How to track ChatGPT visibility

Tarun Kapoor Updated July 25, 2026 12 min read

You cannot track ChatGPT visibility in analytics, because most AI mentions produce no click and therefore no referrer — the brand is named, the user absorbs it, and nothing reaches your server. Measuring it means sampling the conversation from the outside: presence rate across a fixed prompt panel, citation share, description accuracy, and the referral traffic that does arrive. This guide covers the free methods, the first-party report almost nobody uses, the tool landscape, and the cadence that keeps the numbers meaningful.

Why analytics can't see AI visibility

The instinct is to look in Google Analytics for a ChatGPT referral source. That will show you something, but it will systematically understate your actual visibility — often by an order of magnitude — for a structural reason.

When a model names your brand in an answer, three outcomes are possible. The user clicks a citation link, in which case you get a referral. The user reads the recommendation and searches your brand separately, in which case the visit is attributed to organic or direct. Or the user simply absorbs the information and acts later, or not at all — in which case nothing whatsoever reaches your servers.

The second and third outcomes dominate. Zero-click behaviour was already the norm in classical search — clickstream analysis put roughly 68% of Google searches ending without a click in early 2026 — and a conversational interface that answers the question directly increases that share rather than reducing it.

So the measurement problem is not that tracking is badly configured. It is that the event you care about happens somewhere you have no instrumentation: inside a conversation you cannot see. Every method below is a way of sampling that conversation from the outside.

The four things worth measuring

These are deliberately ordered. Presence rate is the primary metric because it maps directly to consideration-set membership — models name only a handful of brands per answer, so being present at all is the binary that matters. Citation share refines it: appearing once among eight sources is weaker than appearing twice among four.

Description accuracy is the one teams skip and later regret. Being named with an outdated price, a discontinued feature, or a competitor's positioning is worse than not being named — and it is invisible unless someone reads the answers rather than the dashboard.

Referral traffic sits last deliberately. It is real, it is the easiest to measure, and it is the least representative. Treating it as your AI visibility metric is the single most common analytical error in this area. For the full metric definitions, see ChatGPT visibility metrics.

Building the prompt set (the hard part)

Every method here depends on a fixed set of prompts you test repeatedly. The prompt set is the instrument, and a badly constructed one produces confident numbers about nothing.

Four rules make it trustworthy:

1. Write prompts a buyer would actually type, not queries a marketer would. Real prompts are long, conversational, and full of constraints: "we're a 12-person dental practice and our front desk is drowning in appointment calls — what should we look at?" Keyword-shaped prompts produce keyword-shaped answers that no real user sees.
2. Cover the whole funnel, then segment. Include early exploratory prompts, mid-funnel comparison prompts, and late-stage specific prompts. Report them separately — aggregate visibility hides the commercial queries where third-party weighting punishes vendor content hardest.
3. Include prompts you expect to lose. A panel of prompts you already win measures nothing but your own comfort. Deliberately include competitor-favouring and category-generic prompts; those are where the growth is.
4. Freeze it, then change it deliberately. The panel must stay constant to be comparable over time. When you add prompts, version the set and report the old and new panels separately for a period rather than silently redefining your baseline.

A workable starting panel is 30–50 prompts: roughly 40% evaluation-stage ("best X for Y", "X vs Z"), 30% problem-stage ("how do I solve Y"), 20% category-generic, 10% brand-specific ("is X any good"). Scale up once the workflow is running; precision improves with panel size but so does cost.

Free methods: what you can do today

Before buying anything, a manual baseline is worth building — partly for cost, mostly because reading actual answers teaches you things a dashboard never will.

The manual prompt audit. Run your panel by hand in a logged-out or temporary chat, capture each answer, and record: were you mentioned, which competitors were mentioned, which sources were cited, and was your description accurate. A spreadsheet with one row per prompt-run is entirely sufficient.

Two methodological cautions. Use a fresh or temporary session so personalisation and memory do not contaminate results — a logged-in account that has discussed your brand before will over-report your presence. And keep location consistent, because answers vary by market.

This is slow: 40 prompts × 3 runs is a couple of hours. But it produces a genuine baseline, it surfaces the qualitative problems tools miss, and it tells you whether paying for automation is worth it before you pay.

The free report almost nobody uses

Bing Webmaster Tools publishes AI citation data for verified sites, and it is the most underused free instrument in this field.

Two reports matter. The AI Page Stats report lists your pages by citation count — how often each was used to ground an AI answer. The AI Search Queries report lists the grounding queries that produced those citations, along with intent classification, topic, your citation count, and — most valuable of all — your citation share on each query.

That last column is the metric everything else in this guide approximates, delivered as first-party data at no cost. It tells you not just that you were cited, but what proportion of all citations for that query were yours — which is the difference between "we appear sometimes" and "we own this question."

Three ways to use it that most teams miss:

• Find your weakest high-volume queries. Sort by citation count descending, then look at share. A query where you hold 200 citations at 25% share has more unclaimed ground than one where you hold 40 at 60%.
• Calculate addressable volume. Citations ÷ share gives the total citations available on that query. Subtract yours and you have a ranked list of what is winnable on questions already proven relevant to you.
• Compare against your Google indexation. These are separate systems. Pages can be heavily cited by AI while absent from Google's index, and vice versa — and knowing which is which changes what you fix.

The obvious caveat: this is Bing and Copilot grounding data, not ChatGPT's own. Because ChatGPT's search has drawn on Bing's index, it is a reasonable directional proxy — but it is a proxy, and it should be triangulated against direct prompt testing rather than treated as ground truth.

Tracking the traffic that does arrive

Referral traffic understates visibility, but it is worth capturing properly because it is the only layer that connects to revenue.

• Build a dedicated channel group in your analytics for AI sources, keyed on the referring hosts of the major assistants. By default this traffic scatters across Referral and Direct and disappears into aggregate.
• Segment it, then judge it on behaviour — pages per session, conversion rate, assisted conversions — not on volume. AI referral volume is small almost everywhere; its value is in quality.
• Add a self-reported attribution field to forms. On this channel it frequently captures more than any referrer ever will, because it catches the branded-search and direct paths that AI mentions actually produce.
• Watch branded search volume in Search Console as a leading indicator. Rising AI presence typically shows up there before it shows up in referrals.

The monitoring tool landscape

Dedicated tools automate the prompt panel across multiple engines and track results over time. Representative options as of 2026:

What actually differentiates them, in order of practical importance:

1. Prompt volume included. Pricing is usually banded by tracked prompts, and a 25-prompt tier will not cover a serious panel. This is the constraint that most often forces an upgrade.
2. Engine coverage. If your buyers use Perplexity or Gemini, ChatGPT-only tracking is a partial view.
3. Geographic coverage. Answers differ by market; single-market tracking misleads international brands.
4. Competitor benchmarking. Share of voice against a named competitor set is the reporting most teams end up needing.
5. Raw answer export. Underrated. If you cannot read the actual answers, you cannot audit description accuracy.

A reasonable adoption path: run the manual audit for one cycle to establish a baseline and learn the qualitative picture, cross-check against free Bing AI reports, then buy the cheapest tool that covers your prompt volume and engines. Buying before you have a considered prompt panel usually means paying to automate the wrong questions.

Measuring citation share correctly

Citation share is the most decision-useful metric in this discipline and the easiest to compute incorrectly.

The definition that matters: of all sources the engine cited when answering a given query, what proportion were yours? Not what proportion of answers mentioned you — that is presence rate. Share measures depth of grounding, and it moves independently of presence.

Why the distinction pays: presence rate tells you whether you are in the consideration set; citation share tells you how load-bearing you are within it. A brand mentioned in 80% of answers but supplying 10% of citations is being name-checked while competitors supply the substance — a fragile position that collapses the moment a competitor publishes something better.

Two practical rules. Segment share by query stage, because informational and commercial queries behave differently and a healthy aggregate routinely conceals a weak commercial position. And track the addressable gap, not just your share: citations ÷ share gives the total available, and the difference is the prize. A query where you hold 33% of 470 available citations has more unclaimed ground than one where you hold 60% of 40.

Cadence: how often to measure

Resist daily measurement. Run-to-run variance is large enough that daily numbers are mostly noise, and the underlying signals — third-party mentions, corroboration, domain credibility — change on a timescale of months.

Six measurement mistakes

1. Using referral traffic as the visibility metric. It measures the small clicking minority and misses the majority who never leave the conversation.
2. Testing in a logged-in account. Personalisation and chat memory inflate your own presence. Use temporary or logged-out sessions.
3. Drawing conclusions from one run. Generative variance is real. Report rates across repeated runs.
4. Measuring only prompts you win. Comfortable panels produce comfortable numbers and no growth.
5. Reporting aggregate visibility. It hides the commercial-query weakness that matters most, because informational wins mask evaluation-stage absence.
6. Editing the prompt panel silently. Every historical comparison becomes invalid. Version it.

Turning measurement into action

Measurement only earns its cost if it changes what you publish. The routing is fairly mechanical once you have presence rate and citation share segmented by query.

The third and fourth rows are where most B2B brands actually sit, and both are fixed off your own domain rather than on it. Why that is, and what to do about it, is covered in why ChatGPT recommends your competitor. For the mechanism underneath all of this, see how brands appear in ChatGPT responses.

Frequently asked questions

Sources and further reading

• Bing Webmaster Tools — AI Page Stats and AI Search Queries reports (first-party citation counts and per-query citation share for verified sites).
• SparkToro — Zero-click search research (clickstream analysis of searches ending without a click).
• Zapier — The best AI visibility tools and Surmado — Profound vs Peec vs Otterly (tool pricing and coverage).
• Context Hints — visibility metrics, how brands appear in ChatGPT, and why ChatGPT recommends your competitor.

Tarun Kapoor

Founder & CEO, Context Hints

Twelve years of media buying across GroupM, WPP, Ogilvy & Mather, and Neil Patel Digital. Has personally owned media for Nestlé, Sage, Qualcomm, Aetna, Weight Watchers, Chubb and Novotel.