To track brand mentions in AI search, write the questions your buyers ask, run them in ChatGPT, Perplexity, Google AI Overviews and Google AI Mode on a schedule, and count how often each answer names your brand and links to your site. That is all AI mention tracking is. There is nothing to crawl, because the answers do not exist until someone asks.
The rest of this guide is the detail that decides whether your numbers mean anything: which two rates to track, how many times to ask, how each engine behaves, and the mistakes we made running this on our own category.
We ran it, and the result was not flattering. Across 97 answers to five questions about our own category, RankScope was named in none of them. More on why below.
What AI Mention Tracking Measures
Google's own AI Overview for "ai mention tracking" splits the job into the same parts we use: mentions, citations, sentiment, share of voice against competitors, and the sources behind the answer. Here is what each one is and how to count it.
| Metric | What it counts | How to calculate it |
|---|---|---|
| Mention rate | Answers that name your brand | Answers naming you ÷ answers collected |
| Citation rate | Answers that link to your domain as a source | Answers citing your domain ÷ answers collected |
| Share of voice | Your mentions against your competitor set | Your mentions ÷ all brand mentions in the set |
| Position | Where you appear in the answer | First named, second named, and so on |
| Sentiment | How the answer frames you | Positive, neutral, negative, or "recommended with a caveat" |
| Sources | Which pages the engine cited when it named someone | Domains cited, counted across answers |
Mentions and citations are different numbers
A mention is your brand name in the answer text. A citation is a link to your page in the sources.
They come apart all the time. An answer can name you while citing a G2 page or someone else's "best tools" list as its source. It can also cite your blog post for a definition and name three competitors as the recommendations. If you only track one, you will misread the other.
When we traced ChatGPT answers about our own category on 2026-11-14, each answer read 12 to 35 pages and cited 3 to 17 of them. Reading is not citing, and citing is not naming. Three different filters, and you only see the last one in the answer.
Why the sources column matters most
When a competitor is named and you are not, the cited sources usually tell you why. In the same trace, one page (Zapier's list of AI visibility tools) was cited 22 times across all four engines. Being on that page is worth more than any single page on your own domain.
That is also why our zero is not a mystery. Those answers cite published lists, and we were on none of the 13 most-cited ones.
Why Tracking AI Mentions Is Different from Brand Monitoring
Traditional brand monitoring tools such as Brand24, Mention, or Brandwatch crawl the web for text that names your brand: social posts, news, forums, reviews. The web is indexed, so you search it. For the wider monitoring workflow, including alerts and reporting, see the guide to LLM monitoring.
AI answers are not stored anywhere public. When someone asks ChatGPT "what's the best tool for X?", it writes a fresh answer from its training data plus whatever it retrieves at that moment. The only way to know whether you appear is to ask.
So three things change:
- Google Alerts and scrapers find nothing. There is no static page.
- One check is not a measurement. The answer can change on the next run.
- You measure prompts, not keywords. The unit is "how often am I named across a set of realistic questions".
The channel is large. Alphabet says Google AI Overviews reach over 1.5 billion users a month, and ChatGPT had 900 million weekly active users in February 2026.
For what you are optimising toward once you have the numbers, the guide to generative engine optimization covers the strategy that tracking feeds.
A Worked Example: What Two Runs of the Same 200 Answers Showed
Before trusting any tracker, including ours, we wanted to know how much of a change between two scans is real and how much is the engines rolling dice. So we ran the same scan twice.
Setup (2026-11-02 and 2026-11-03). 10 brands, 5 buyer questions each, 4 engines (ChatGPT, Google AI Overviews, Google AI Mode, Perplexity). That is 200 answers per pass. The second pass ran about 3 hours after the first. We compared only the 168 answers that came back in both passes.
| What we compared | Result | What it means |
|---|---|---|
| One answer: named or not | 16 of 168 flipped (10%) | Never read a single answer as a verdict |
| Flips by engine | AI Overviews 3 of 19, AI Mode 5 of 49, ChatGPT 5 of 50, Perplexity 3 of 50 | Every engine moves, AI Overviews the most |
| Per-brand total ("named in X of ~16") | changed by a median of 1, at most 3 | Totals are usable if you allow for about ±2 |
| Which brands appear in one answer | overlap of 0.25 (AI Overviews) to 0.67 (Perplexity) between runs | The list inside a single answer is unstable |
| Top 5 competitors, aggregated | 4 or 5 of 5 the same in 9 of 10 brands | Aggregated competitor rankings are stable |
| Our own brand | named in 1 of 18 answers, then 0 of 18 | Consistently invisible in our own category |
These are small samples, and they come from one pair of runs on one day. Treat them as an order of magnitude, not a constant. The direction was consistent everywhere we looked, though: single answers are noisy and totals are not.
A separate study of local plumbers (20 cities, 5 questions, 4 engines, run twice) showed the same pattern at a larger scale. Two engines answering the same question named largely different businesses, and that disagreement was 2 to 3 times larger than one engine's own run-to-run noise. So a result on one engine tells you little about another.
What went wrong, and what it cost
On the first pass, 22 of our 50 Google calls failed with a provider error. Our script logged each failure as "no AI Overview shown".
That is a 44% failure rate silently recorded as a finding. Every brand looked less visible in AI Overviews than it was, and nothing on the report said so. With up to 3 retries per call the failure rate fell to 10% (5 of 50), and only 6 of 50 questions genuinely showed no AI Overview. Retries raised the cost per AI Overview call from $0.0030 to $0.0044. Cheap, next to a wrong chart.
If you track AI Overviews, by hand or by script, keep "the engine showed no AI Overview" and "the check failed" as two separate outcomes. They mean opposite things.
Our opinion: more questions, fewer repeats
Many guides, including an earlier version of this one, say to run each prompt 10 to 50 times. We think that is the wrong place to spend the effort.
Our data says a single answer flips about 10% of the time and a per-brand total over 16 or so answers is stable to within 2 or 3. Running one question 50 times buys a very precise number for one question. Running 30 questions 3 times buys a usable number for the 30 different ways buyers actually ask. The second is more useful, and it costs less.
This could be wrong for categories where answers are far less stable than ours, so check your own flip rate before you settle on it: run your set twice and count how many single answers changed.
The Four AI Engines to Monitor
Each engine works differently, and the differences decide how you track it.
ChatGPT
With search on, ChatGPT retrieves live web pages and lists its sources. With search off, it answers from training data. Track the search-on version: it is what most people asking buying questions see, and it is the one your new pages can reach.
Two things from our trace are worth knowing. On shortlist questions like "alternatives to Profound", ChatGPT's own searches already contained brand names in 10 of 10 runs: it searched to confirm and price a list it already held. And the 54 domains it read for one question had zero overlap with Bing's top 50 for the same query. Ranking on Bing did not predict what ChatGPT read.
If ChatGPT matters most to you, the ChatGPT SEO tools comparison covers the trackers and what they cost, and the ChatGPT rank tracker page shows how RankScope measures it. Whether two people get the same answer at all is covered in does ChatGPT give the same answer to everyone.
Google AI Overviews
AI Overviews sit at the top of normal Google results. Google does not publish how often they appear; Ahrefs found AI Overviews on 20.5% of 146 million desktop keywords it checked in September 2025.
This is the engine brands forget, because it looks like "just Google". It was also the least stable engine in our test, and the one where failed checks hide most easily. To see which sources the AI Overview cites for your keywords today, the free AI Overview checker loads the signed-out result for up to 10 keywords, with AI Mode's sources beside it. An AI Overviews tracker repeats that check daily.
Perplexity
Perplexity shows numbered sources on every answer, so it is the easiest engine to audit for citations. It was also the most consistent between our runs.
On "alternatives to X" questions it leaned hard on vendor-owned alternative pages: in our trace it cited nine of them on all 5 runs of one question. That is useful if you publish those pages, and a warning if a competitor does and you do not.
Google AI Mode
AI Mode is Google's conversational search, separate from AI Overviews. It handles follow-up questions and writes longer answers. A brand named in the first answer can drop out on the follow-up, so test short conversations as well as single questions.
How to Track AI Mentions by Hand
Manual tracking is where everyone starts. It is the right way to get a first baseline of 10 to 20 questions, and it teaches you how the engines behave before you pay for a tool.
In ChatGPT
- Open a fresh chat each time (signed out, or with memory off) and make sure search is on.
- Ask an unbranded buyer question: "What are the best tools for [your category]?" or "How do I solve [specific problem]?"
- Record: named or not, position, what the answer says about you, which competitors are named, and the sources listed.
- Repeat 2 to 5 times in fresh chats.
In Perplexity
- Run the same questions on perplexity.ai.
- Record the answer text and the numbered sources separately.
- Note whether your domain is a source, whether or not your brand is named.
- Repeat across sessions.
In Google AI Overviews
- Search the question on google.com, signed out if you can.
- Note whether an AI Overview appears at all. Write "none shown" if it does not; do not leave the cell blank.
- Record brand mentions in the summary and the sources it lists.
- Try two or three phrasings, because AI Overviews do not trigger on every query.
In Google AI Mode
- Open AI Mode from the Google search bar.
- Ask a broad question, then a follow-up ("which of those suit a team of five?").
- Record mentions and sources for each turn.
Where manual tracking breaks
The arithmetic gets you first. 30 questions × 4 engines × 3 runs is 360 answers per check. Weekly, that is a part-time job.
Then you lose history. A manual check says where you stand today. It cannot tell you whether last month's article moved anything, or when a competitor started appearing.
And the data degrades. Past 20 or so questions, people start running each question once, skipping an engine, or checking less often. The spreadsheet still fills up, but the numbers stop meaning what they say.
Build the Question Set
Your question set is the foundation of any AI mention tracking, manual or automated. It should read like the questions buyers in your market type.
Good questions look like:
- "What are the best [category] tools for [use case]?"
- "How do [your buyers] deal with [specific problem]?"
- "What software do marketing teams use to [outcome]?"
- "[Competitor] vs [competitor]: which is better for [situation]?"
- "What should I look for in a [category] platform?"
Leave these out of mention rate:
- "Tell me about [your brand]". Useful for checking how you are described, but no buyer who has not heard of you asks it.
- Questions so narrow that only you could be the answer. They inflate the rate and describe no real buyer.
- Questions so broad they produce noise, like "what are all the software tools?"
Aim for 20 to 40 questions, weighted toward unbranded discovery. Record every answer in a sheet with these columns: question, engine, run, answer shown (yes / none shown / check failed), named (Y/N), position, your domain cited (Y/N), competitors named, sources cited, sentiment.
That "answer shown" column is the one we wish we had added first.
The Metrics, in More Detail
Mention rate
The share of answers that name you, over all answers collected. Quote the raw count with it: "10 of 40" tells a reader how much weight the 25% can carry. On a few dozen answers, a move of 2 or 3 can be noise.
We will not give you a benchmark like "30% is good". We have not seen a source for one that holds across categories, and in our own measurements the number depends far more on which questions you chose than on anything else.
Citation rate
The share of answers that link to your domain as a source. It tells you whether your own pages are being used, which is the part you control most directly.
Share of voice
Your mentions as a share of all mentions across your competitor set. Mention rate is absolute; share of voice is relative. A 40% mention rate where the leader has 80% is a different position from 40% where nobody passes 20%. The share of voice guide covers the calculation and its traps.
Position
Where you appear in the answer. Being named first is different from being named eighth, and a high mention rate at low positions usually means you are seen as a secondary option.
Sentiment
Whether the answer frames you well. "RankScope is one option, though some users find the pricing steep" is a mention doing negative work. Be careful with sentiment read from keywords, though. In our plumber study, the biggest brand had 10 mentions with negative-sounding words nearby. On reading, all 10 were recommendations with a caveat ("ask for a flat-rate breakdown"). None said to avoid it. Read the answers before acting on a sentiment score.
Engine-level breakdown
All of the above, per engine. Engines disagree about who is authoritative, so a blended number can hide a strong ChatGPT result and an empty AI Overview.
Automating AI Mention Tracking
Once you pass about 20 questions or two engines, automate. A tool changes three things.
History. Daily scans build a timeline, so you can see whether mentions moved after you published something, and the noise in single runs averages out.
Change detection. When an answer changes (you are added or dropped, a competitor's framing shifts, a new source appears), you can see the before and after text. That is how you connect a piece of work to an outcome instead of guessing.
Competitors at scale. Tracking yourself and 10 competitors across 150 questions and 4 engines by hand is not realistic.
What to look for in a tool
- Engine coverage. ChatGPT, Google AI Overviews, Perplexity and Google AI Mode, or at least the ones your buyers use. Check which engines each plan includes, not only the top plan.
- Your own prompts. Brand mentions are prompt-based. You should be able to write the exact questions.
- Mentions and citations as separate numbers. If a tool reports one blended "visibility" score, ask how to see both.
- Answer history. You want to see what changed in the text, not only that a score moved.
- Failed checks reported as failures. Ask the vendor how a failed request shows up. Our own pipeline got this wrong once.
- Competitor tracking in the same view, since share of voice needs it.
What RankScope does
RankScope runs your prompt set daily and includes share of voice trending, sentiment scoring and Answer Diff (how each answer changed) on every plan. Starter is $49/mo, or $39/mo billed annually, for 40 prompts on ChatGPT and Google AI Overviews with 3 competitors. Professional is $149/mo, or $124/mo billed annually, for 150 prompts on all four engines (ChatGPT, AI Overviews, Perplexity, Google AI Mode) with 10 competitors, plus competitive alerts. There is no free trial.
If you want a one-off look before paying anything, the free AI visibility checker and AI citation tracker need no signup. For the other tools in this category with their prices and trade-offs, see the AI brand monitoring tools comparison, and for citations specifically, the AI citation tracking guide. The AI rank tracker guide goes deeper on measurement per engine.
What to Do With What You Find
Tracking is the input. Here is how to use it.
Step 1: Map the gaps
After your first full pass you will have a list of questions where a competitor is named and you are not. For each one:
- Which competitors were named instead?
- Which pages did the engine cite for that answer?
- What do those pages cover that yours do not, and are you listed on them?
Step 2: Work out what is driving the competitor's mentions
There are two routes into an answer.
Your own pages. The engine retrieves and cites your content directly. You control this one through what you publish.
Third-party pages. The engine has seen your brand on review sites, comparison articles, industry lists and community threads. On shortlist questions this matters more: our ChatGPT trace showed the model searching for brands it already had in mind. A brand repeated across many independent lists is the one it holds.
If a competitor is named because of a good comparison page on their own site, you can publish one. If they are named because 15 roundups list them and none list you, the work is getting onto those roundups. For the outreach side of that, MentionAgent's roundup of backlink software options compares the tools built for earning editorial links from relevant blogs.
The guide on how to get cited by ChatGPT covers both routes in detail.
Step 3: Watch for the change
After you publish a page or land a listing, you want to know whether it moved anything. With answer history you can see the specific change: when it happened, what text changed, whether you were added or moved up. Without it you wait for a monthly average to drift and guess at the cause. Cited sources also change quickly; how often AI citations change has the numbers.
Step 4: Check access before blaming content
If you are absent from every answer on every engine despite solid content, check whether the engines can reach you at all. Look at robots.txt for GPTBot, OAI-SearchBot and PerplexityBot. Some hosting setups block them by default. The free AI crawler checker tests your robots.txt against the AI crawlers.
Also confirm your key pages are indexed in Bing Webmaster Tools. Indexing is a precondition for some retrieval. Just do not expect Bing rank to predict what ChatGPT reads; in our trace it did not.
A Four-Week Setup
Week 1: Question set and baseline
- Write 20 to 40 unbranded buyer questions.
- Run each on ChatGPT (search on), Perplexity, Google AI Overviews and Google AI Mode, twice.
- Record named, position, cited, competitors, sources, and whether an answer was shown.
- Calculate mention rate, citation rate and share of voice. Write down the raw counts.
- Pick your top 5 gaps: questions where a competitor appears in both runs and you appear in neither.
Week 2: Audit the gaps
- Open every page the engines cited for those 5 questions.
- Note what they cover, how they are structured, and whether you are listed.
- Check whether you have a page on the same topic, and what it is missing.
- Check
robots.txtand Bing indexing.
Week 3: Fix and publish
- Fix any crawler blocking.
- Update or publish one page for the highest-priority gap, with a direct answer at the top of each section.
- Start outreach to the two or three third-party pages that were cited most often.
Week 4: Monitor
- Re-run the full set and compare against week 1, allowing for ±2 on small totals.
- Set a schedule: weekly by hand, or daily with a tool once you pass 20 questions or two engines.
- Set alerts for drops in mention rate and jumps in a competitor's share of voice. The AI search monitoring guide has suggested thresholds.
Common Tracking Mistakes
Reading one answer as a result. Ask ChatGPT the same question five times and you will likely see different brands. In our test, 1 in 10 single answers flipped between two runs 3 hours apart.
Recording a failed check as "not shown". We did it on 22 of 50 Google calls. Keep failures separate.
Tracking only branded questions. "Tell me about [brand]" shows how you are described. It does not show whether buyers find you.
Treating mentions and citations as one number. They move separately and mean different things.
Skipping AI Overviews. It looks like ordinary Google, so it gets left out, and it was the least stable engine in our test.
Tracking yourself without competitors. 20% means one thing when the leader has 80% and another when nobody passes 15%.
Doing it by hand past the point it works. Once you are running questions once instead of three times to save time, the data has stopped being trustworthy.
Frequently Asked Questions
How do you track brand mentions in AI search?
Write 20 to 40 unbranded questions your buyers ask, run each one in ChatGPT (search on), Perplexity, Google AI Overviews and Google AI Mode, and record whether each answer names your brand, links to your site, and which competitors it names. Run each question 2 to 5 times and report totals, because single answers change between runs. For more than about 20 questions or two engines, use an AI visibility tool that runs the set on a schedule.
What is AI mention tracking?
AI mention tracking is monitoring how often, where, and how AI engines such as ChatGPT, Perplexity, Google AI Overviews and Google AI Mode name your brand in their answers. Unlike web or social listening, there is no page to crawl: the answers are generated on demand, so you have to ask the engines a fixed set of questions on a schedule and count the results. The core numbers are mention rate, citation rate, share of voice, position and sentiment.
What is mention rate in AI search?
Mention rate is the share of AI answers that name your brand, out of all the answers you collected for a set of questions. If you collect 40 answers and 10 name you, your mention rate is 25%. Quote it with the raw counts (10 of 40) and over enough answers to be stable: in our own test, one brand's count moved by up to 3 out of about 16 answers between two runs 3 hours apart.
What is the difference between a mention and a citation in AI search?
A mention is the engine naming your brand in the answer text. A citation is the engine linking to a page on your domain as a source. They are separate numbers: an answer can name you while citing a third-party list, or cite your page without naming you. Track both, and track which third-party pages are cited when a competitor is named instead of you.
How do I track brand mentions in ChatGPT?
Turn search on, ask your unbranded buyer questions in a fresh chat each time (signed out or with memory off), and record whether your brand is named, its position, and the sources ChatGPT lists. Run each question several times, because ChatGPT's answers vary between sessions. To track it continuously, use a tool that runs your questions on ChatGPT daily.
How do I track brand mentions in Perplexity?
Run the same questions on perplexity.ai and record two things for every answer: whether your brand is named in the text, and whether your domain appears in the numbered sources. Perplexity shows its sources on every answer, which makes it the easiest engine to audit for citations. In our tests it was also the most consistent engine between runs.
Can Google Search Console track AI brand mentions?
No. Search Console reports impressions and clicks for your pages in Google Search, including when a page appears in AI features, but it does not tell you whether an AI answer names your brand, and it says nothing about ChatGPT or Perplexity. For mentions you have to query the engines directly or use an AI visibility tool that does.
How often should I track brand mentions in AI search?
Weekly at minimum if you are actively publishing or doing outreach, because cited sources change from day to day. Daily scans give you a trend line that smooths out single-run noise. Monthly checks show direction but will not tell you which change caused a move.
Do I need to track every AI engine?
Track the engines your buyers use. For most brands that means ChatGPT, Google AI Overviews, Perplexity and Google AI Mode. They disagree: in our tests, two engines answering the same question named largely different brands, so a result on one engine does not tell you much about another.
If you manage visibility for a brand, RankScope for brands runs your questions daily across ChatGPT and Google AI Overviews on Starter, and all four engines from Professional, with share of voice trending, sentiment scoring and Answer Diff on every plan. On Professional and above, RankScope's GEO agents can run on a set schedule and generate action plans when your mention patterns shift. Get started at rankscope.ai.