How to check if ChatGPT mentions your brand (and track it month to month)

“Ask ChatGPT about us” is now a line in most marketing meetings, and the answer usually comes back as one screenshot, taken once, by someone logged into an account that has discussed the company for months. That is not a measurement. This guide shows how to get a number you can act on, what the citations tell you, how the other assistants differ, and what the tracking tools cost. If you want the other half of the question, how to make the answer better, read how to get ChatGPT to recommend your business.

Why one check proves nothing

In January 2026 SparkToro and Gumshoe.ai published the largest public test of AI brand recommendations: 2,961 runs of 12 prompts, each asked 60 to 100 times by around 600 volunteers on ChatGPT, Claude and Google's AI features. ChatGPT and Google returned an identical list of brands in under 1% of repeats and the same order in under 0.1%. One hospital appeared in 69 of 71 ChatGPT answers to the same question but was the first name in only 25 of them. The authors' advice is the basis of everything below: track the percentage of runs that mention you, and ignore “rank”.

Three things make the answers move:

• Search is sometimes on, sometimes off. OpenAI says ChatGPT “will automatically search the web if your question might benefit”; you can force it with the web icon or by typing a slash. A searched answer cites pages; an unsearched one comes from what the model learned in training.
• Memory and personalisation. ChatGPT stores saved memories and can draw on your chat history, and it uses neither in a temporary chat set to unpersonalised. It also takes a rough location from your IP address and may pass it to its search providers, so a check from Lahore and a check from Chicago are different checks.
• Sampling. Even with search off and memory off, the model writes each answer fresh. Repetition is the only way to see the pattern.

The 30-minute manual check

1. Write ten prompts in four shapes

Use the words customers use, not your category jargon. Cover the four ways people ask an assistant about a purchase:

Three recommendation prompts, two comparisons, three trust prompts and two alternatives prompts make a good ten.

2. Open a temporary chat and switch personalisation off

Temporary chats stay out of your history and, in the unpersonalised mode, use no saved memories and no custom instructions. OpenAI notes that it may keep a copy of a temporary chat for up to 30 days for safety purposes, which does not affect the result. If you want to check what a customer in another country sees, you need a connection from that country; a VPN is the simple way.

3. Run each prompt five times, with search on

Start a fresh temporary chat for each run, so nothing carries over. Force web search for at least three of the five runs; let the other two go wherever ChatGPT decides, which is closer to what a casual user gets. Fifty runs sounds like a lot. It is about 25 minutes.

4. Log every run

A spreadsheet with one row per run is enough. The columns that matter:

5. Turn the log into three numbers

• Visibility: mentioned runs divided by all runs. Peec AI, one of the tracking tools, defines it the same way, “how often a brand is mentioned in AI answers”: named in 4 of 10 chats is 40%.
• Share of voice: your mentions divided by all brand mentions in the same runs. It tells you whether you are one of three names or one of twelve.
• Source list: the cited URLs, grouped by domain. This is the part most people skip and the part that tells you what to do next.

Reading the citations

When ChatGPT searches, the answer carries inline citations and a Sources panel. Those pages are, for that answer, the sum of what ChatGPT knows about you. Three patterns come up again and again:

• It cites your own site for facts and someone else for judgement. Normal. Semrush's 2025 study of AI answers found that official company sites rarely appear in the top source lists; reviews, communities and news do. If the judgement sources are thin or hostile, that is the work.
• It cites an old page. A 2023 press release, a stale directory entry, a product you no longer sell. Update the page or publish a current one on a site the assistant already reads.
• It cites nothing about you and names three competitors. Either your site is blocked from OpenAI's search crawler or there is nothing independent for it to read. Check robots.txt for OAI-SearchBot first; it takes a minute and is the most common cause we see.

Which domains get cited shifts fast. Semrush tracked 230,000 prompts between July and October 2025 and saw Reddit cited in close to 60% of ChatGPT answers in early August and about 10% by mid-September. Profound's analysis of 680 million citations put Wikipedia at 7.8% of ChatGPT's, Reddit at 1.8%, Forbes and G2 at 1.1% each. The lesson is not to chase one domain; it is to be described accurately on several.

The same check on the other assistants

Customers do not only use ChatGPT. Each assistant has its own sources view and its own personalisation trap:

One free report deserves a line of its own. In February 2026 Bing added an AI Performance report to Bing Webmaster Tools with total citations, cited pages and the queries that triggered them, across Copilot and Bing's AI features. Bing says it is a sample, not a full count, but it is the only first-party citation data any assistant publishes, and it costs nothing.

Measure the traffic too

Visibility is the input; visits are the output. Links ChatGPT sends carry utm_source=chatgpt.com, so they appear as a campaign source in GA4 and most analytics tools. Perplexity, Copilot and Gemini show up as ordinary referrals from their domains. Google's AI features do not get a separate line anywhere: the clicks sit inside normal web search traffic in Search Console. Keep expectations in proportion; for most sites AI referrals are a small, fast-growing share, not a replacement for search.

Tools that do the runs for you

Trackers run your prompts on a schedule across several assistants and chart visibility, share of voice and sources. Prices below are from each vendor's pricing page on 23 October 2026, monthly unless noted:

Buy a tool when you track more than one brand, more than one country, or more than about 20 prompts; until then the spreadsheet is faster and you learn more from reading the answers yourself. Whatever you use, keep the prompt set fixed. Changing the questions changes the score.

What to do with the result

1. Visibility near zero. Confirm OpenAI's search crawler can reach you. OpenAI's crawler documentation says sites that opt out of OAI-SearchBot are not shown in search answers, that GPTBot (training) is a separate opt-out, and that robots.txt changes take about a day to apply. Then check what independent pages exist about you; if the answer is none, that is the next job.
2. Mentioned but never recommended. The assistant knows you exist and has no reason to pick you. Reviews, comparisons and roundups on sites it already cites are what move this; the batch of tactics is in how to get ChatGPT to recommend your business.
3. Mentioned with wrong facts. Find the cited page that carries the error and fix it, or publish a current, plainly written page on a site the assistant reads. A press release on a news domain is one way to put a dated, factual description where assistants look; how press releases get brands cited by AI covers the format.
4. Mentioned with a warning. The trust prompts surfaced a complaint page or a scam-alert site. Answer it where it lives, then build the positive sources around it; assistants weigh several sources, so one page rarely decides the answer for long.