September 26, 2026, and published this guide on .

I ran this check on my own company before I wrote it down for anyone else. Our weekly probe sends 14 buyer questions to Perplexity through its API. It records whether each answer names us and whether it cites one of our pages.

A second test limits Perplexity's search to a single domain. An answer with no sources is a strong sign that Perplexity has not indexed that site yet. I trust that test most, since its answer does not change with how many times you ask.

Audit practice treats a single screenshot as an anecdote, and I hold my own team to the same rule. OpenAI writes that search results and citations can be incomplete, outdated or incorrect (Searching the web with ChatGPT). Google says AI Mode and AI Overviews may use different models, so the responses and links they show will vary (AI features and your website). Microsoft lists changing models, freshness and user behavior among the reasons citation patterns shift (Bing Search Blog, June 16, 2026). Answers that shift this much can only be reported as a rate from a dated sample.

Before you start

  • You need a list of 10 to 30 questions real buyers ask, written the way they type them. That range is our rule of thumb for a small business, since no assistant vendor publishes a number.
  • You need access to each assistant you want to measure, with web search available, and a way to run it without your saved chats or memories. The table below shows the control each vendor documents.
  • You need a spreadsheet with one row per run and the columns listed in Step 5.
Each control is named as the vendor's own help page names it, read September 26, 2026. Links are in the Sources list.
Assistant Run it without your history Where the sources show
ChatGPT Open a new chat and select Temporary. Choose Unpersonalized before your first message, since a temporary chat set to Personalized can still use your existing memories. People who are not signed in can also use web search. Inline citations and the Sources button. Sources also lists other relevant links, which are not citations.
Claude An incognito chat, which is not saved to your chat history or to Claude's memory. Direct citations and source links. Turn on Web search from the + button, or in the new Claude experience let Claude search when it helps.
Perplexity Incognito mode. Perplexity says incognito sessions expire within 24 hours. Numbered citations in every answer.
Gemini Next to New chat, click Temporary chat. It needs a personal Google Account, and it gives no personalized responses. The Sources button at the bottom of the response or inline. No button means no links for that response.
Microsoft Copilot Open Settings, select Personalization and turn off the Saved memories toggle. This path applies to the app version of August 18, 2026 or later. Hyperlinked citations below the text when the answer is grounded in web search.

Steps

Step 1. Write the question set

Take the questions from your sales calls, your Search Console queries and the community threads your buyers read. Use three kinds. Category questions ask for the best option for a need. Comparison questions put two options side by side. Brand questions name you and ask what you do, what you cost and how you compare. Freeze the set for at least three months, and change it only on a date you write down.

Step 2. Fix the conditions

Write down the date, the assistant, the plan and whether you were signed in. Record whether web search ran and whether memory was off. Record where you ran it from as well. OpenAI says ChatGPT may use an approximate location based on your IP address, and that a VPN can change it. If memory is on, saved memories can shape the search query itself.

Step 3. Make sure web search ran

ChatGPT may search on its own when a question needs current information. To force it, select View all tools, then Search, or type / in the message box and select Search. In Claude, turn on Web search from the + button, unless you have the new Claude experience, which has no toggle. An answer with no sources may still be useful, but it tells you what the model remembers, not what it found today. Mark those runs separately.

Step 4. Run each question three to five times

Open a fresh chat for every run. Three to five runs per question and assistant is our starting point, and the check at the end of this guide tells you whether your site needs more.

Step 5. Record every run

Use one row per run. Record the question, the assistant, the date and your settings. Mark whether your business was named and whether one of your pages was cited. List the competitors named and the domains cited. In ChatGPT, count a page as cited only when the answer cites it, not when it appears among the other relevant links. If a sponsored ad appears below a ChatGPT answer, record it in its own column. OpenAI says ads are labeled as sponsored, appear for users on the Free and Go plans and do not influence ChatGPT's answers.

Step 6. Check what the answer says about you

List the 10 to 15 facts that matter most: your name, what you do, locations, prices or plans, founders and policies. Mark each fact in each answer as correct, outdated or wrong. For every wrong one, copy the source the answer cited. Then look for the same wrong fact on your own old pages, profiles and directory listings. If it sits on a page you control, correct it there first.

Step 7. Calculate rates across runs

Mention rate is the share of runs that name you. Citation rate is the share of runs that cite one of your pages. Share of voice is your mentions divided by all brand mentions across the same runs. Calculate all three across every run of a question, not per answer. Report each rate with the number of runs behind it.

Step 8. Repeat on a schedule

Run the same set on the same weekday each month. Put the first party reports next to your manual results: the Generative AI performance report in Search Console for Google, and the AI Performance report in Bing Webmaster Tools for Microsoft. Both come from the systems that showed your pages, so I trust their counts over any manual run, although neither shows what the answer said about you.

Check that it worked

Run the whole question set twice on the same day and compare the two mention rates. If they differ by more than the change you want to detect, your sample is too small. Add runs until the two rates settle.

If it did not work

  • If you tested from your own signed in account with memory on, the assistant may already know you, and the result looks better than a buyer's would. Use the controls in the table above.
  • If every link in the ChatGPT Sources panel went into your count, recount with only the sources the answer cites.
  • If two months used different question sets, they cannot be compared. Freeze the set, and date every change to it.
  • If a single run is being read as a ranking, report rates with the number of runs instead.
  • If a sponsored ad was counted as a mention, move it to its own column.

Platform notes

For this check, the assistant matters far more than the platform your site runs on, and the table above covers the assistants. Your own platform matters only when a wrong fact comes from your site. A Shopify store, a WordPress site and a hand coded site fix it the same way: correct the page, then check again next month.

What Visibility Mesh checks and what it does not

Our customer portal has an AI mentions check that you start yourself. It asks your questions of ChatGPT and Claude with web search and of Perplexity, and it records whether your brand is named. The Shopify app runs it as AI Mentions, which shows whether AI names your store or a rival. Neither one shows which URLs an answer cites, nothing runs on a schedule yet, and our checks do not cover Gemini or Google AI Overviews today. Our web scan answers the question that comes before all of this: whether AI crawlers can read your pages at all. If an assistant never cites you, run the scan first, since a page crawlers cannot read gives an assistant nothing to quote.

Questions people ask

Should a small business check what ChatGPT says about it?

Yes, and monthly is better than a few times a year. A customer may read an answer about you before they ever open your website. Ask what a customer would ask: what you do, your prices, your hours and how you compare. Do it in ChatGPT, Google and one or two other assistants. Write down anything wrong or out of date and where the answer seems to get it. Many errors trace back to an old page, a listing or a profile you can correct yourself.

How many questions should I track to measure AI visibility?

For a small business, I start with 10 to 30 questions that real buyers ask, and add more for each product line or market. That range is our rule of thumb, since no assistant vendor publishes a number. Mix category, comparison and brand questions. Freeze the set for at least three months so the months stay comparable, and change it only on a date you record.

How many times should I repeat a question to measure AI visibility?

Start with three to five runs per question in each assistant, each in a fresh chat. That is our starting point, not a platform rule. Run the whole set twice on the same day and compare the two mention rates. If they differ by more than the change you want to detect, add runs until they settle. Report every rate with the number of runs behind it.

Sources

This guide is one part of how to measure AI visibility. When an answer gets a fact wrong, correcting AI when it is wrong covers the fix, and your citation footprint covers where assistants find you.

Work through the six checks in How to check if AI can read your website, or hand them to your developer.

You can also run a free scan. We read five of your pages the way an AI crawler reads them, and you do not need a card or a call.
See whether this applies to your site

The free scan reads your own pages and shows which of these problems it finds on your site.

The free scan reads 5 pages of any website as AI crawlers receive them and returns a scorecard with 3 complete findings, each naming the page and the fix. No install, no call, no card.

Run the free scan