How to Actually Measure AI Search Traffic

The most common question we were asked this year, almost word for word: sessions are down about a fifth, enquiries are the same or better, and nobody can tell me why.

That pattern is now common enough to have a cause. People are asking an assistant instead of running three searches, so the browsing that used to happen on your site happens in a chat window first, and the person who arrives has already read about you. Fewer visits, further along.

For most of the last two years there was no way to see any of that in a report. In 2026 that changed twice, and both changes are narrower than the headlines suggested.

What actually shipped this year

May 13, 2026: GA4 added an AI Assistant channel. It is part of the Default Channel Group, so it appears in your acquisition reports with no setup. Google's release note describes it as a way to "identify how users are discovering your site through chatbots like ChatGPT, Gemini, and Claude". When an arriving session's referrer matches a recognised assistant, GA4 stamps the medium ai-assistant, files it under the AI Assistant channel, and tags the campaign (ai-assistant).

June 3, 2026: Search Console added a generative AI performance report. It shows how often your pages appeared inside Google's own AI features. It rolled out to a subset of properties first, so not every account has it.

These two reports are frequently talked about as though they are the same measurement. They are close to opposites.

The distinction that trips everyone up

Google's AI Overviews and AI Mode traffic does not appear in GA4's AI Assistant channel. The documentation for the default channel group says so plainly: the AI Assistants channel "excludes Google's AI Overviews and AI Mode", and Organic Search is the channel that includes them.

That is defensible (a click from an AI Overview is a click from a Google search results page) but it means the obvious mental model is wrong. AI Assistant is not "traffic from AI". It is traffic from everyone's AI except Google's.

The report that has no clicks in it

Search Console's generative AI report gives you impressions and nothing else. No clicks, no click-through rate, no average position. You can break impressions down by page, country, device and date, and that is the full set.

An impression here means links to your site were shown to somebody inside a generative AI feature on Google Search. So the report answers one question well, which of my pages does Google pull into its AI answers, and cannot answer the question anyone actually asks next, which is what that was worth.

Data from Search Labs experiments is excluded, and a site that has opted out of Google's AI features is not eligible for the report at all.

It is still the most useful thing in the account for this work. If three of your service pages carry almost all your AI impressions, that tells you what Google considers your quotable material, and it is usually not the pages you would have guessed.

We have written before that Search Console is not the place to look for AI results. That remains true of the specific thing that post was about. A citation inside a ChatGPT answer will never show up in Search Console, because it did not happen on Google. The new report covers Google's own AI surfaces, which is a different question and now a measurable one.

Most of your chatbot traffic is filed as Direct

Here is the part no report fixes.

A referrer is a header the browser sends saying where the click came from. Analytics has no other way to attribute a visit. Assistants lose that header constantly:

• In-app browsers. A citation tapped inside the ChatGPT, Claude or Perplexity mobile app often opens in an in-app or system browser that starts a fresh session with no referrer.
• Copy and paste. Somebody reads an answer, copies the URL, and opens it in a new tab. There is no referrer in that flow at all, and it is a very normal way to use an assistant.
• Deliberate stripping. Some assistant links carry rel="noreferrer", which removes the header by design.

Every one of those lands in Direct. Published estimates of how much AI referral traffic disappears this way run from about a third to around 70%; the highest of them, a vendor analysis of 446,405 visits published in February 2026, put it at 70.6%. Treat any single figure as unproven, since it is vendor data and the real answer depends on how your audience splits between app and desktop, but treat the direction as settled. Whatever your AI Assistant channel says, the real figure is materially higher.

This also means a rising Direct line is not automatically brand strength. It might be. It might be ChatGPT with its return address removed.

The one setup step worth doing

Google's recognised list is not exhaustive. The examples in its own documentation name ChatGPT, Gemini, Claude, DeepSeek, Copilot and Grok. Perplexity is not among them, and Perplexity is one of the assistants most likely to send an actual click, because it shows citations by default.

So build a custom channel group in GA4 that matches on source, and include everything you care about rather than everything Google has got round to:

chatgpt.com
chat.openai.com
perplexity.ai
claude.ai
gemini.google.com
copilot.microsoft.com
grok.com
deepseek.com
meta.ai

Admin > Data display > Channel groups > Create new channel group. One rule, source matching that list, and you have a line you can trust more than the default one. Add to it when a new assistant starts sending traffic; this list will be out of date within the year.

What to put in the report you actually send

Sessions are the wrong headline for this. They are falling for structural reasons across most of the web, and a chart of them tells a business owner they are losing when they may be winning.

Four things are worth reporting, in this order:

1. Enquiries by channel, not sessions by channel. If AI-referred visits are 3% of sessions and 9% of form fills, that is the whole story and the session count is a distraction.
2. Direct, watched as a trend and split by landing page. Direct traffic landing on a deep service page is not somebody typing your domain from memory. That is displaced AI referral, and the landing page tells you which answer sent it.
3. Branded search volume in Search Console. People who meet you in an AI answer routinely search your name afterwards to check you are real. Brand queries rising while non-brand falls is a healthy pattern, not a decline.
4. Generative AI impressions, by page. Not as a performance metric. As a map of which pages Google finds quotable.

The check no dashboard performs

Once a quarter, open ChatGPT, Gemini, Claude and Perplexity, and ask each one the question a customer would ask. Then ask it what it knows about your business by name.

You are looking for three things: whether you come up at all, whether what it says about you is accurate, and which sources it cites. The third is the most useful. It tells you which pages and which third-party listings are doing the work, and it is the only view of the citation side you can get, since none of it will ever reach an analytics report.

Write down what you find. Five minutes a quarter, in a document, is a better record of your AI visibility than any tool currently sold for the purpose.

Measurement is the part clients ask for last and need first, because without it every other decision is a guess about a channel nobody can see. Setting it up properly is part of our AI SEO and analysis service, and if the traffic is not arriving in the first place, start here instead.