AI Search Tracking: How to Tell If ChatGPT, Perplexity & AI Overviews Actually Cite You

Start with the mental model, because it explains every measurement gap that follows. Three separate things can happen involving your content and an AI engine, and they happen at radically different frequencies.

The gap between the first and the last is enormous. Cloudflare's analysis of its own network traffic — published in July 2025 and covering the week of 19–26 June 2025 — found Anthropic's crawler made roughly 71,000 HTML page requests for every single referral it sent back, while Mistral sat at the opposite extreme, sending about ten referrals for every crawl request. The specific ratios move constantly and vary by platform, so treat the numbers as a snapshot rather than a constant; the structural point is what matters and it has not changed. Cloudflare's own summary of the trend was that these models "continue to consume more content, more frequently, despite sending the same or less traffic to the source of its content."

So server logs showing heavy AI crawler activity tell you almost nothing about whether you are being recommended. Conversely, a citation that a user reads and acts on without clicking — the AI answered their question, they now know your name — leaves no trace in any analytics tool you own. That invisible middle is precisely why the third layer of the stack exists.

On 3 June 2026 Google announced Search Generative AI performance reports in Search Console: dedicated views of impressions within generative AI features on Search — AI Overviews and AI Mode — plus generative AI features in Discover. Google said it was rolling them out to a subset of websites to test and gather feedback before making them widely available, so if you don't see the report yet, that is expected rather than a configuration problem.

Two clarifications save a lot of confused analysis. First, these dedicated views are impressions-focused — they answer "did I appear?" rather than "what did that earn me?" Treat them as a visibility trend, not a traffic report, and don't build revenue attribution on them. Second, and more important: this is not a new pool of data sitting outside your normal numbers. Google states the generative-AI data is also included in the overall performance report, where it continues to contribute to your site's overall Search visibility. Clicks that came via an AI Overview have been in your main Search Console totals all along, blended in under the Web search type.

That second point quietly resolves a question we get asked constantly — "is AI eating my Search Console clicks?" You cannot answer it by subtracting one report from the other, because they overlap by design. What you can do is watch whether your AI-feature impressions grow while total clicks stay flat, which is the honest shape of the zero-click concern.

Google Analytics added an AI Assistant channel to its default channel group, confirmed in its Analytics Help documentation on 13 May 2026 and applied globally in the weeks after. It requires no configuration: when Google Analytics detects a referrer matching its list of AI assistants, it sets the session medium to `ai-assistant` and the campaign to `(ai-assistant)`, and the session is grouped under AI Assistant in default channel reports.

Useful — but there are three limitations that materially change how you read the number, and none of them are obvious from the report itself.

The most consequential is the exclusion. Google's documentation is explicit that this channel does not cover its own AI Overviews and AI Mode, which are treated as Search rather than as an AI assistant referral. So your AI Assistant channel is a measure of third-party assistants only. If you present it internally as "our AI traffic," you are systematically underreporting, because the largest AI surface by user count is definitionally excluded from it.

The second is coverage. Google names ChatGPT, Gemini, DeepSeek, Copilot and Grok as examples of recognised assistants. Perplexity is not named in that documentation — so unless and until it is added, Perplexity sessions will keep landing in generic Referral. If Perplexity matters to you, build a custom channel group rather than assuming the default catches it.

The third is the date. The channel only classifies traffic from the point it went live; it does not retroactively reclassify your history. Any before-and-after comparison spanning mid-May 2026 is comparing two different definitions, not two different realities — and that artefact will look exactly like a growth story if you let it.

Layers 1 and 2 both measure consequences — an impression, a session. Neither can tell you the thing you most want to know: when a buyer asks an engine the question your business exists to answer, does your name come up? Most of the time that event produces no click at all, so the only reliable way to observe it is to ask the engines yourself, systematically, and record what happens.

That is what a prompt panel is: a fixed list of the questions your buyers actually ask, run against each engine on a schedule, in fresh sessions, with the results logged. It is unglamorous and it is the highest-signal thing on this page. Done properly it turns a vague anxiety into a tracked metric — a citation rate per engine that moves over time.

We built one for ourselves rather than buying one, because we wanted the panel fixed and the history ours: a set list of prompts spanning brand questions, category questions with no brand named, and the specific topics we publish about, run against multiple engines and scored for whether the site is cited and whether the brand is mentioned. Building it is genuinely simple — the discipline is in never changing the prompt list, because the moment you edit it you lose comparability with everything you recorded before.

The rules that make it trustworthy are worth stating, because each one corresponds to a way people accidentally fool themselves:

Eventually, maybe. Not to start. The paid platforms in this space mostly automate exactly what the prompt panel does — run prompts at scale, across more engines and locations than you would manage by hand, and chart it. That scale is real value once AI search is a material channel for you, and it is worth paying for at that point.

But there is a sequencing argument for building the cheap version first. A hand-run panel of twenty or thirty prompts costs you an hour a month and teaches you which questions actually matter — and that judgement is what makes a paid tool useful later. Buying the dashboard first tends to produce a beautifully charted number that nobody can act on.

The other reason to start manual is honesty about the underlying instability. These engines are non-deterministic and their retrieval changes without notice; any vendor implying a precise, stable share-of-voice figure is overselling what the data can support. Read every AI-visibility number — ours included — as a directional trend.

Pulling the three layers together, this is the minimum viable dashboard. It fits on one page, takes about an hour a month, and each row answers a different question — which is the entire point of a layered stack.

Every one of these has come up in real conversations with founders this year. They are worth knowing before you present anything to a board or a client.

We set the measurement up before touching content, because the alternative is spending three months on visibility work with no way to prove what it did. In practice that means wiring the GA4 channel groups properly (including the custom rules the defaults miss), confirming Search Console access and reading the generative-AI report if the site has it, checking server logs for what the AI crawlers can actually reach, and standing up a fixed prompt panel with a recorded baseline.

It is deliberately unglamorous work, and it is the part that makes everything afterwards accountable. It also frequently surfaces the real problem on day one — several times this year the finding has been that the client's key pages were not renderable without JavaScript, which is a build issue rather than a marketing one, and one we can simply fix because we do both.

As a scope and price anchor: a measurement setup with a documented baseline is typically a few days of work from about ₹25,000 (~$350) for a standard business site, quoted fixed after a free call. Ongoing AI-search work is scoped per engagement. What we will not do is sell a monthly retainer against a number nobody can validate — if the panel says you are absent from every unbranded prompt in your category, we would rather show you that, honestly, and talk about what would change it.