How to Track AI Mentions Across ChatGPT, Claude, Gemini, Grok, and Perplexity

Why AI mentions need their own measurement system

For twenty years, "are we visible?" had a clean answer: check your rank. You sat at position three for a keyword, you watched it move, and the metric was continuous and stable. AI answers broke that model completely. When someone asks ChatGPT "what's the best CRM for a solo real-estate agent?", the assistant returns one paragraph naming two or three products. You are either in that paragraph or you do not exist — there is no page two, no scroll, no "almost ranked." Visibility went from a position on a list to a binary event, and most brands have no idea how often the coin lands in their favor.

It gets harder. The same prompt typed twice can return different brands, because the models are non-deterministic. Each engine sources its answer differently — Perplexity reads the live web, Claude leans on training memory, Grok weights real-time chatter on X. And because there is no "AI rank tracker" baked into Google Search Console, the only way to know whether ChatGPT, Claude, Gemini, Grok, and Perplexity recommend you is to measure it deliberately and repeatedly.

This guide is the framework for that. It is five components — a prompt set, a sampling method, a scoring rubric, a cross-platform baseline, and a cadence — that together turn "I hope we're getting recommended" into a number you can defend in a board meeting. If you are still mapping how this whole channel differs from classic search, our explainer on how GEO differs from traditional SEO is the right primer before you start measuring.

1. Build a prompt set that mirrors how buyers actually ask

The challenge it solves. Most teams "check" their AI visibility by typing their own brand name into ChatGPT — which proves nothing, because the model will obviously describe a company you just named. Real buyers do not search your brand; they describe a problem and ask for recommendations. If your measurement uses vanity prompts, you will record a flattering mention rate that has zero relationship to whether you win unbranded, high-intent conversations.

The fix. Construct a fixed prompt set of the questions a *prospect who has never heard of you* would actually ask. These are conversational, long-tail, and intent-rich — the same phrasing logic we cover in targeting long-tail over short-tail queries. Organize them by funnel stage so you can see where you appear and where you vanish: discovery ("what tools help with X?"), comparison ("X vs Y for small teams"), and decision ("best X for [specific role/use case]").

Implementation steps

1. 1Write 20–40 unbranded buyer questions, each one a query where your product is a *legitimate* answer — not aspirational ones you have no business winning.
2. 2Tag each prompt by funnel stage (discovery, comparison, decision) and by the customer segment it represents.
3. 3Include 3–5 explicit comparison prompts that name a key competitor, so you can see how assistants frame you head-to-head.
4. 4Freeze the list. The prompt set must stay identical run to run, or your trend data is measuring your edits instead of your visibility.

2. Sample each platform enough to beat randomness

The challenge it solves. A single screenshot of ChatGPT naming you is not data — it is an anecdote. Because the models sample from a probability distribution, the same prompt can name you on Monday and skip you on Tuesday with no change to your site at all. Teams that report a one-shot result either celebrate a fluke or panic over noise, and both lead to bad decisions.

The fix. Treat each prompt like a coin you flip several times. Run every prompt 3–5 times per platform and report the percentage of runs you appear in rather than a yes/no. This converts a jittery single answer into a stable mention rate. Where it matters, vary the account and the region too, because logged-in personalization and geography shift recommendations — an answer in a fresh incognito session is your truest "cold buyer" reading.

Implementation steps

1. 1Run each prompt at least 3 times per engine; for high-stakes decision prompts, push to 5 or more.
2. 2Use a clean, logged-out or fresh session per run so prior chat history does not bias the model toward what you discussed before.
3. 3Record every run's raw output verbatim, not just your verdict — you will want the exact wording later for sentiment and competitor analysis.
4. 4Calculate mention rate per prompt = (runs you appeared in ÷ total runs), then average across the set for a headline number.

3. Score every answer with a four-part rubric

The challenge it solves. "Did they mention us?" is too blunt to act on. A grudging mention buried at the bottom of a list, with no link and a lukewarm tone, is worth far less than a top-line recommendation with a citation to your pricing page — yet a yes/no log scores them identically. Without a rubric, you cannot tell an improving position from a deteriorating one.

The fix. Score each answer on four axes. Named — does your brand appear at all? Cited — is a specific URL on your site linked as a source? Sentiment — is the framing positive, neutral, or a caveat ("X is cheaper but limited")? Competitive set — who appears *alongside* you, and in what order? The gap between "named" and "cited" is the most important distinction in AI tracking: a citation drives referral traffic and proves the engine retrieved your page, while a name-only mention rides on training memory. Our guide to earning citations with llms.txt and schema explains how to convert the former into the latter.

Implementation steps

1. 1For every recorded answer, log four fields: named (y/n), cited (y/n + the exact URL), sentiment (positive/neutral/negative), and the ranked list of brands mentioned.
2. 2Track which *page* gets cited, not just that you were cited — this tells you exactly which content the engines trust.
3. 3Compute a "share of voice": your appearances ÷ total brand appearances across the answer set, so you measure dominance, not just presence.
4. 4Flag negative or caveated mentions for review — a confidently wrong or unflattering AI description is a fixable content problem, not just a low score.

4. Normalize across platforms so the numbers are comparable

The challenge it solves. The five engines are not interchangeable, and comparing them naively misleads you. Perplexity and Gemini cite live URLs you can click; ChatGPT and Claude often recommend from training data with no link; Grok weights real-time posts on X. A "low Claude score" and a "low Perplexity score" mean completely different things — one is a training-coverage gap, the other a retrievability gap — and a single blended average hides both.

The fix. Track each engine on its own line *and* roll up a normalized cross-platform index, with each platform's quirks annotated so a low score points to the right fix. Interpret per-engine results diagnostically: weak Perplexity/Gemini means your pages are not retrievable or quotable; weak ChatGPT/Claude means the broader web does not establish you strongly enough for the model's training to have absorbed you; weak Grok means thin social presence.

Implementation steps

1. 1Maintain a per-engine dashboard (ChatGPT, Claude, Gemini, Grok, Perplexity) plus one normalized roll-up, never a single blended figure that erases the differences.
2. 2Weight the roll-up by where *your* buyers actually are — a B2B SaaS may weight ChatGPT and Perplexity heavily; a consumer brand may care more about Gemini in AI Overviews.
3. 3Annotate each engine: live-web (Perplexity, Gemini), training-led (Claude, ChatGPT default), real-time-social (Grok), so every score maps to a fix type.
4. 4For the citation-capable engines, cross-check your cited URLs against server logs to confirm the AI crawlers can actually reach those pages.

5. Run it on a cadence and watch the trend, not the snapshot

The challenge it solves. A one-time audit is a vanity exercise. AI answers drift constantly — models get updated, the live web is re-crawled, competitors publish new content — so a number captured once is stale within weeks. Teams that check sporadically can never tell whether a change in mentions came from their own work, a competitor's move, or a model update.

The fix. Lock in a recurring schedule and let the trend line do the talking. Run a priority subset weekly and the full prompt set monthly, always with the identical method, so any movement reflects reality rather than a change in how you tested. The goal is a chart of mention rate over time per engine, with annotations for what you shipped — that is the artifact that proves your GEO work is paying off.

Implementation steps

1. 1Schedule a fixed cadence (weekly priority set, monthly full set) and treat it as a standing operational task, not an occasional curiosity.
2. 2Keep an annotation log: every time you publish or fix a page, mark the date on the chart so you can connect cause to effect.
3. 3Watch competitor share of voice on the same chart — a drop in your rate while a rival's climbs is an early warning worth acting on fast.
4. 4Set a simple alert threshold (e.g., mention rate falls below X% on any engine) so a slide triggers investigation instead of going unnoticed.

Putting it all together

Tracking AI mentions is not one check — it is a small, disciplined system: a frozen prompt set that mirrors real buyer questions, enough sampling to beat the randomness, a four-part rubric that separates a cited recommendation from a throwaway name-drop, a per-engine view that turns each score into a specific fix, and a cadence that produces a trend you can act on. Build that once and "are AI assistants recommending us?" stops being a guess and becomes a number you manage like any other funnel metric.

The catch is that doing all of this by hand — prompting five engines, several runs each, dozens of prompts, every week — is exactly the kind of repetitive work that quietly stops happening when the team gets busy. That is the gap CookMyRank was built to close: it runs your prompt set across ChatGPT, Claude, Gemini, Grok, Perplexity, and Google AI Overviews, scores the mentions and citations, and shows you the trend and the open lanes over time. Start by running a free AI-visibility audit to see where you stand today, then close the gaps with one-click SEO and GEO fixes and read more playbooks on the CookMyRank blog. We measure and improve your AI visibility — we will never promise to "guarantee" a recommendation — but we will show you, prompt by prompt, exactly whether the engines are putting you in the answer. When you are ready to monitor continuously, our plans and pricing lay out the options.