---
title: "How to track brand mentions in AI search and ChatGPT"
description: "How to track brand mentions in AI search: which prompts, how many runs per engine, what to count and how long to watch, from TIN's studies and primary research."
url: "https://theinternetninja.com/guides/how-to-track-brand-in-ai-search/"
updated: "2026-10-16"
---

# How to track brand mentions in AI search and ChatGPT

By the TIN research desk. Updated 2026-10-16. A guide in AI search and visibility (https://theinternetninja.com/guides/ai-search-visibility/), The Internet Ninja.

To track your brand, run a fixed set of buyer questions several times a day on each engine, report each engine on its own, and read a rolling figure over two to four weeks, never a single check [source](https://arxiv.org/abs/2604.07585). In a study of four engines over a window of 45 to 46 days, the sources cited for the same prompt had a day-to-day Jaccard similarity of only 0.34 to 0.42 [source](https://arxiv.org/abs/2604.07585). Leave your name out of the questions [source](https://arxiv.org/abs/2601.00912), write three to five wordings of each [source](https://arxiv.org/abs/2607.14035), and keep the answers that cite nothing in the count [source](https://arxiv.org/abs/2607.14035).

| Decision | What to fix before the first run | Evidence |
|---|---|---|
| Prompts | Buyer questions without your name, chosen by a written rule, three to five wordings per need | Recognition vs discovery, [survey protocol](https://arxiv.org/abs/2607.14035) |
| Engines | Each engine named, each with its own baseline, never blended | [TIN citation sources study](https://theinternetninja.com/recon/ai-citation-sources-2026/), [Schulte et al.](https://arxiv.org/abs/2604.07585) |
| Country and account | Country, language, signed in or not, recorded on every run | [TIN citation sources study](https://theinternetninja.com/recon/ai-citation-sources-2026/), [survey protocol](https://arxiv.org/abs/2607.14035) |
| Runs | Several per prompt per day; a pilot to see how wide the interval is | [Schulte et al.](https://arxiv.org/abs/2604.07585), [survey protocol](https://arxiv.org/abs/2607.14035) |
| Metric and denominator | Mentioned, cited, or share of voice, each with its denominator stated; no-citation answers stay in | TIN Reddit post, [survey protocol](https://arxiv.org/abs/2607.14035) |
| Window | Two to four weeks, rolling | [Schulte et al.](https://arxiv.org/abs/2604.07585) |

## Why one check is not a measurement

Ask the same engine the same question on two days and you get a different set of sources. A team at the University of St. Gallen ran eight prompts in each of four verticals through ChatGPT, Gemini, Google AI Mode and Perplexity from 24 January to 20 March 2026 [source](https://arxiv.org/abs/2604.07585). "The day-to-day Jaccard similarity for cited sources averages between 0.34 and 0.42" [source](https://arxiv.org/abs/2604.07585). Brands held steadier than URLs: brand-level day-to-day stability ran at a Jaccard of 0.45 to 0.59, which the authors read as "brand presence aggregated over a campaign is a more reliable KPI than tracking individual cited URLs" [source](https://arxiv.org/abs/2604.07585).

A 2026 survey of the field states the consequence: "Visibility is a distribution. It depends on the engine, date, location, query formulation, whether search is actually activated, and stochasticity in generation. A point estimate is not a stable indicator" [source](https://arxiv.org/abs/2607.14035).

TIN's own studies say the same of themselves. The citation sources study's pre-registered method reads "Two weekly runs: a snapshot, not a trend", while the same page states that its first edition "is one weekly run (amendment 3), a snapshot, not a trend" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/), and the AI Overviews study prints "AI Overview answers vary between runs; this is a snapshot" [source](https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/).

There is a second trap before variance even starts. Asking an engine about your own name measures recognition, not discovery. A January 2026 preprint tested 112 startups across 2,240 queries: asked about a product by name, ChatGPT recognised it 99.4% of the time and Perplexity 94.3%; asked the buyer's question without the name, the products appeared 3.32% and 8.29% of the time [source](https://arxiv.org/abs/2601.00912). TIN's GEO post flags it as "one author, one preprint" and calls this "a structural point rather than a settled number" source. The question that pays is the one that leaves your name out.

## Write the method before you look

The cheapest protection against reading noise as progress is a rule written before the data exists. If the prompts, engines and window were fixed in advance, a good week cannot quietly become the new baseline and a bad prompt cannot be dropped because it looked bad.

Every TIN Recon study works this way. The citation sources study states that "the method, the category scheme and the audience layer were committed before the data they govern was read", and later changes are published as dated amendments [source](https://theinternetninja.com/recon/ai-citation-sources-2026/).

Your version is one page, dated before the first run:

- The prompt list, with the rule that produced it.
- The engines, and the mode for each (search on or off, if the engine exposes it).
- The country, language and account state.
- Runs per prompt per day.
- The metric, with its numerator and denominator written out.
- The window, and what size of change you will treat as real.

When you change any of it, add a dated line rather than editing the original.

## Step 1: choose the prompts

### Ask the buyer's question, not your name

The recognition result above decides the first rule. "What does [your brand] do" will come back positive almost every time and tells you nothing about whether a buyer who has never heard of you is shown your name. Write the questions your buyers type before they know you exist.

### Use real searched questions, chosen by a rule

Pick prompts by a rule you can write down, not by taste. TIN's citation sources study took "the 4 highest-volume questions in each of 15 everyday topics" from a US keyword database, "after removing brand, person, one-country and date-dependent questions by a rule fixed before selection" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). The St. Gallen team started from high-volume keywords entered into Google, and "the “People Also Ask” feature was subsequently used to identify and generate relevant prompts" [source](https://arxiv.org/abs/2604.07585). Either route gives a list someone else could rebuild.

### Several wordings per need

One wording measures the wording. Prompt-level agreement varied "from below 0.2 to above 0.8" within a single campaign, and the authors conclude that "monitoring based on one or two prompts will reflect the idiosyncrasies of those prompts rather than campaign-level visibility" [source](https://arxiv.org/abs/2604.07585). The survey's protocol asks for "three to five paraphrases per information need" [source](https://arxiv.org/abs/2607.14035). Group the wordings under the need they express and report by need, so a single odd phrasing cannot swing the total.

## Step 2: choose the engines and keep them apart

Engines do not share sources, so they should not share a score. In TIN's study of seven engines on the same 600 prompts, "1,091 of 1,395 cited sites (78%) appear in only one engine's answers" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). "The closest pair, AI Overviews and AI Mode, share 8 of their 25 most-cited sites; ChatGPT shares at most 1 with any other engine" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/).

They differ in volume too. The citations behind each engine in that study were "ChatGPT 299, Perplexity 1,099, Gemini 218, Copilot 523, Claude 47, AI Overviews 184, AI Mode 487" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). A blended share is mostly the high-volume engines' share.

The same split shows up when you count brands. Of 238 AI and automation agencies, 173 were never cited by ChatGPT, 140 never by Google AI Overviews, and 77 never by any platform [source](https://theinternetninja.com/recon/agency-ai-citations-2026/). A zero on one engine is not a zero on another. Those are counts for one industry on one index and one date, and the all-platform figure is not a ChatGPT figure.

The St. Gallen authors put the rule plainly: "Marketers should set engine-specific visibility baselines rather than applying a single threshold across all AI search products" [source](https://arxiv.org/abs/2604.07585).

### ChatGPT

Many ChatGPT answers do not search at all. In the St. Gallen configuration, "ChatGPT activates web search only for specific queries, leaving 57.8% of its runs with zero citations" [source](https://arxiv.org/abs/2604.07585). That is one study's set-up on Swiss prompts, not ChatGPT's general rate. Record, for every run, whether search ran. A brand named in an answer that never searched is a different signal from a brand named beside a list of sources.

For the checks specific to ChatGPT, see how to check whether ChatGPT cites you.

### Google AI Overviews and AI Mode

Track them as two surfaces, each with its own baseline. AI Overviews do not appear on every search, so the number of searches that showed one belongs in the report. In TIN's AI Overviews study, 113 of 151 usable searches had an AI Overview [source](https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/).

For the Search Console steps, see how to check whether AI Overviews cite you.

### Perplexity, Gemini, Copilot and Claude

Each needs its own baseline. Perplexity returned the most citations of the seven. Claude returned 47 in total, and none for Germany, Italy, Sweden or Japan [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). TIN's inference, not a measured finding: an engine that returns few citations needs more runs before its share means anything.

## Step 3: fix the country and the account

The same question gets a different answer in a different country. In TIN's study, "Japan (84%), Sweden (79%), Italy (70%) and Germany (65%) mostly get sites serving their own country; France (32%) and Spain (25%) do not", and in the English-speaking markets outside the US the local share was "the UK 30%, Australia 29%, India 8%" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). Track each market you sell into as its own series.

Know what your instrument means by country. In TIN's study, country is the country the index files the prompt under, not proof of where the asker was [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). In the AI Overviews study, "the source is a crawler's view of the results page, not a real user's personalised result" [source](https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/). Neither is what your buyer sees when signed in.

Record on every run what the survey lists: "prompts, raw responses, citations, search status, timestamps, locale, account type, user agent, and canonicalization rules" [source](https://arxiv.org/abs/2607.14035). In a spreadsheet that is one row per run, with columns for prompt, wording, engine, country, language, account state, date and time, run number, whether search ran, the full answer, the cited URLs, and whether your brand was named. Keep the raw answers. A metric you decide on later can only be computed from what you kept.

## Step 4: run each prompt enough times

The St. Gallen team measured how the standard error of a per-brand detection rate falls as runs per prompt rise, across 1,216 per-brand series [source](https://arxiv.org/abs/2604.07585).

| Runs per prompt | Standard error | 95% interval half-width |
|---|---|---|
| 1 | 0.370 | 0.724 |
| 5 | 0.123 | 0.241 |
| 7 | 0.081 | 0.158 |
| 8 | 0.062 | 0.121 |

Source: Schulte, Bleeker and Kaufmann, Table 16 [source](https://arxiv.org/abs/2604.07585). The paper notes that its value at 9 runs is subject to a finite population correction and underestimates the error.

One run is close to useless: its interval covers almost every possible rate. From that table the authors recommend "at least 7 runs per prompt per day for brand visibility monitoring, and at least 8 runs when source-level coverage matters" [source](https://arxiv.org/abs/2604.07585).

Do not treat seven as a law. The survey is direct about it: "This number is not a universal standard: it derives from a small universe of Swiss queries and at most ten repetitions. The appropriate practice is sequential precision analysis: repeat the measurement until the interval around the estimand is sufficiently narrow for the decision at hand" [source](https://arxiv.org/abs/2607.14035). In practice: run a pilot on a few prompts, see how much your own rate moves between runs, and set the run count from that. A decision between "present" and "absent" needs fewer runs than a ranking of five close competitors.

### Wait for a complete run

Read a run only once it has finished. TIN's citation sources study published only after its first weekly run "has stopped changing (checked 2026-10-01 and 2026-10-02: identical counts per engine)" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/).

## Step 5: decide what you count, and keep the empties

"Visibility" is several different quantities. Pick one, and write its denominator next to it.

| Metric | Numerator | Denominator |
|---|---|---|
| Mentioned | Answers that name your brand | All answers on the panel, including those that cite nothing |
| Cited | Answers whose sources include your URL | All answers on the panel, including those that cite nothing |
| Share of voice | Mentions of your brand | Mentions of all brands in the same answers |
| Passage used | Answers that drew on a specific block of your page | Answers that cite your URL |

These definitions are TIN's, written for this guide; no study above defines share of voice for you.

TIN's Reddit post found that "Reddit's pre-drop share is therefore 2.05%, 3.83% or 11.97% depending on whose instrument you read" source. It explains why: "Share of citations among responses that included at least one citation is a different quantity from share of responses that cite a domain at least once, and both differ from share across a fixed prompt panel. The units are rarely stated" source. Its conclusion is the rule for your own report source: "“Our ChatGPT citation share is 2%” is not a number. Whose panel, which prompts, share of what denominator."

Keep the answers that cite nothing. The survey: "Outputs without search, without citations, or with errors are outcomes, not data to be discarded", and "a dashboard cannot calculate a “share of citations” only among responses that contain citations and then interpret it as overall visibility" [source](https://arxiv.org/abs/2607.14035). With 57.8% of ChatGPT runs in the St. Gallen configuration citing nothing [source](https://arxiv.org/abs/2604.07585), dropping the empty answers changes the figure, not just the precision. The St. Gallen paper itself excludes zero-citation runs from its source-similarity analysis [source](https://arxiv.org/abs/2604.07585). That suits a comparison of source sets. It does not suit a figure for how visible a brand is.

TIN's studies show what keeping the empties looks like. In the citation sources study, Claude and the two Google surfaces returned nothing for some countries: "Those eight cells are left empty, never filled" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). In the AI Overviews study, 438 searches were run; 227 returned no stored result and 60 were more than 30 days old, "both are counted and left out, never replaced", leaving 151 usable searches [source](https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/).

Be exact about what "cited" means. TIN's retrieval post: "“Returned by an engine” is not “fetched” is not “opened” is not “surfaced in the sidebar,” and “cited” is not “was the lead source behind a citation”" source. Its advice for anyone already counting citations: "Stop measuring whether you were cited, and start measuring which block was cited" source. That is the last row of the table above, and the one that tells you what to change on the page.

### Report medians when comparing brands

Citation counts are skewed, so averages mislead. In TIN's agency study, "161 of 238 agencies were cited at least once by an AI platform, but the median agency was cited 3 times", and "the mean is 124.9 and the median 3: a handful of very large counts pull the average far above what a typical agency gets, which is why this page reports medians" [source](https://theinternetninja.com/recon/agency-ai-citations-2026/). When you set your brand against a competitor set, compare against the median.

## Step 6: read it over weeks, not days

Runs per day fix the noise within a day. The window fixes the noise between days. For a per-brand rate, the St. Gallen team found the standard error "drops below 0.10 at d = 10 days and below 0.05 at d = 24 days", and that "a 14-day window still leaves SE at 0.080 (± 0.157), sufficient for directional monitoring but not for fine-grained brand comparison" [source](https://arxiv.org/abs/2604.07585). Their recommendation: "Rolling aggregation over two to four weeks" [source](https://arxiv.org/abs/2604.07585).

So report a rolling two to four week figure, and decide in your method note how large a change has to be before you call it one. A move that sits inside the interval is not a result. TIN's GEO post puts the same point about vendor claims: "A vendor reporting a round multiplier from one run is reporting the weather" source. The survey adds that a measurement "must therefore vary along at least four dimensions: run, paraphrase, date, and engine" [source](https://arxiv.org/abs/2607.14035).

## Signals you already own

Your server logs carry one signal no panel does. OpenAI documents that "when users ask ChatGPT or a CustomGPT a question, it may visit a web page with a ChatGPT-User agent", and that "ChatGPT-User is not used for crawling the web in an automatic fashion" [source](https://platform.openai.com/docs/bots). OpenAI uses ChatGPT-User "for certain user actions in ChatGPT and Custom GPTs" [source](https://platform.openai.com/docs/bots), so a ChatGPT-User request in your logs means a user action led ChatGPT to fetch that page, not an automatic crawl. It does not say the page was cited or that the user saw your name.

The same document notes that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links" [source](https://platform.openai.com/docs/bots). Check that you have not blocked it before you start counting ChatGPT answers; otherwise you are measuring your own robots.txt.

## How to judge any tracking tool's numbers

TIN does not rank tracking tools, and this guide names none. If you already pay for one, ask it to state each decision in the method table at the top of this page. A dashboard that cannot is showing you a number without its unit.

## What this guide cannot tell you

- **The run and window numbers come from one study.** The St. Gallen paper used Swiss servers and locale, eight prompts per vertical, four verticals and one window of 45 to 46 days, and detected brands by "substring matching of a fixed lexicon" [source](https://arxiv.org/abs/2604.07585). It says itself that "results may not generalise to other regional or linguistic markets" [source](https://arxiv.org/abs/2604.07585). Its first author is also affiliated with Aurora Intelligence [source](https://arxiv.org/abs/2604.07585). The survey calls its run count a starting point, not a standard [source](https://arxiv.org/abs/2607.14035).
- **TIN's studies are snapshots of particular populations.** The citation sources study covers 60 everyday questions, "not of everything people ask AI. Business-to-business questions are not covered" [source](https://theinternetninja.com/recon/ai-citation-sources-2026/). The agency study covers AI and automation agencies listed in seven partner directories, and its citation counts come from one commercial index on one date, a sample of the answers that index tracks [source](https://theinternetninja.com/recon/agency-ai-citations-2026/). The AI Overviews study is Google only.
- **The metric table is TIN's inference, not a measured finding.** The definitions follow from the denominators the sources argue about; no study here tested which metric best predicts buyers reaching you.
- **Tracking is not getting cited.** Nothing here says what makes an engine name you.
- **TIN's own number.** TIN publishes its own AI visibility baseline: "our AI visibility baseline is 0%" source. This guide is a method, not a claim that TIN has solved the problem it measures.

**About this guide.** Written by TIN. Every figure links to the page or paper it was read from on 2026-10-16, and TIN's own studies publish their method and dataset. Last updated 2026-10-16.

## Read next

- [Which websites AI engines cite, TIN's study](https://theinternetninja.com/recon/ai-citation-sources-2026/): the worked example of fixed prompts, engines and countries, with the dataset.
- [Which agencies AI engines cite](https://theinternetninja.com/recon/agency-ai-citations-2026/): counting brand citations across a population, and why it reports medians.
- [Google top 10 against AI Overviews](https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/): how to count the searches where nothing came back.
- ChatGPT stopped citing Reddit: one share, three instruments, and why a share needs its unit.
- Generative engine optimization: the controlled evidence: why most reported results are noise.
- How ChatGPT ranks passages: what "cited" means, and which passage got used.
- How to get your company cited by ChatGPT: the ChatGPT checks this page only points to.
- How to show up in Google AI Overviews: the Search Console steps for AI Overviews.
- AI search optimization, the hub: where tracking fits among the other guides.

**Measure your market.** If you want this method run on your own market's buyer questions, [tell us what you sell](https://theinternetninja.com/about/#work-with-us).

## Sources

1. Schulte, Bleeker and Kaufmann, "Don't Measure Once: Measuring Visibility in AI Search (GEO)", arXiv 2604.07585, 8 April 2026. https://arxiv.org/abs/2604.07585
2. Martinez, "Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)", arXiv 2607.14035, 15 July 2026. https://arxiv.org/abs/2607.14035
3. Amit Prakash Sharma, "The Discovery Gap: How Product Hunt Startups Vanish in LLM Organic Discovery Queries", preprint, January 2026. https://arxiv.org/abs/2601.00912
4. OpenAI, "Overview of OpenAI Crawlers", read 2026-10-16. https://platform.openai.com/docs/bots
5. TIN, "AI citation sources" study, edition 2026-10-02. https://theinternetninja.com/recon/ai-citation-sources-2026/
6. TIN, "Agency AI citations" study, edition 2026-09-27. https://theinternetninja.com/recon/agency-ai-citations-2026/
7. TIN, "Google top 10 vs AI Overviews" study, edition 2026-10-02. https://theinternetninja.com/recon/google-top10-vs-ai-overviews-2026/
8. TIN, "ChatGPT stopped citing Reddit". https://theinternetninja.com/blog/chatgpt-stopped-citing-reddit-substitute-platforms-2026/
9. TIN, "Generative engine optimization: what the controlled evidence shows". https://theinternetninja.com/blog/generative-engine-optimization-what-the-controlled-evidence-shows-2026/
10. TIN, "ChatGPT ranks passages, not pages". https://theinternetninja.com/blog/ai-search-optimization-what-the-retrieval-record-shows-2026/

## Questions

**How do you measure brand visibility in AI search?**

Write down the method before the first run: the buyer questions, three to five wordings of each, the engines, the country and account, the runs per prompt, the metric with its denominator, and the window. Ask questions that do not name you, keep each engine's figure separate, keep the answers that cite nothing in the denominator, and read the result over two to four weeks rather than from one check.

**How can I track my brand mentions across AI platforms like ChatGPT and Perplexity?**

Run the same fixed prompts on each platform and report each one on its own. In TIN's study of seven engines, 1,091 of 1,395 cited sites (78%) appeared in only one engine's answers, so a blended score hides the engine that matters. For ChatGPT, record whether the answer searched the web at all: in one study's configuration, 57.8% of its runs had zero citations.

**How many times should I run the same prompt?**

One 2026 study of four engines recommends at least 7 runs per prompt per day for brand visibility monitoring, and at least 8 when the cited sources matter. A survey of the field says that number is not a universal standard, because it comes from a small set of Swiss queries and at most ten repetitions. Use it as a starting point and repeat until the interval is narrow enough for your decision.

**Why do AI answers change every time I ask?**

Because visibility is a distribution, not a point. In a study of four engines over a window of 45 to 46 days, the sources cited for the same prompt overlapped at a day-to-day Jaccard similarity of 0.34 to 0.42. A survey of the field lists what it depends on: the engine, the date, the location, the wording, whether search ran at all, and randomness in generation.

**How do I track AI Overview rankings over time?**

Treat AI Overviews and AI Mode as two surfaces with two baselines: in TIN's study of seven engines, they shared 8 of their 25 most-cited sites. Use fixed searches, record the date and the country, count the searches where no answer came back, and read the trend over two to four weeks. A crawler's view of the results page is not a real user's personalised result.

**How do I track my brand in Google AI Mode?**

The same way as any engine: fixed buyer questions, several runs, the country and account recorded, and its own baseline. Do not merge it with AI Overviews.

**What is AI share of voice?**

Your brand's mentions divided by all brand mentions in the same set of answers, on a stated prompt panel, engine and date range. Without its instrument and unit it is not a number: TIN found three published figures for Reddit's pre-drop ChatGPT share, 2.05%, 3.83% and 11.97%, depending on whose instrument was read.

**Do I need a paid tool to track AI visibility?**

Not to start. The method on this page runs by hand in a spreadsheet at small scale. Whatever you use should state each decision in the method table on this page.

## Method and author

Written by the TIN research desk. Every figure carries its source link, and nothing is rounded or invented. How TIN works: https://theinternetninja.com/about/