You have read the warnings by now: impressions up, clicks down, AI eating the traffic, search running out of road as a channel. Each one more urgent than the last — we have run pieces like that here ourselves.
So you opened Google Search Console, and the curve had not really collapsed. Then you opened your analytics, looked up chatgpt.com, and found a number small enough to ignore — a few dozen sessions, a percentage point either way, possibly zero. Add up every AI source and it still looks nothing like the flood those articles described.
Which leaves you with a perfectly reasonable conclusion: the shift is real, but it is not urgent, and there is time to assess it properly.
But that conclusion came off a broken instrument.
The same Google Search Console holds a second report. The generative AI features report tells you how many of your impressions occurred inside AI Overviews and AI Mode — for our own site over the past three months, that figure sits at roughly 14%.
Here is the part that is easy to miss: those impressions are not extra. They were already counted in the very table you just looked at. Google’s own documentation says the generative AI report draws its data from the Web search type in the Performance report. Which means the curve you read as reassuringly intact is being held up, in part, by AI.
The shift is not pending. It already happened inside the number you took as evidence that nothing had happened.
Fourteen percent of impressions running on AI surfaces, while the AI referrals your analytics can name round to nothing. Both cannot be true. The analytics number is the wrong one, and it is wrong systematically.
The referrer field records the last click
The browser’s rule for sending a referrer is simple: whatever page you clicked the link on is what gets recorded. It answers “where did this click happen,” and it has never been able to answer “how did this person hear about you.”
For twenty years those two questions overlapped closely enough that nobody had to separate them. Someone searched, saw a result, clicked it — discovery and arrival were the same gesture, one second apart, and the referrer captured the whole causal chain honestly. AI split that second into two halves that can sit hours apart.
Three paths that launder AI traffic into something else
One: the answer contains no link. Similarweb’s July 2026 generative AI report tracked citation presence in US ChatGPT prompts: the share of answers carrying a clickable source link rose from roughly 1.6% in June 2025 to about 6.8% by May 2026. Fast growth — and read the other way round, it means more than 93% of answers give the user nothing to click. They remember the brand name, close the tab, and type that name into a search box the next morning. Google books it as organic search.
Two: the click happens inside a mobile app. Cloudflare put it plainly back in July 2025: traffic referred by Claude’s native app carries no Referer header, and they judged the same to hold for other native apps. The visitor genuinely arrived from an AI answer and lands on your site with no label attached, filed under direct traffic. There is no credible public percentage for this path — nobody can count how much traffic vanishes this way, which is the subject of this article.
Three: the visitor doesn’t trust the answer. This is the one people underestimate, and it runs two ways.
The on-site version happens on Google’s own page. An AI Overview puts the answer, and your name, right at the top. The visitor reads it, isn’t quite convinced, and scrolls down to click a blue link instead. That click books as ordinary organic search, so your report tells you the SEO is working — when the thing that actually did the work was your presence in the AI answer. You double down on rankings, and the money goes to the wrong box on the page.
The cross-site version moves the visitor somewhere else. Shown a recommendation in a chat product, they open a search engine first to check the company is real, look for complaints, see what the website looks like, and only then click through. An April 2026 survey by Rithum and Retail Dive (1,046 online shoppers across the US and UK) asked the people who do verify AI recommendations what they do first: 28% go to a search engine, ahead of reading reviews at 19%, asking friends or family at 17%, and visiting the brand’s own site at 5%.
Both versions share a shape: the decision was made by the AI, and the click landed on a Google results page.
There is a third outcome that is worse. They never click at all. But the AI has already said your name out loud in front of them — how warmly, alongside whom, in what tone — and all of that went in. Three months later, when they go looking for this kind of service, whether your name surfaces first depends on those impressions they never clicked. No analytics package on earth records that. Your report counts clicks, and half the value of being named in an AI answer was never going to show up as one.
Those figures do not share a denominator and will never sum to a hundred. What they tell you is which path gets walked most often: the first one dominates on ChatGPT, simply because the overwhelming majority of answers hand the reader nothing to click.
All three paths end in the same place. Your analytics attribute the work to Google and to direct traffic, and they do it with total confidence, because the referrer field really does say google.com.
How much of your direct traffic came from AI?
Nobody can calculate it, and that includes you.
Go back to your analytics and look at the direct bucket. Whether it sits at thirty percent or sixty, that bucket is a junk drawer — regulars with a bookmark, links opened inside an app, URLs someone copied and pasted, QR-code scans, all piled into one row with no field that separates them.
This is not a backlog problem or a tracking-setup problem that some better tool will fix. The number is structurally unavailable: a field the browser never sent cannot be reconstructed on the server afterwards.
So when a vendor tells you they can report your “share of traffic from AI,” ask them one question: how did you attribute the sessions with no referrer? The answer is almost always a model estimate. Estimates are fine for watching a trend and unfit to be quoted as a number.
Ask it, and their reaction is the best filter available to you. An honest one will tell you the figure is modelled, name the assumptions behind it, and mark its own margin of error before you have to ask. The one who hands you a clean percentage, calls it measurement, and cannot explain how it was derived is not merely being sloppy. They are lying to you. Google cannot obtain this number, and your vendor does not happen to be the exception.
Google gave us 14%. Every other engine gives nothing.
Google Search Console’s generative AI performance report is the only official number in existence right now.
It matters less for its own sake than for being the only such number that exists anywhere. It covers only Google’s own AI surfaces. It reports impressions with no clicks, no click-through rate, and no query strings — though by the logic above, impressions are the honest unit for this anyway. And Google is the only company publishing it at all: ChatGPT, Perplexity, Grok, DeepSeek and the other major AI engines give you no console whatsoever. How often you get mentioned there, how you get described, which competitors you get listed alongside — officially, zero data.
Put those two facts side by side. The one gateway willing to show you a number has already conceded 14%. The gateways that show you nothing add up to an amount no one can state. That polished traffic report on your desk is measuring a world that keeps getting smaller.
The only remaining way to get a number is to ask
Since nobody is going to deliver this data to your analytics, the alternative is to go and interrogate the engines yourself.
The method is not exotic. Assemble a set of questions your customers would actually type — not the ones you wish they typed — hold them fixed, run them against every engine on a regular cycle, and record whether you appear, how early you appear, what adjectives get attached to you, and who else shows up in the same answer. This is sampling, and it works the way political polling works: you never reach everyone, but with fixed questions, a fixed cadence and enough runs, the trend line is real.
It is also the only visibility number we track for ourselves. We still read the referrer report — it tells us whether the site is healthy. It no longer tells us anything about AI visibility, because it cannot.
Why this is hard to run yourself
You can absolutely open ChatGPT right now and ask one question to see whether you show up. Do it — you will come away with a feel for where you stand.
The difficulty starts after that. Ask the same question three times and the answers drift, so a single run is an anecdote rather than a measurement. Eight engines have to be asked separately, because they source material differently and leading on one says nothing about the others. It has to repeat weekly before a trend exists. And competitors have to be measured in the same pass, because visibility is relative — a flat score while a rival climbs means you are losing ground.
Then comes the genuinely hard part. When a score drops, someone has to determine whether your site broke or the engine changed its rules. Those two diagnoses call for opposite responses, and getting it wrong means spending a quarter repairing something that was never broken.
One manual test gives you a snapshot. What this needs is a continuous line, plus somebody who can read what the line is saying.
One thing to do this week
Open the generative AI features report in your own Google Search Console and find your percentage.
Then work out what that figure is saying, and what it is not. It says how much of your Google exposure already happens on AI surfaces — and since those impressions were already inside your totals, it measures how far the shift has progressed, not how much new channel you have gained.
What it does not say is how you are faring on ChatGPT or Perplexity. Those are different denominators, and the first does not bound the second. Only the degree of mediation is certain: Google remains the most conservative gateway of them all, still setting conventional results beside its AI answers. A pure chat interface offers no blue link to fall back on — the user asks once, receives one answer, and whether you get named is entirely the model’s call. So the share of discovery that is AI-mediated can only be higher there.
Whether you get praised or skipped in those answers, this number says nothing at all — and it may well be worse than your Google showing, because the two draw on different material.
To see how AI currently describes you, run an analysis at geoweb.tw. To have someone track your multi-engine trend line against your competitors — and judge whether a movement warrants action — take a look at managed GEO.