How to Measure AI Search Traffic: Four Systems

I went looking for one number and came back with four that do not add up.

The question behind it is ordinary. What share of a site's traffic now comes from AI search? You will be asked some version of it in a meeting this year if you have not been already, and I wanted to know whether I could answer it honestly. On paper the timing looked good. Bing Webmaster Tools introduced an AI Performance report on February 10, 2026. Google Analytics 4 added an AI Assistants channel to its default channel group. Google Search Console began rolling out a generative AI performance report in June 2026. Three new reports inside one half-year, sitting on top of server logs, the oldest layer, which were always there and remain the one most teams never read. I assumed the answer was waiting in a console I had not opened yet.

It was not. I spent July 29, 2026 reading the documentation instead of the dashboards, and that is where it became obvious: the four do not measure the same thing. One counts HTTP requests, one impressions, one citations, one sessions. No honest arithmetic joins them. And the single system with "AI" in its name, GA4's AI Assistants channel, cannot see the AI surface most people mean when they ask the question.

More reporting made the reporting worse. That surprised me, and on reflection it should not have.

So when a GA4 property is presented to me as measuring AI search, the first thing I check is not the number. It is which system produced it, and what that system cannot see.

Quick answer

Four systems now measure AI search, each in its own unit: server logs count crawler requests, Search Console's generative AI report counts impressions, Bing Webmaster Tools counts citations, and GA4's AI Assistants channel counts sessions. I report them separately, because they cannot be summed or reconciled. GA4's AI channel explicitly excludes Google's own AI Overviews and AI Mode, so it cannot see AI inside Google Search at all.

Does GA4's AI Assistants channel include AI Overviews?

No, and Google's own default channel group documentation, which I read on July 29, 2026, says so plainly: "AI Assistants is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok. It excludes Google's AI Overviews and AI Mode." The definition has two parts. The channel matches when "The medium exactly matches 'ai-assistant'," and GA4 sets that medium itself: "The medium is set to 'ai-assistant' and the campaign is set to '(ai-assistant)' if the referrer matches a list of AI Assistants." It is referrer detection over a list Google maintains — which is exactly why it cannot see a surface whose clicks carry an ordinary google.com referrer. The same page defines Organic Search as arrivals "via non-ad links in organic-search results, including Google's AI Overviews and AI Mode."

So the one channel GA4 built to represent AI is blind to the AI surface inside Google Search. I do not think that is a bug. The channel does exactly what its definition says it does. But the name will be misread by nearly everyone who sees it, and I would not blame a single one of them. AI Overviews and AI Mode sessions land in Organic Search, indistinguishable from ordinary blue-link clicks.

A slide presenting GA4's AI Assistants number as "our AI traffic" is describing referral traffic from a few named chat assistants. If your dashboard is labelled that way, it overstates one thing while hiding another. I write the exclusion into the metric definition itself so it travels with the number, per my GA4 and GTM notes.

What does the Search Console generative AI report actually show?

Impressions. I went looking for more and there is not much more.

Google's Search Console Help page states that the report "includes impressions for the following generative AI capabilities on Google Search: AI Overviews" and "AI Mode," grouped by page, country, date, and device. I read it in full on July 29, 2026: no clicks, no click-through rate, no position, no queries. It also excludes "data from experiments in Search Labs."

The sentence that changed how I read this report is in a different document. Google's AI features documentation, last updated December 10, 2025, states that "sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console" and are reported "within the 'Web' search type."

That sentence does more work than its length suggests. AI clicks are already inside your Web totals, undifferentiated. The new report is an impressions-only breakout sitting on top of a click total you cannot decompose. Divide one by the other and you get a ratio with no referent, because there is no AI click figure to divide by and Google does not publish one.

Why does Bing tell you more than Google?

Because Microsoft chose to publish a smaller unit: the citation. Microsoft's February 10, 2026 announcement introduced Total Citations — the "total number of citations that are displayed as sources in AI-generated answers during the selected time frame" — plus grounding queries, the "key phrases the AI used when retrieving content that was referenced in AI-generated answers." On June 16, 2026 Microsoft added Intents, Topics, Compare, and Citation Share: "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query." Google publishes no equivalent.

Bing gives me more detail than the other three combined. I would still refuse to use it to describe Google. Microsoft calls Citation Share "designed as an observational metric – not a ranking system or a competitive scoreboard," and the February preview says of grounding queries that "The data shown represents a sample of overall citation activity." Different retrieval system, different query population, and the vendor itself calls the sample a sample.

What can only server logs see?

Crawlers. Nothing else observes them, and GA4 structurally cannot. GA4's standard web collection is a client-side tag: something has to execute the page's JavaScript for a measurement call to be sent, so a system that fetches your HTML without running scripts leaves no GA4 record at all. I could not find vendor documentation confirming which AI crawlers execute JavaScript, so I will not repeat the common claim that most do not.

Logs also show you something no dashboard will, which is whether the published commitments hold on your own domain. OpenAI documents four crawlers with different jobs: GPTBot for model training, OAI-SearchBot for ChatGPT search results, OAI-AdsBot for ad safety, and ChatGPT-User for user-triggered visits. Of the last, OpenAI says: "Because these actions are initiated by a user, robots.txt rules may not apply." Meta is equally direct: "The Meta-ExternalFetcher crawler may bypass robots.txt because it performs fetches that were requested by the user."

Vendors differ on robots.txt more than the listicles suggest, and I checked rather than assumed. Anthropic's crawler documentation, dated April 7, 2026, describes ClaudeBot, Claude-User, and Claude-SearchBot. I went in looking specifically for a user-initiated robots.txt exemption and found none: "Anthropic's Bots respect 'do not crawl' signals by honoring industry standard directives in robots.txt." Grouping Claude-User with ChatGPT-User and Meta-ExternalFetcher would make a tidier paragraph. Anthropic's documentation does not support it, so I am not writing it.

Google documents its own tokens separately — Googlebot for Search, and Google-Extended as a standalone control for Gemini training and grounding that Google says does not affect Search inclusion or ranking. Only your logs show which of these commitments is actually kept on your domain. That is why I treat log retention as part of the site's architecture rather than an operations detail, alongside the items in my SEO architecture checklist.

Which AI measurement system answers which question?

Sort them by the question each one can actually answer and most of the confusion goes away.

Logs are a record of who fetched what: which AI systems pulled which URLs, how often, with what status code. Search Console tells you how often Google displayed your pages inside AI Overviews and AI Mode, then stops. Bing is the only one of the four that will name the pages it cited in Microsoft's AI answers and give you a share figure against a grounding query. GA4? Sessions from assistant referrers, plus everything those people did next, which is genuinely valuable as long as nobody calls it the whole picture.

No single console holds all four. I have stopped waiting for one that does.

Layer	Unit it counts	What it cannot see
Server logs	One HTTP request, by user agent and IP	Whether a human ever saw the answer; anything served from CDN cache
Search Console generative AI report	One impression in AI Overviews or AI Mode	Clicks, CTR, position, queries, Search Labs
Bing Webmaster Tools AI Performance	One citation, tied to a grounding query	Any Google surface; sessions; whether anyone read it
GA4 AI Assistants channel	One session, medium exactly matching `ai-assistant`	AI Overviews and AI Mode, routed to Organic Search; crawlers that do not run JavaScript

Never add server log requests, Search Console impressions, Bing citations, and GA4 sessions together. If your dashboard sums them, it is wrong, and the error is not a rounding error. One AI answer can produce a server log line (the fetch), a Search Console impression (the display), a Bing citation (the attribution), and zero GA4 sessions because nobody clicked. It can equally produce a GA4 session with none of the other three, because someone pasted your URL into a chat window days later. The intersections are overlapping and unmeasurable, so summing them double-counts an unknown amount and invents the rest.

What about third-party AI visibility tools?

Third-party AI visibility platforms sell a fifth unit: a prompt-sampling score built by querying assistants on your behalf and counting how often you appear. None of the four systems above produces that unit, and none of them can verify it. Before I put one in a client's stack I want the sample size, the prompt set, and the run-to-run variance. Those three answers tell me more about a tool than its score ever will.

How to set up a defensible baseline this quarter

Write down which system answers which question before you build the dashboard. Not after. This is where it goes wrong in exactly the way described in why analytics implementations fail: instrument first, define later, then discover the metrics disagree and argue about which one is lying.

Turn on server log retention. Keep at least ninety days. Classify requests by the user agents each vendor documents, and recheck those pages quarterly, because the user-agent tokens change. Log evidence is the only defense against a "we blocked them" claim that was never true.
Connect the two vendor reports. Pull in the Search Console generative AI report and Bing's AI Performance report, and store the exports separately. Never in the same table or on the same chart axis. Put the unit, impressions or citations, in the column header.
Annotate the GA4 channel. Make its definition travel with it: "AI Assistants (referrals only: excludes Google AI Overviews and AI Mode)." Then state openly that Google's AI sessions live in Organic Search. My analytics implementation checklist covers agreeing definitions and naming an owner before launch.
Take a position on Google's Search generative AI control. Google's Help documentation states that for sites using it, "You won't receive any traffic or impressions from these features," and that it "isn't used as a ranking or inclusion signal affecting other parts of Search."

You will finish with four numbers and no total. I know that is a worse-looking slide than a single AI traffic percentage. It is also the only one you can defend when somebody asks where the number came from. If AI also produces the content you are measuring, the governance argument in AI agents for SEO content applies here too.

My working take

If I could fix one thing this quarter, it would not be a dashboard. It would be log retention, because logs are the only layer nobody else can redefine on me. Google can change what the generative AI report contains. Microsoft can change what counts as a citation. Google wrote the GA4 channel definition and can rewrite it. Nobody edits my access logs.

Everything after that follows from one decision: report the four numbers and refuse to produce a fifth that combines them. I know that is the unpopular answer. A single AI traffic percentage is easy to present and impossible to defend, and the four-number version is the exact opposite on both counts.

The part I am least comfortable with is GA4, and it is a naming problem rather than a data problem. A channel called AI Assistants that excludes Google's own AI will be misread by almost everyone who sees it, and the misreading is reasonable. Until the name changes, an annotation is carrying weight the name should be carrying.

And when someone asks what share of their traffic comes from AI search, my honest answer right now is that nobody can tell them. Not Google, not Microsoft, not a visibility vendor. I would rather say that in the first minute of the meeting than hand over a confident number that quietly means something narrower.

Frequently asked questions about measuring AI search traffic

Does GA4 track AI Overviews traffic?

Not under that label. Google's default channel group documentation states that the AI Assistants channel "excludes Google's AI Overviews and AI Mode," and that Organic Search includes them. Once those sessions land in Organic Search, no dimension separates them again.

Can I see clicks from AI Overviews in Search Console?

No. The generative AI performance report gives impressions only, and Google's AI features documentation states that AI feature traffic is "included in the overall search traffic in Search Console" — inside your Web totals, and not separable from them.

Why do my AI traffic numbers never match across tools?

They are counting different events: HTTP requests, impressions, citations, sessions. Report each separately.

Does blocking AI crawlers in robots.txt actually work?

Partially, and it depends entirely on the vendor. OpenAI documents that for user-initiated fetches, "Because these actions are initiated by a user, robots.txt rules may not apply." Meta documents that Meta-ExternalFetcher "may bypass robots.txt." Anthropic's April 7, 2026 documentation offers no such exemption. Your logs are the only place you will find out which of those commitments held on your site.

Which AI measurement system should I start with?

Server logs. They are the only layer that shows crawler behavior, and the only one nobody else can revoke or redefine. Search Console and Bing Webmaster Tools come next, both free and vendor-authoritative for their own surface. Treat GA4's AI Assistants channel as what it is: a referral-traffic report with a narrow scope.

Sources I used

Verified at source: July 29, 2026.

One sourcing note. I left out widely circulated AI-search statistics such as crawl-to-referral ratios and AI Overviews query-coverage percentages, because I could not reach their primary sources.