Guide

How to tell whether AI search is sending you customers.

The short answer: split assistant referrals into their own channel, measure how often you are cited across a fixed panel of buyer prompts, and use branded search to corroborate. Clicks alone will understate AI search badly, because the most valuable citations never produce a visit. Here is how to run each of those, and what none of them can settle.

Split AI referrals out of your existing analytics

Where it goes wrong

When someone clicks a link inside ChatGPT, Perplexity, Gemini or Copilot, the visit usually arrives with a referrer from that assistant's own domain. Most analytics setups bucket those hosts under generic referral or, worse, under direct, so the traffic exists in your data but has no name and nobody looks at it.

What to do instead

Create one channel grouping that captures the assistant hostnames you care about, chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com among them, and report it as its own line rather than folding it into referral. Keep the host list in version control and revisit it, because new assistants appear and existing ones change domains. The number will look small at first. Small and named beats large and unattributed.

Accept that AI Overviews traffic is not separable

Where it goes wrong

Google AI Overviews appear inside the normal results page, so a click from an Overview reaches you as ordinary google.com organic traffic. There is no referrer that distinguishes it. Any tool claiming to report your AI Overviews clicks as a clean, separate number is inferring, not measuring.

What to do instead

Stop trying to isolate it and watch the shape of the queries instead. In Search Console, look for pages where impressions hold steady or rise while clicks fall. That gap is the signature of an answer being read on the results page rather than clicked through. It is evidence, not proof, so treat it as a direction rather than a figure.

Measure share of answer, not just clicks

Where it goes wrong

The most valuable outcome in AI search often produces no visit at all. The model reads your page, uses it, cites you, and the person forms a view of you without ever landing on your site. Judge AI search purely on sessions and you will conclude it does nothing, right up until it decides a deal.

What to do instead

Build a panel of the prompts your actual buyers would type, the comparison questions, the category questions, the objections. Run the panel across the engines your market uses and record who gets cited, in what position and in what framing. Your share of answer is the percentage of that panel where you appear as a source. Score your competitors on the same panel or the number means nothing on its own.

Use branded search as the corroborating signal

Where it goes wrong

Citation without a click still leaves a trace, just not where you are looking. Someone who reads about you in an AI answer and is persuaded tends to search your name afterwards. That shows up as branded search, and it is easy to misread as a general brand-awareness win.

What to do instead

Track branded query volume against the periods when your share of answer moved. If citations rise and branded search follows, you have two independent signals pointing the same way, which is far stronger than either alone. If share of answer rises and branded search does not, your citations are probably passing mentions rather than recommendations, which is a content problem, not a measurement one.

Write down what you still cannot know

Where it goes wrong

AI answers vary between runs. The same prompt can cite you today and not tomorrow, with no change on your side. There is no equivalent of rank tracking that is stable enough to treat as ground truth, and no engine publishes a complete picture of when it used your content.

What to do instead

Sample rather than chase completeness. Run the same panel on the same cadence, report the trend across runs instead of any single result, and state the sample size next to the number. A share-of-answer figure quoted without a panel size and a date is not a measurement. Being explicit about the error bars is what makes the rest of the number usable.

Three weak signals beat one confident number

None of the four measurements above is conclusive by itself. Assistant referrals undercount, the impressions-minus-clicks gap is circumstantial, share of answer is a sample, and branded search has other causes. Read together and moving together, they are a reliable read on whether AI search is working for you. Read individually and quoted to two decimal places, any one of them will mislead you.

That is the honest position on AI search measurement today, and it is worth saying plainly that our own GEO tracking is on the roadmap rather than live. SEO and AEO measurement in Discoverability works today.

Related reading: Five marketing numbers that are probably lying to you, and the Analytics module on attributing revenue you can actually trace.

Know where your discoverability actually stands.

Decifer measures SEO and AEO from your own data today, and is candid about what AI search can and can't yet be measured on.

Request access How the engine works