Two things changed in 2026 that make this question answerable in a way it was not before. Google Analytics added a dedicated AI Assistant channel to its default channel group on 13 May 2026, and Search Console launched generative AI performance reports on 3 June 2026.
Both are genuine improvements. Neither answers the question people think it answers. Between them you can now see roughly how often your pages appear inside Google's AI features, and roughly how many sessions arrived from an assistant that passed a referrer. You still cannot see how many people bought from you because ChatGPT told them you existed, and no tool on the market can.
That gap matters commercially, because it is the gap every "AI visibility" dashboard quietly fills with estimates. This post is about reading those numbers honestly. If you want the setup steps — channel groups, regex, the events to fire — we keep them in a separate guide to measuring AI search traffic in GA4. This is the companion piece about whether to believe what you then see.
Does Search Console separate AI Overviews and AI Mode from normal search?
Partly, and the distinction is worth getting exactly right because a lot of advice gets it wrong in both directions.
Impressions: yes, in a separate report. The new generative AI performance reports give you a dedicated view of impressions within AI Overviews and AI Mode in Search, plus generative AI features in Discover. Google's announcement lists what is in them: impressions, pages, countries, devices, and dates down to hourly granularity. The reports were rolling out to a subset of properties first, so check whether yours has them yet.
Clicks: no. The announcement is explicit that the metric set is impressions, with a note that Google is "continuing to work with website owners to understand what insights and data would be most helpful", including "adding additional metrics over time". Clicks are not in that list today. And the AI data is not held separately from everything else — Google says it "is included in the overall performance report", so clicks from an AI Overview are counted, they are simply mixed in with every other web click.
Google's own AI features and your website documentation says the same thing from the other direction: sites appearing in AI features "are included in the overall search traffic in Search Console", reported "in the Performance report, within the 'Web' search type".
The counting rules are worth reading too, because they explain why some AI numbers look strange. Per Search Console's performance data documentation, clicking an external link in an AI Overview or in AI Mode counts as a click. An AI Overview "occupies a single position in search results, and all links in the AI Overview are assigned that same position" — so your average position can move for reasons that have nothing to do with ranking. And in AI Mode, a follow-up question is treated as a new query, with its own impressions and clicks.
So the honest summary: Search Console can tell you whether Google's AI features show your pages, and it cannot tell you what share of your Google clicks came through them.
Where do assistant referrals actually appear in GA4?
GA4's default channel group documentation now defines the AI Assistant channel as the one "by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok", and states that it "excludes Google's AI Overviews and AI Mode". Those clicks stay in Organic Search, which the same page confirms includes AI Overviews and AI Mode.
Mechanically, GA4 sets medium to ai-assistant and campaign to (ai-assistant) when the referrer matches its list of recognised AI assistants. Three practical consequences:
- The list is not published in full. The channel description names five examples; Google's release note names ChatGPT, Gemini and Claude. Perplexity appears in neither. Anything unrecognised falls into Referral under its own hostname, so keep looking at your referral report rather than trusting the channel to be complete.
- It is not retroactive. The channel exists from 13 May 2026 forward. Comparing this quarter's AI Assistant channel against last year's is comparing a thing against its own absence.
- Google's AI traffic is deliberately somewhere else. If someone shows you a single "AI traffic" figure that includes AI Overviews, they have built it themselves out of two sources that count differently.
What turns into Direct, and why that is not fixable
GA4 attributes a session to Direct when it has nothing else to go on. Several common AI paths produce exactly that:
- App link taps. The ChatGPT, Gemini, Perplexity and Claude mobile apps open links in in-app browsers, and these commonly send no
Refererheader at all. - Referrer policies. A page served with a restrictive
Referrer-Policystrips the referring hostname on cross-origin navigation by design. This is a privacy feature, not a bug, and you have no say in it. - Copy and paste. Someone reads an answer, types your name into a browser, and arrives. No referrer exists to send.
- Voice and summary answers with no click at all. The assistant says the thing. Nobody visits. The influence is real and the session count is zero.
One partial rescue: query parameters survive where referrers do not, and ChatGPT has been widely observed appending utm_source=chatgpt.com to cited links in its search-grounded answers. GA4 prefers campaign parameters over the referrer, so those sessions attribute correctly even from an app. That behaviour is not documented by OpenAI as a commitment, so treat it as something to verify in your own logs rather than something to build a report on. Search your landing-page report for the parameter and see for yourself.
The number nobody has
The best public evidence on the size of this gap comes from Cloudflare, which publishes a crawl-to-refer ratio on Radar's AI Insights — how many HTML page requests a platform's crawlers make for every one HTML request arriving with that platform's hostname as the referrer. Their methodology post explains the calculation, and the ratios are large and move month to month, so look up the current figure rather than quoting one from a blog post.
The caveat Cloudflare attaches to its own data is the most useful sentence in this whole area: because referral counts only include traffic from the providers' web-based tools, the calculations "may overstate the respective ratios, but it is unclear by how much". Cloudflare sees a meaningful fraction of the web and still says the direction is clear and the magnitude is not.
That is the correct posture to have about your own numbers.
What each source can and cannot tell you
| Source | Can tell you | Cannot tell you |
|---|---|---|
| Search Console generative AI report | Impressions in AI Overviews, AI Mode and Discover AI features; which pages; country; device | Clicks, CTR, or what share of your Google clicks came from AI features |
| Search Console Performance report | Total Google clicks and impressions, Web search type | Which of those clicks came through an AI feature |
| GA4 AI Assistant channel | Sessions from recognised assistants that passed a referrer, from 13 May 2026 | Unrecognised assistants, app traffic with no referrer, anything before that date, Google's AI features |
| GA4 Direct channel | That something unattributable happened | Whether it was an assistant, a bookmark, an email client or a QR code |
| Server logs | Which AI crawlers fetch which URLs, and how often | Whether any of it led to a human visit |
| Third-party AI visibility tools | Whether your brand appears in answers to prompts they test | Whether real people asked those prompts, or what happened next |
Read down the right-hand column before you commission a dashboard.
How to judge an AI traffic number you have been shown
Four questions, in order. They are fast, and they are the ones vendors least enjoy.
1. Is it measured or modelled? A session count from a referrer is measured. "AI-influenced revenue" and "estimated AI visibility share" are models. Models can be useful, but a model presented without its assumptions is a sales aid.
2. Which bucket does it come from? Search Console impressions, GA4 sessions, and prompt-testing tools are three different populations. Any single figure combining them needs its construction shown.
3. Does the denominator exist? Share-of-voice percentages need a total. Nobody knows how many times people asked an assistant about your category this month — the assistants do not publish it. A share calculated against a sample of prompts the vendor chose is a measure of that prompt list.
4. Does it move when nothing happened? AI answer sets are unstable. Ask the same assistant the same question twice and the citations can differ. A tracker showing a clean upward line through that noise is smoothing something.
If the answers are good, the number is probably worth watching. Our own view is that AI search is worth real investment in GEO, and the reason to be strict about measurement is precisely that — you cannot direct spend with a number you cannot interrogate.
What to measure instead
Since attribution is partly unavailable, measure the things that are not.
Ask people. A single optional "How did you hear about us?" field on your enquiry form, with free text, will out-perform your analytics stack for this specific question. It is self-reported and imperfect, and "ChatGPT recommended you" is still a stronger signal than a Direct session. For most Indian B2B enquiry volumes this is the highest-value measurement change available.
Watch Direct as a trend, not a number. Direct sessions rising while branded search and referral stay flat is weak evidence of unattributed discovery. It is not proof, and it will never become proof. Treat it as a reason to look, not a result to report.
Check that the right crawlers can reach you. This is the part that is fully in your control and fully verifiable. Each provider documents a separate search crawler and asks you to manage it in robots.txt: OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity, Claude-SearchBot for Claude, and Googlebot for Google's AI features. Blocking the search bot while allowing the training bot is a common accidental configuration, usually inherited from a security plugin or a CDN rule, and it is the one measurable own goal in this whole area. Check it today; it takes ten minutes.
Count enquiries and revenue. Unfashionable, and it survives every change Google and OpenAI make to their reporting.
When this does not apply
If your Direct channel is small in absolute terms, stop. Reorganising attribution for a channel sending you fifteen sessions a month is a hobby. The measurement work earns its cost when AI referrals are large enough to shift a budget decision, and below that threshold the honest answer is "we cannot tell yet, and it does not matter yet".
If you sell in a category people do not research conversationally — emergency plumbing, walk-in retail, anything bought on proximity — the mechanism barely applies, whatever a dashboard shows. Local search and maps still do the work.
And if someone offers to guarantee a position in AI answers, note that the citation sets are unstable and the ranking systems are undocumented. Nobody can guarantee that. We publish what GEO work costs and what it consists of, and the first deliverable is always a measurement baseline, because you cannot judge anything afterwards without one.
The honest state of play is that AI search is clearly influencing demand, the tooling to prove it is six months old and incomplete, and anyone presenting a precise figure is either modelling or guessing. That is not a reason to ignore the channel. It is a reason to be unusually careful about the numbers attached to it.
Send us your current numbers and we will tell you which parts of them are measured, which are modelled, and whether there is enough signal to act on yet — including when the answer is that there is not.
Bharat Vasireddy
Founder & Digital Marketing Strategist8+ years of experience in performance marketing and web development. I help businesses scale through data-driven SEO, Google Ads, and high-performance websites.