How to track brand mentions in AI search, honestly

What tracking brand mentions in AI search actually measures

To track brand mentions in AI search is to sample the answers that assistants such as ChatGPT, Google AI Overviews, Gemini, Perplexity, Claude and Microsoft Copilot give to the questions your buyers ask, and record how often your brand appears, how it is described and which sources are cited. The output is a rate across many runs, never a rank.

Four separate things get called a “mention”, and a tracking sheet that blurs them produces a number nobody can act on. Record each in its own column.

Signal	What it means	Why it is tracked separately
Mention	Your brand name appears in the answer text	The broadest signal. A mention can be neutral, negative or wrong
Recommendation	The answer suggests your product as an option for the question	The commercially valuable subset of mentions
Citation	A link to one of your pages appears as a source	Shows which page did the work. You can be cited without being named, and named without being cited
Accuracy	What the answer says about you is correct	A frequent, confident, wrong mention is worse than none

The distinction between mention and citation matters more than most tracking tools let on. A brand can be recommended because review sites and forum threads describe it, with none of its own pages cited, and a brand can have a blog post cited on a how-to question without its product being named. Those two situations call for different work, which is why the column split comes first.

Why AI answers cannot be rank-tracked

AI answers cannot be rank-tracked because the same question, asked repeatedly, returns a different set of brands in a different order almost every time. The stable quantity is how often a brand appears across many runs, not where it appears in any one of them.

The clearest public evidence is a study by Rand Fishkin of SparkToro and Patrick O’Donnell of Gumshoe.ai, published in January 2026. As Search Engine Journal’s summary of the SparkToro study reports, volunteers ran 12 brand-recommendation prompts 60 to 100 times each across ChatGPT, Claude and Google’s AI Overviews or AI Mode, 2,961 runs in all. ChatGPT and Google’s AI returned the same list of brands in fewer than 1% of repeated runs, and the chance of the same list in the same order was under 0.1%.

The same study found the useful counterpart: visibility percentage held up better than position. Across 994 headphone responses, four brands each appeared in 55% to 77% of answers. Search Engine Land’s write-up draws the conclusion in six words: “Repeat presence means something. Exact rank does not.” Two cautions belong with it. The report is not peer-reviewed, and Gumshoe.ai sells AI tracking tools, which Fishkin disclosed.

Phrasing

“Best CRM for startups” and “what CRM should a startup use” are the same question to a person and can retrieve different sources. In the SparkToro study, 142 people writing their own prompt for the same need produced prompts with a semantic similarity score of 0.081. Any measurement is a measurement of the phrasings you chose.

Time and model version

Retrieval indexes update and models are replaced, often without notice. A drop between two months may be your content or a version change you were never told about.

Account context

OpenAI’s help page on memory says that when memory is on, ChatGPT can use relevant details from it to write more useful search queries. The same question inside a personalised account is not the same query as in a clean session.

How to build the prompt set

The prompt set is the measurement: because nobody can sample every question, the questions you choose decide what the numbers mean. Build it from questions buyers actually ask, not from the keywords you most want to win, and freeze it once written.

Where to find real buyer questions

• Search Console queries. Filter the Performance report for queries of six words or more, or those starting with how, what, which, best or vs. They are the closest thing to conversational phrasing you can get with real volume behind it.
• Sales and support transcripts. The questions prospects ask on discovery calls, in their own words. These are the phrasings no keyword tool shows.
• People Also Ask and forum threads. Reddit, Stack Exchange and community forums in your category, which show how people phrase the problem before they know the product category exists.
• Competitor comparison pages. Every “X alternatives” page a competitor publishes is a list of prompts they expect to be asked.

What goes in a 25-prompt set

Prompt type	How many	Example shape	What it tells you
Category recommendation	8	What is the best tool for X for a team of 20	Whether you are in the consideration set. High value, slow to move
Problem or job	7	How do I stop X from happening	Whether your content is cited before the buyer knows your category
Comparison	5	X vs Y for small teams	Clear signal: you appear or you do not, and how you are framed
Alternatives	3	Alternatives to [competitor]	Whether you are named when a rival is the starting point
Brand control	2	What is [your brand] and what does it cost	Accuracy. If this fails, nothing else on the sheet matters

Proportions are a starting point, not a rule. Weight toward the types your sales team hears most

Rules that keep the set honest

1. Never put your brand name in a non-control prompt. Naming the product guarantees it appears and measures nothing.
2. Write each prompt the way a buyer would type it, including the context they would give (team size, budget, industry).
3. Tag every prompt with its type and topic, so later you can see whether a decline is concentrated in one topic.
4. Freeze the wording. If a prompt has to change, retire it and add the new one as a new row, so the old trend line stays comparable.
5. Review the set once a quarter, not once a month.

How many runs and platforms you need

The number of runs you need depends on how small a change you want to detect, because an appearance rate measured on a few runs is mostly noise. The arithmetic is ordinary binomial sampling, and it is worth doing once before choosing a cadence or buying a tool.

For a rate p measured over n runs, the standard error is roughly the square root of p(1 minus p) divided by n, and a 95% margin is about twice that. Runs of the same prompt are not fully independent and prompts differ in difficulty, so treat the result as a floor on the noise, not an exact figure.

Prompts and runs both add responses, but they do not add the same thing. Some guides start much larger: Alex Birkett’s guide to AI share of voice suggests 100 to 200 prompts to begin with. For the same total, more distinct prompts usually beat more repeats, because repeats of one prompt are correlated with each other and a wider set covers more of what buyers ask. Repeats still matter for reading any single prompt; three or more per period is a sensible floor.

Worked example (hypothetical)

Take a hypothetical brand that appears in 1 of 4 runs of a single prompt, a 25% rate. The standard error is the square root of (0.25 times 0.75 divided by 4), which is about 0.22, so the 95% margin (1.96 times that) is roughly plus or minus 42 points. A single prompt checked four times tells you almost nothing.

Now pool the whole set. The same hypothetical brand runs 25 non-control prompts on 3 platforms, 4 times each: 300 responses. It appears in 84 of them, a 28% appearance rate. The standard error is the square root of (0.28 times 0.72 divided by 300), about 0.026, so the margin is roughly plus or minus 5 points.

Next month the hypothetical rate is 33%. Is that real? The standard error of the difference between two such months is the square root of (0.000672 plus 0.000737), about 0.038, so a change needs to exceed roughly 7.4 points (1.96 times 3.8) to clear the 95% bar. A 5-point rise is not yet a finding. If it holds or grows over the next two runs, it becomes one.

Responses per period	Typical setup	Approx. 95% margin at a 30% rate	Month-on-month change needed
40	20 prompts, 1 platform, 2 runs	plus or minus 14 points	about 20 points
120	20 prompts, 2 platforms, 3 runs	plus or minus 8 points	about 12 points
300	25 prompts, 3 platforms, 4 runs	plus or minus 5 points	about 7 points
1,000	50 prompts, 4 platforms, 5 runs	plus or minus 3 points	about 4 points

Binomial approximation, calculated for this post. Treat as a lower bound on noise

The decision rule: choose the smallest setup whose “change needed” column is smaller than the change you expect your work to produce. If you expect a content programme to move you a few points a quarter, 40 responses a month cannot see it, and the honest options are more runs or a tool that does the runs for you. The column is the bar a change has to clear; to be fairly sure of catching a real change of exactly that size, four times in five, you need roughly twice the responses, which is what the last line of the calculator below works out.

Check your own numbers

The calculator below applies the same arithmetic to your own counts. It works in percentage points: the rate is p = mentions / responses * 100; the margin is 1.96 * sqrt(p * (100 - p) / n); the change needed between two periods is 1.96 * sqrt(p1 * (100 - p1) / n1 + p2 * (100 - p2) / n2); and the responses needed per period to detect a change of d points is the standard two-proportion sample size, 2 * (1.96 + 0.84)^2 * p * (100 - p) / d^2, using last period’s rate for both periods: 95% confidence with 80% power, so a real change of that size shows up as significant in about four comparisons out of five. The signal-to-noise figure is the change divided by its standard error; above 1.96, the change clears the 95% bar.

Enter the response counts from your log for two periods.

Responses last periodResponses mentioning you, last periodResponses this periodResponses mentioning you, this periodSmallest change you want to detectpoints

Appearance rate, last period

28.0%

Appearance rate, this period

33.0%

95% margin this period, plus or minus (points)

5.3

Change (points)

5.0

Change needed to be real at 95% (points)

7.4

Signal-to-noise (z)

1.33

Above 1.96: a real change. Below: wait for two more runs.

Responses needed per period to detect your change

1,266

95% confidence, 80% power, based on last period’s rate

Hypothetical defaults, matching the worked example above. Change any number. Normal approximation: treat the result as a floor on the noise, because runs of the same prompt are not fully independent.

Which platforms to include

Include a platform when your buyers use it and its answers can differ from the others. For most B2B SaaS brands that means ChatGPT and Google (AI Overviews and AI Mode) at minimum, then Perplexity, Gemini, Copilot or Claude depending on the audience. Report each platform separately before pooling, because a gain on one and a loss on another can cancel out into a flat total that hides both.

How to run brand mention checks by hand

Running brand mention checks by hand means executing the frozen prompt set in clean sessions on a fixed schedule and logging every answer in the same template. It is slower than a tool and has one advantage no dashboard matches: you read the raw answers and see what was cited instead of you.

The procedure

1. Open a clean session for each platform. In ChatGPT, use a Temporary Chat with personalisation off: OpenAI’s Temporary Chat FAQ says unpersonalised temporary chats do not use memory or custom instructions. ChatGPT search is also available to people who are not signed in, which is another clean option.
2. For Google AI Overviews and AI Mode, use a private window, signed out, and note the country and device. Run every check from the same location each time.
3. Paste each prompt exactly as written. Do not add follow-up turns unless the follow-up is itself part of the frozen set.
4. Repeat each prompt the planned number of times, in a fresh session each run.
5. Log the result in the template below before moving to the next prompt. Logging later from memory is where errors creep in.
6. Save the full answer text or a screenshot. You will want it when a number moves.

The logging template

Each row of the tracking sheet is one response. These columns are enough to calculate every metric in the next section:

• date, platform, model or mode (as shown in the interface)
• prompt_id, prompt_type, topic
• run_number
• search_ran (yes if the answer shows linked sources)
• mentioned (yes/no), recommended (yes/no), position (order in the list, if any)
• cited_own_url (the URL, or blank)
• competitors_mentioned (comma-separated)
• sources_cited (every URL the answer cited)
• accuracy (correct, partly wrong, wrong, not applicable) and a note on what was wrong

The sources_cited column is the valuable one. Over a few months it becomes a ranked list of the pages and domains that assistants trust for your topics, which is a better brief for content and outreach than any keyword report.

What it costs in time

Assume, hypothetically, about a minute per response to run, read and log. The 120-response setup in the table above is then around two hours a month, which delegates well once the prompt set exists. Beyond a few hundred responses a month the manual method stops being practical, and that is the point where paying for automation is justified.

How to calculate AI share of voice and the other metrics

AI share of voice is your brand’s mentions divided by the total mentions of your brand and a fixed list of competitors, across the same responses. It is one of four metrics worth calculating from the log, and each answers a different question.

Metric	Formula	Question it answers
Appearance rate	Responses mentioning you / all non-control responses	How often are we in the answer at all?
Share of voice	Your mentions / (your mentions + each tracked competitor’s mentions)	How much of the conversation do we hold against named rivals?
Citation rate	Responses citing one of your URLs / responses where a search ran	When the assistant reads the web, does it read us?
Accuracy rate	Brand-control responses marked correct / all brand-control responses	Is what it says about us true?

Metrics to treat with suspicion

Average position inside an AI answer is the most misleading number in this field. The SparkToro study found order repeated in fewer than 0.1% of runs, so an average position mostly measures which runs happened to be sampled. Sentiment scores have a similar problem: unless the tool publishes its scoring method, a sentiment figure is an opinion with a decimal point. Report both, if at all, as secondary.

What Search Console, Bing and GA4 show about AI search

Search Console, Bing Webmaster Tools and GA4 each show part of your AI search footprint without any sampling of your own. None of them shows mentions without a click or a citation, so they complement a prompt set rather than replace it.

Report	What it shows	What it cannot show
Google Search Console, Performance (Web)	Clicks and impressions from AI Overviews and AI Mode, folded into the Web search type	Which of those clicks came from AI features
Google Search Console, Generative AI report (beta)	Impressions in AI Overviews and AI Mode by page, country, device and date	Clicks, click-through rate and the queries that triggered the appearance
Bing Webmaster Tools, AI Performance	Citations in Copilot and Bing AI summaries, cited pages, a sample of grounding queries	Clicks, and any platform outside Microsoft’s
GA4	Sessions referred from assistants, now grouped in an AI Assistant channel	Answers that were read without a click, which is most of them

Google’s documentation on AI features and your website states that sites appearing in AI Overviews and AI Mode are included in the overall search traffic in Search Console, reported under the Web search type. That means you can see the effect on clicks and click-through rate, but not isolate it. In June 2026 Google added a beta Generative AI performance report which, as Search Engine Roundtable reported, shows impressions in AI features by page, country, device and date, with no click or query data yet. Google said on 31 August 2026 that the reports had rolled out to all websites worldwide, as Search Engine Journal reported. It tells you which pages appear in AI features, not what that appearance is worth. For clicks, the practical workaround is to watch click-through rate on the queries you know trigger an AI Overview, as covered in the guide to AI Overviews and what they do to your clicks.

Microsoft introduced AI Performance in Bing Webmaster Tools in February 2026. It reports total citations, the average number of your pages cited per day, page-level citation counts and grounding queries, which Microsoft describes as the key phrases its AI used when retrieving your content and notes are a sample. Unlike Search Console and GA4, it reports citations rather than clicks, which makes it worth checking monthly alongside the prompt set.

In GA4, Search Engine Journal reported in May 2026 that Google added an AI Assistant channel to the default channel group, naming ChatGPT, Gemini and Claude as examples of recognised referrers. The list has since changed: a July 2026 Search Engine Journal analysis found the documentation naming ChatGPT, Gemini, DeepSeek, Copilot and Grok, with Claude removed and Perplexity still landing in Referral, so the default channel undercounts. ChatGPT also tags its outbound links: OpenAI’s publisher FAQ says it automatically adds utm_source=chatgpt.com to referral URLs. If you need a custom channel for older data or for referrers the default list misses, a session source regex such as chatgpt\.com|perplexity\.ai|gemini\.google\.com|claude\.ai|copilot\.microsoft\.com is a reasonable start; check your own referral report for the exact hostnames before relying on it.

Treat assistant referral sessions as a floor. Most AI answers are read without a click, so the count is real but always partial, and it says nothing about the buyers who read an answer, remembered the name and came back later through search or direct.

How to evaluate AI visibility tracking tools

AI visibility tracking tools run a prompt set against assistants on a schedule and report how often a brand appears, which is the same method as the manual version, automated. What you pay for is volume of runs, storage and presentation, so the evaluation is about whether the sampling is transparent and large enough.

Two of the established SEO suites show the range. Ahrefs says its Brand Radar models prompts from real searches in its keyword database and runs them across AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini and Copilot, with Claude available for custom prompts, and offers an Estimated Impressions figure weighted by the search volume behind each prompt. Semrush’s AI Visibility Toolkit documentation describes daily prompt tracking for Google AI Mode and ChatGPT, starting at 25 tracked prompts on the base plan. Dedicated trackers exist in growing numbers; the questions below apply to all of them. Google’s own AI optimization guide adds a caution worth keeping in mind: “No third-party tool has access to our internal ranking or AI systems”. Every tracker is sampling answers from outside, as the manual method does.

Manual prompt panel

Strengths

• Full answer text and every cited URL in front of you
• Free apart from time
• You control the prompts, sessions and run count

Limits

• Roughly a minute per response, so impractical beyond a few hundred a month
• Easy to drift if the procedure is not written down

Start here, for at least one month

AI visibility tracker

Strengths

• Hundreds or thousands of runs without staff time
• History, competitor views and alerts

Limits

• Sampling method often undisclosed
• Volume-weighted figures are estimates of interest, not counts of people

Buy once the run count you need outgrows a spreadsheet

First-party reports

Strengths

• No sampling of your own: Search Console, Bing AI Performance and GA4
• Real clicks, impressions or citations

Limits

• Nothing on mentions without a click or citation
• Google’s Generative AI report has no clicks or queries yet

Always on, alongside either of the others

Questions to ask any vendor

• Which prompts are sampled, where do they come from, and can I add and freeze my own?
• Which platforms, and which model or mode, for each one? Is it the logged-out consumer interface or an API?
• How many runs per prompt per period? If the answer is one, the score is a single sample presented as a metric.
• Can I export response-level data, including the full answer text and every cited URL?
• How are brand mentions detected? Does it handle abbreviations, misspellings and product names?
• How is any “impressions” or volume figure calculated? A prompt weighted by keyword volume is an estimate of interest in the topic, not a count of people who saw your brand.

When to buy, and when not to

The decision comes down to three questions, in order. Use the calculator above to find the run count your expected change needs, then answer them.

Have you run at least one full month of the prompt set by hand?

No

Do not buy yet. Run one month by hand to learn what the answers look like and which competitors and sources appear.

Yes

Does the change you want to detect need more than a few hundred responses a month, or more than three platforms?

No

Stay manual. At that volume the spreadsheet is enough, and it is more transparent than any dashboard.

Yes

Can the vendor export response-level data, including the full answer text and every cited URL?

Yes

Buy. The run count is what makes the numbers usable, and the export lets you check them.

No

Do not buy that tool. You cannot check a number you cannot see; ask the next vendor the same question.

How to read the results without fooling yourself

Reading AI visibility results well means separating noise from change, and change caused by you from change caused by the platform. Most bad decisions in this area come from reacting to a single run or assuming a broad drop was self-inflicted.

Pattern	Likely reading	Action
One prompt drops out, returns next run	Noise	Nothing
Change smaller than the change needed (see the sample table)	Noise until repeated	Wait for two more runs
Several prompts on one topic decline over three runs	Content or competitive change on that topic	Read the cited-instead column for those prompts and act on that topic
Every prompt declines at once, competitors too	Platform or model change	Annotate the date and do not rewrite anything yet
Every prompt declines at once, competitors flat or up	Something specific to you	Check crawl access, recent site changes and brand-control answers
Brand-control answer becomes wrong	Accuracy problem	Act immediately, as below

Keep a change log next to the data: your own publishing and site changes, and any announced model or product changes on each platform. Without it, every movement is ambiguous.

The wider attribution problem behind all of this, and why last-touch reporting undercounts anything that happens early in a buying journey, is covered in measuring organic when attribution is already broken.

How to check what AI says about your brand

Checking what AI says about your brand means running brand-control prompts about your pricing, capabilities and fit every month and scoring each answer for accuracy. A wrong answer about your own product is repeated to everyone who asks, which makes it the highest-priority line on the tracking sheet.

1. Write three or four brand-control prompts: what the product is and who it is for, how pricing works, what it integrates with, and what it does not do.
2. Run them in clean sessions on every tracked platform.
3. Score each answer: correct, partly wrong or wrong, and quote the wrong sentence.
4. If a search ran, note which source the wrong claim came from. That source is usually what needs correcting.
5. If no search ran, the claim comes from the model’s training data, which changes only when the model does; the response is slower and indirect.

The difference between those last two cases, and what you can do about each, is the core of the guide to ChatGPT visibility and how it picks sources. The short version: searched answers are fixed by correcting the page being read, trained answers by making your product’s facts consistent wherever it is described.

Indirect signals worth tracking alongside

Indirect signals are measures that move when AI visibility changes but do not depend on catching an assistant in the act: branded search demand, web mentions of the brand and what buyers report. They are less precise and much less noisy than prompt sampling, so the two together are more trustworthy than either alone.

• Branded search impressions in Search Console. If people who did not know you start searching your name, something upstream created that demand. Filter the Performance report to queries containing your brand and watch the monthly trend.
• Web mentions of your brand. In Ahrefs’ May 2025 study of 75,000 brands in AI Overviews, branded web mentions had the strongest correlation with AI Overview brand visibility of the factors tested (Spearman 0.664), ahead of branded anchors (0.527), branded search volume (0.392) and Domain Rating (0.326). A December 2025 follow-up across ChatGPT, AI Mode and AI Overviews found YouTube mentions correlated slightly more strongly still. Ahrefs notes that correlation is not causation, but it makes web and video mention counts sensible leading indicators to track.
• Self-reported attribution. Add “AI assistant (ChatGPT, Gemini, etc.)” as an option to the “How did you hear about us?” field on demo and sign-up forms, and ask the same question on discovery calls. It has the weaknesses of any self-report and none of the sampling problems.
• Assistant referral sessions in GA4. Covered above. A floor, not a total.

Self-reported attribution is the one most teams skip and the one worth adding first. One extra option on a form produces, within a quarter, a record of how many pipeline opportunities name an assistant, which a finance team is far more likely to accept than a visibility score.

What to do when your brand is not appearing

When your brand is not appearing in AI answers, the tracking data should tell you why before you change anything: whether the right pages are retrievable, whether you have anything citable, and what the assistant cites instead. The fixes themselves belong to a separate guide; this section is about reading the diagnosis.

• Check access first. If ChatGPT never cites your URLs even when a search runs, confirm robots.txt does not block OAI-SearchBot, which OpenAI’s publisher FAQ says must be allowed for content to appear in ChatGPT’s summaries and snippets.
• Read the cited-instead column. The sources appearing in place of yours show what the system treats as a good answer to that prompt. If a forum thread wins, the question is being asked in a register your pages do not match. If a review site wins, the gap is third-party coverage, not content.
• Compare entities. Paste a page that is being cited and your nearest equivalent into the Entity Extractor to see which tools, standards and concepts the cited page names that yours does not.
• Check passage structure. Assistants retrieve passages rather than whole pages, as set out in what AI search actually retrieves from your page, so a correct answer buried in paragraph five of a section is often an answer that is never read.

The changes that follow from that diagnosis, from restructuring pages to earning third-party mentions, are set out in how to improve brand visibility in AI search. If you would rather have the tracking and the fixes run as one programme, that is what the AI search visibility service covers.

A first-quarter plan for AI brand mention tracking

A first-quarter plan for AI brand mention tracking builds the prompt set and baseline in week one, establishes the routine and first-party reports in month one, and reaches a trend you can act on by the end of quarter one. Most of the judgement is in the first week; after that the work is routine.

Week 1

Build the set and a baseline

Collect real buyer questions, write and tag 20 to 25 prompts, choose platforms and competitors, size the run count with the margin table, and run the first full baseline by hand.

Month 1

Connect the first-party data

Verify the site in Bing Webmaster Tools and check AI Performance. Confirm assistant referrals land in GA4’s AI Assistant channel. Add the AI option to the how-did-you-hear field. Run the second full set.

Quarter 1

Read the trend and decide

With three runs, read changes against the change-needed figure, review the cited-instead column by topic, fix any brand-control inaccuracies, and decide whether the run count you need justifies a tool.

Use the checklist below to see how much of that setup is in place. It covers the measurement only, not the content work it should lead to.

Prompt set

20 to 25 prompts written from real buyer questions No brand name in any prompt except the brand-control ones Every prompt tagged with type and topic Wording frozen; changed prompts retired and added as new rows

Sampling

Run count sized so the change needed is smaller than the change you expect Clean sessions on every platform (Temporary Chat, signed out, private window) Same location and device for every run Every platform reported separately before pooling

Logging

One row per response in a fixed template Cited sources recorded for every answer Full answer text or a screenshot saved Change log of your site changes and announced platform changes

First-party data

Site verified in Bing Webmaster Tools and AI Performance checked monthly Assistant referrals landing in GA4’s AI Assistant channel “AI assistant” option on the how-did-you-hear field Brand-control answers scored for accuracy every month

Ticks are saved in your browser.

Tracking is not the work, and it is easy to let it become the work. If three months of data have not changed what you publish, correct or pitch, either the prompt set is measuring the wrong questions or nothing needs changing yet. Both are worth knowing, and neither is a reason to add more dashboards.