AI citation tracking: how to track brand mentions in AI search

Mentions and citations fail differently, and most tracking setups never separate them.

Key takeaways

• An AI mention is your brand named in the answer text. An AI citation is your URL used as a source. You can be mentioned constantly and cited never.
• Tracking works by sampling, not logging: no pixel sits inside a model's answer, so every figure is an estimate with a range.
• Manual tracking is free, but 40 prompts across 4 engines weekly is 160 responses to read.
• Myrah scores the signals separately: Presence 55% of an engine score, Prominence 30%, Citation 15%.
• Mention gaps and citation gaps need different fixes, so one blended number hides the work.

Yes, you can track how AI assistants talk about your brand: run a fixed set of unbranded prompts across engines on a schedule and parse what comes back. What you cannot do is pull a log. Count two things separately: how often a model names you in the answer, and how often it links your site. A brand can sit in the answer to "best project management tools for agencies" every week and never get a source link, meaning the model knows your category position but takes its facts elsewhere. This guide is for marketers and agency strategists whose brand already shows up in ChatGPT and Perplexity. Scoping note: it covers AI assistants citing your brand and website, not academic citation indexing.

What AI citation tracking is

AI citation tracking is the practice of monitoring how often AI assistants name your brand in their answers and how often they use your URLs as linked sources. It matters because a model can describe your category position accurately while sending every supporting link to a competitor's page.

The wider category sits in our guide to what AI visibility means; this page is the measurement half.

Mention vs citation

AI brand mentions are the times your brand name appears in the text of an AI assistant's answer to a question you did not ask about yourself. They measure whether a model associates you with a category or use case, independent of whether any link back to your site appeared.

Keep them apart, because they fail differently. A mention reflects what the model associates with your name. A citation reflects a retrieval step picking a URL. Two numbers, two owners, two backlogs.

Table 1. Mention and citation compared, including what neither can tell you.

Is it possible to track brand mentions in AI search?

Yes, but by sampling, not counting. That should change how you read every report you get, including ours.

No analytics pixel sits inside a model's answer. Those conversations happen inside OpenAI, Google LLC, Anthropic and Perplexity AI, and none publish per-brand mention feeds. You reconstruct the picture: send a fixed prompt set to each engine repeatedly and parse the responses for your brand name and your links.

That gives an estimate with a range, not a count. Ask the same model the same question twice and the wording, the brand order and the source list can all shift. Variability is a property of these systems, which is why Myrah's published principles include showing it rather than hiding it. So report "roughly 4 of 10 category answers this month, up from about 2 of 10", and treat small wobbles as noise. The variance math sits in our guide to how brands measure visibility in AI.

Anyone selling an exact count of who saw your brand in ChatGPT is selling a modelled number. Ask what the sample was.

How to track brand mentions in AI search by hand

Start with a spreadsheet and no budget. Run this before buying anything: it teaches you what your prompt set should contain.

Prerequisites, before step 1:

• 20 to 40 buyer questions in your category, none containing your brand name.
• A fixed set of 3 to 5 competitors, held for the quarter.
• Accounts on the engines you track, and a browser profile you keep logged out.
• The 12-column sheet below, and one named owner.

The procedure:

1. Write prompts as category, comparison and use-case questions: "best X for Y", "X vs Z", "how do I do Y". Keep branded prompts out: asking an assistant about your own company guarantees a mention and measures nothing.
2. Open a fresh session per engine, logged out where you can, with memory and personalization off, and no chained follow-ups.
3. Paste the prompt unmodified. Anything you add changes the answer you are measuring.
4. Save the answer verbatim, then fill one row per prompt per engine.
5. Repeat three times across a sample, to see how far answers move between runs, then re-run the identical list on a fixed cadence. A changed list resets your trend line.

The tracking spreadsheet: 12 columns

Those values are one worked row for "best AI visibility tools for agencies": named third, described fairly, link to a review site. A citation gap, not a mention gap.

Why manual tracking breaks

It breaks on arithmetic. Four forces compound.

Variance between runs. One observation is a single draw from a distribution, so you need repeats, and repeats multiply reading time.

Model version drift. Vendors update the models behind consumer products. Without the version recorded, a version change and a real drop look the same.

Personalization. Logged-in sessions, memories, location and prior turns shape answers. Your own account is not a neutral instrument.

Combinatorial cost. 40 prompts across 4 engines is 160 responses a week, or 1,920 cells at 12 columns each. At 90 seconds a response that is 4 hours a week of data entry, and 12 hours once you add three repeats.

Manual tracking is the right way to learn the problem and the wrong way to run it for a year.

What automated tracking measures

Automation earns its place by holding conditions still. Myrah audits run in four stages, set out on the Myrah methodology page.

Prompt construction. Prompts come from your niche, competitor positioning and buyer-intent phrases, organized as Category, Compare and Use case, and you edit them before a run. Branded prompts are excluded: a prompt with your name in it produces a mention by construction.

Provider selection. Four engines are queried, OpenAI, Google Gemini, Anthropic Claude and Perplexity AI, on fast generally available tiers, with model versions published so runs replicate.

Query execution. Prompts go out individually, with no system prompts, no conversation history and no appended instructions, and with search tools enabled. Brand names are matched only in responses, never injected into prompts. Each control removes a way of manufacturing the result: history creates carryover, system prompts bias selection, naming your brand guarantees the hit.

Response parsing. Mentions, ranking positions, sentiment context and citations are extracted by deterministic rules, plus LLM passes for ambiguous cases.

Scoring keeps the signals apart per engine: Presence 55% for how often you were mentioned, pulled toward a fixed per-engine baseline; Prominence 30% for mean reciprocal rank among the brands the engine named, scoring zero where you were absent; Citation 15% for how often you were cited where citations were on the table. Engine scores average equally, and results map to 0–100 on a fixed curve, versioned as scoring v2.0.0. Note the denominator: citation counts only against answers that showed sources, and fades into presence as that opportunity disappears.

That is why "just ask ChatGPT about your brand" is not measurement. One logged-in, branded, conversational query breaks four controls at once.

Which AI engines show their sources

Citation behaviour differs by product. These are documented behaviours as of mid-2026, not ranking claims; re-check the docs.

Table 2. Where each product's source links come from, per vendor documentation.

Two consequences. A "not cited" result means nothing until you know whether the answer showed sources at all, which is what the third state in your sheet is for. And blocking the wrong agent drops you out of citation eligibility, while model-development controls are separate: Google says Google-Extended does not affect Google Search inclusion or ranking.

Presentation shifts too. In May 2026, Google expanded Preferred Sources into AI Overviews and AI Mode and added link carousels and "Highly Cited" labels: presentation changes, not a ranking factor, that can still move your measured citation rate.

Turning a citation gap into work

A citation gap arrives with its brief written: your sheet holds the question, the engine, and the URL that won instead of yours. Filter for rows where mentioned = y and cited = n, group by theme, and compare the winning pages with yours.

Usually the difference is specificity. The cited page answers the literal question in its first paragraph, in a structure a model can lift. The durable guidance from Google agrees, rejecting llms.txt, artificial content chunking, AI-only rewrites and special AI schema as requirements. Sequencing those fixes is covered in our guide to improving brand visibility in AI search engines.

Mention gaps are different work: category framing and third-party coverage. Compare your rate of being named with your rivals' using competitor benchmarking.

Reporting cadence

Weekly runs while a campaign is live, monthly for reporting. Faster measures noise; slower and you cannot tie a change to a cause.

A healthy setup has six properties: a fixed, versioned prompt set; a competitor set held for a quarter; model versions recorded; mentions and citations reported separately; ranges shown with the sample size; and one change shipped at a time.

A metric without its denominator is a rumour, so report the sample size beside every percentage.

Frequently asked questions

How do I see if AI mentions my brand?

Run 20 to 40 unbranded buyer questions through each assistant in a logged-out session and record whether your name appears in the answer. Log the engine, model version, date and locale. One check tells you nothing; a repeated fixed set gives you a rate.

What's the difference between an AI mention and an AI citation?

A mention is your brand named in the answer text. A citation is your URL displayed as a source. Mentions show the model ties you to the category. Citations show a page of yours was the most useful thing to quote. You can have one without the other.

Can I track brand mentions in ChatGPT for free?

Yes. Paste a fixed prompt list into a logged-out session, record each answer in a spreadsheet, and repeat on a schedule. The limits are time and consistency: 40 prompts across 4 engines weekly is 160 responses to read. Free suits learning, not a year of it.

How often should I run AI mention tracking?

Weekly during an active campaign, monthly as a reporting rhythm. Use the identical prompt set each time, because a changed list resets the trend. Re-run after a publishing push or PR moment, and record the model version so drift is not read as performance.

Why do my results change between runs?

Because generated answers are sampled, not read from a fixed index. Wording, brand order and source lists shift between identical prompts, and vendors update the models behind their products. Treat every figure as an estimate, use repeat runs, and ignore movement inside your spread.

Which AI engines show their sources?

Perplexity AI shows sources by default, Google's AI Overviews draw supporting links from Google Search, ChatGPT Search links when it has searched, and the Gemini API's google_search grounding returns inline citation annotations. Answers produced without retrieval show no sources at all.

Conclusion

AI citation tracking answers two questions most reports collapse into one: does the model name you, and does it link you. Separate them from day one. A mention gap sends you to positioning and third-party coverage; a citation gap sends you to one page that must answer one question better. Build the 12-column sheet this week with 20 to 40 unbranded prompts, then automate once the arithmetic outgrows what a person can read. Read the Myrah methodology for the controls a run holds constant.

Sources

• Google — AI features and your website
• Google — AI optimization guide
• Google — Google-Extended
• Google — Original, high-quality content
• Google — Grounding with Google Search
• OpenAI — OpenAI bots
• Perplexity AI — Perplexity crawlers
• Anthropic — Anthropic web crawlers