How to track brand mentions in ChatGPT and other AI assistants
AI answers change every time you ask, so one screenshot proves nothing. Here is a repeatable method for tracking whether ChatGPT, Perplexity, Gemini, Claude and Google's AI name your brand, how often, against which rivals and from which sources.
The seoscore team 10 min read
On this page
- Why tracking AI mentions is not like tracking Google rankings
- Step 1: build a prompt set from questions your customers really ask
- Step 2: run the prompts the same way every time
- Step 3: record what each answer says about you
- Step 4: turn answers into metrics
- Tracking Google AI Overviews and AI Mode citations
- Measuring AI referral traffic in GA4
- Reading the cited sources: where your mentions really come from
- What to do if AI assistants don't mention you
- How often to check, and what counts as real change
- When manual tracking stops working
- Track AI mentions with seoscore
- Next steps
To track brand mentions in ChatGPT and other AI assistants, ask them the questions your customers ask, in the same controlled way, several times each, and record whether you are named, where in the answer, how accurately and which sources are cited. Answers change from one run to the next, so the number worth trusting is a mention rate across many prompts and runs, not a single screenshot. Add the signals you can measure directly, such as AI referral traffic in GA4 and citations in Google's AI Overviews, and you have a picture you can repeat every month.
Why tracking AI mentions is not like tracking Google rankings
GEO (generative engine optimization) is about whether AI assistants name and cite your brand in their answers. Measuring it is harder than measuring a Google ranking, because there is no fixed position to check.
The clearest evidence comes from a study by SparkToro and Gumshoe.ai. Around 600 volunteers ran 12 identical prompts through ChatGPT, Claude and Google's AI (AI Overviews and AI Mode) about 2,961 times. The authors found there is less than a 1 in 100 chance that ChatGPT or Google's AI gives the same list of brands in any two answers, and the same list in the same order came up fewer than 1 in 1,000 times. Answers also differed in how many brands they listed. The study did not test Perplexity or the Gemini app, and it has not been peer reviewed, but its direction matches what anyone who has asked the same question twice will have seen.
Other things move the answer too:
- Memory and personalisation. A logged-in user with saved memories or past chats can get a different answer from a new user.
- Location and language. Recommendations for local or regional services shift with the country.
- Web search. When an assistant searches the web, it leans on pages that rank well now. When it answers from what it learned in training, it leans on brands that were well known when the model was built.
- Different models. ChatGPT, Claude, Gemini and Perplexity are separate systems with separate sources. Doing well in one says little about the others.
This is a different question from AEO (answer engine optimization), which asks whether AI engines can read, trust and quote a given page. AEO is the page side; GEO tracking measures the outcome. Our GEO guide and the article on how to do generative engine optimization cover the full strategy. This article is only about measuring it properly.
Step 1: build a prompt set from questions your customers really ask
Write 20 to 50 prompts. Fewer than 20 and one odd answer skews the result; more than 50 is hard to run by hand.
Source them from real language, not keyword lists:
- Sales calls and demo notes: how prospects describe their problem before they know your name.
- Support tickets and live chat: the questions people ask after buying, which others ask before buying.
- People Also Ask boxes in Google for your main topics.
- Keyword research: seoscore's keyword research turns a seed keyword into related ideas with search volume, which shows which questions are common enough to track.
Write each prompt the way a customer would type it into a chat, in full sentences with their constraints (budget, size, location). Group them by intent:
| Intent | Example prompt | Funnel stage |
|---|---|---|
| Category | What is the best accounting software for a UK sole trader? | Consideration |
| Comparison | Is Brand X or Brand Y better for a small team? | Decision |
| Problem | How do I stop my invoices being paid late? | Awareness |
| Alternatives | What are good alternatives to Brand Y? | Decision |
| Branded | Is Brand X any good? What does it cost? | Decision |
Keep branded prompts in their own group. An assistant will almost always mention you when your name is in the question, so mixing them in inflates your mention rate. Use branded prompts to check accuracy instead.
Step 2: run the prompts the same way every time
The aim is that a change in your numbers reflects a change in the assistants, not in how you asked.
- Use a clean account or no account. A dedicated test account with memory turned off, or a logged-out session, avoids answers shaped by your own history. In ChatGPT, a Temporary Chat does not use saved memories.
- Start a fresh chat for every prompt. Earlier messages in a conversation change later answers.
- Fix the country. Run from the same location every time, or note it if you track several markets.
- Record whether the answer used web search. Where you can switch search on or off, keep it the same each time, and log answers that cite live sources separately from those that don't.
- Use the same models. Note the model name, because a model update can shift results overnight.
Then run each prompt several times. A single run is a sample of one. As a rough guide from basic statistics, if your true mention rate is around 30%, 50 answers give a margin of error of about plus or minus 13 points, 100 answers about 9 points and 150 answers about 7 points. So 30 prompts run five times each, giving 150 answers per assistant, is a sensible minimum for a number you can compare month to month.
Step 3: record what each answer says about you
One row per answer. A spreadsheet with these columns is enough to start:
date, assistant, model, web_search, prompt_id, run, mentioned, position, recommended_or_listed, facts_accurate, sentiment, competitors_named, cited_urls, notes
- mentioned: yes or no.
- position: 1 if you were the first brand named, 2 if second, and so on.
- recommended_or_listed: were you the suggestion, or one name in a list of ten?
- facts_accurate: were pricing, features, location and ownership correct? Note anything wrong.
- sentiment: positive, neutral or negative.
- competitors_named: every other brand in the answer.
- cited_urls: every source the assistant linked to.
Step 4: turn answers into metrics
From the sheet, calculate four numbers per assistant:
- Mention rate = answers that name you ÷ total answers.
- Average position when mentioned = mean of the position column, counting only answers that name you.
- Citation rate = answers that cite a URL on your domain ÷ answers that cite any source.
- AI share of voice = your mentions ÷ mentions of you plus your named competitors.
Treat position with care. SparkToro's conclusion is that visibility percentage across many prompts run multiple times is a reasonable metric, and Rand Fishkin has said that any tool offering a fixed ranking position in AI is not to be trusted. An average position across 150 answers tells you something; a single "you are number 3 in ChatGPT" does not.
Here is an illustrative example: 30 unbranded prompts, five runs each, 150 answers per assistant, for a brand and three rivals.
| Brand | ChatGPT | Perplexity | Gemini | Claude | Share of voice |
|---|---|---|---|---|---|
| You | 30% | 44% | 24% | 20% | 21% |
| Rival A | 60% | 50% | 64% | 56% | 40% |
| Rival B | 40% | 30% | 36% | 44% | 26% |
| Rival C | 20% | 26% | 16% | 12% | 13% |
Share of voice here pools all four assistants: you were named 177 times out of 858 brand mentions in total. With 150 answers per assistant, the 14-point gap between Perplexity and ChatGPT is likely to be real, and worth investigating through the sources Perplexity cites. The 4-point gap between Gemini and Claude is within the noise.
Tracking Google AI Overviews and AI Mode citations
AI Overviews are the AI summaries at the top of some Google results; AI Mode is Google's chat-style search. For both, the questions are whether an AI answer appears for your keywords and whether your site is cited in it.
To check by hand, search your target keywords in a private window from the country you care about and note, for each: AI Overview shown or not, your site cited or not, and which sites were cited. Repeat on a schedule, because both the appearance of an Overview and its sources change by query, country and day.
Search Console has long counted clicks and impressions from these AI features inside its ordinary web search totals rather than as a separate report. Check the Performance report help page for what it breaks out for your property today, and don't assume a rise in impressions came from AI unless the report says so.
If you track many keywords, seoscore's rank tracking checks your Google position for the keywords and country you choose, weekly or daily, and records whether Google showed an AI Overview and cited your site. It is on Starter and up, with a free trial.
Measuring AI referral traffic in GA4
Some AI answers send visitors, and those visits show up in Google Analytics 4 as referrals. Look for sources such as chatgpt.com, chat.openai.com (older traffic), perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. ChatGPT often adds utm_source=chatgpt.com to links in its answers, which helps.
To group them, create a custom channel group in GA4 (Admin, then Data display, then Channel groups) and add a channel called AI assistants where the session source matches this regular expression:
.*(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai).*
Move the new channel above Referral so these sessions are not swallowed by it. An exploration filtered on the same pattern works if you prefer not to change channel groups.
Know the limits. Many AI answers send no click at all. Some apps pass no referrer, so the visit lands in Direct. Clicks from AI Overviews and AI Mode come from google.com and look like ordinary organic search. Referral traffic is a floor, not the full picture, which is why the mention tracking above matters.
Reading the cited sources: where your mentions really come from
The cited URLs column is the most useful one in your sheet. Pool every cited domain across your prompts and sort by how often it appears. Then label each by type: review sites, "best of" lists, forums and communities, news and trade publications, competitors' own pages, and your own site.
This shows you:
- Outreach targets. A list article cited in a third of the answers that names your rivals but not you is worth a pitch.
- Where wrong facts come from. If assistants quote an old price, find the cited page carrying it and get it corrected, starting with your own site, directory listings and profiles you control.
- Which of your pages earn citations, so you can keep them current.
What to do if AI assistants don't mention you
Work through this in order:
- Check the page side. The free AEO and GEO checker runs seoscore's AEO checks on any page: AI crawler access in robots.txt, content that loads without JavaScript, entity markup, FAQ markup and answer blocks. If AI crawlers are blocked, see how to allow AI crawlers in robots.txt. For the wider page work, read what answer engine optimization is.
- Get mentioned where the answers come from. Use the cited-source list above to pick review sites, lists and communities to appear on.
- Keep brand facts consistent. The same name, description, prices and locations everywhere you control.
The GEO guide covers each of these in more depth.
How often to check, and what counts as real change
Monthly suits most brands. Check weekly during a launch, a rebrand or a PR campaign, when you expect movement and want to see it early.
To tell real change from noise:
- Compare the same prompt set, run the same way, with the same number of runs.
- Ignore moves smaller than your margin of error. At 150 answers per assistant, a 5-point swing is not news.
- Look for direction across several months and several assistants, not one jump.
- Note model updates beside your numbers, so a sudden shift has an explanation.
When you report to a boss or client, lead with mention rate and share of voice against named rivals, per assistant, with the trend. Add the top cited sources and any wrong facts found. Leave out screenshots of single answers; they invite the wrong conclusions.
When manual tracking stops working
The method above is sound, but it does not scale. Thirty prompts, five runs and four assistants is 600 answers a month, each read and coded by hand. Add a second market or a few more rivals and it becomes a part-time job, and coding errors creep in.
A tool for this should ask several assistants your chosen questions, keep every answer and cited source so you can check its scoring, and score how often and how early you are named, in a way you can compare over time.
Track AI mentions with seoscore
seoscore's AI visibility does the work of Steps 2 to 4. You choose the customer questions; it asks ChatGPT, Claude, Gemini and Perplexity, scores how often and how early each names your site, and keeps every answer with the sources it cited. It is on every plan, including Free, so you can replace the spreadsheet with your existing prompt set. The AI visibility guide explains the set-up.
For citations, AI mentions reports where Google AI Overviews and ChatGPT cite your site: how often, which of your pages, the questions people asked and the other sites cited beside you. Run it on demand or monthly. It is on Starter and up, with a free trial, as is rank tracking for AI Overview citations by keyword.
Next steps
- Write 20 to 50 prompts from sales calls, support tickets and People Also Ask, with branded ones kept separate.
- Run each prompt at least five times per assistant under fixed conditions and log every answer.
- Calculate mention rate, average position, citation rate and share of voice against three rivals.
- Set up the AI assistants channel in GA4.
- Run your key pages through the AEO and GEO checker, then work on the sources assistants cite most.
- Repeat monthly and report the trend, not the snapshot.
Questions people ask
Can ChatGPT tell me whether it mentions my brand to other people?
No. ChatGPT has no record it can share of what it told other users, and asking it about your own brand only shows what it says to you in that chat. The only way to know is to sample answers to the questions customers ask, repeatedly, and measure the pattern.
Why does ChatGPT get facts about my company wrong?
It may be relying on what it learned in training, which can be out of date, or on a web page it found that carries old or incorrect details. Check which sources it cites, correct the facts on pages you control first, and keep your name, prices and locations the same everywhere they appear.
Is AI share of voice the same as share of voice in search?
The idea is the same, your share of the attention in your category, but the inputs differ. In search it is usually based on rankings and estimated clicks. In AI assistants it is based on how often you are named in answers compared with your competitors across a fixed set of prompts.
Do I need to track every AI assistant?
Start with the ones your customers use. ChatGPT, Gemini, Perplexity, Claude and Google's AI Overviews cover most of the ground for most businesses. Results differ a lot between them, so tracking only one can give a misleading picture.
Sources
- NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility - SparkToro sparktoro.com
- AI brand recommendations vary, marketers warned on tracking tools - Communicate Online communicateonline.me
- Explaining AI visibility ppc.land
- support.google.com/webmasters/answer/7576553 support.google.com
seoscore
Do AI assistants mention you?
Track how often ChatGPT, Claude and Gemini name your site, and what to change.
Get my free score