Transcript
Intro
One percent. That is how often people clicked a link inside Google's AI summary, in a study that followed the real browsing of nine hundred American adults for a month. Cue a fresh wave of posts announcing that SEO is dead.
Those posts skip the next two lines of the same data. People still clicked a normal result on eight percent of searches that had a summary on top. And sites that get cited inside the AI answer do clearly better than sites that get left out.
Here is the frame for this whole part of the course. AI search is not one machine. It is several different retrieval systems behind similar chat boxes, and each one hands out clicks differently. Today we measure where the clicks went, what a citation is worth, and three claims that need a verdict.
The click studies, as a range
Start with real people. Pew Research logged almost sixty nine thousand Google searches in March 2025. With no AI summary on the page, people clicked a normal result fifteen percent of the time. With a summary, eight percent. A link inside the summary got one percent.
Now the scary one. Ahrefs looked at three hundred thousand keywords. When an AI Overview showed, the top result got fifty eight percent fewer clicks than you would expect. That is desktop data from December 2025, against a baseline from before AI Overviews existed.
It is a real number from a credible method, and it gets quoted as if it were the whole story. It is one number, from one method, on one kind of query. Keep that in your pocket for a minute.
Then comes the plot twist. A study tracking fifty three brands found that the organic click rate on AI Overview queries fell from about three point two percent in January 2025 to one point three in December. By February 2026 it had bounced back to about two point four. Nobody's headline predicted the bounce.
And one big study points the other way entirely. Semrush took keywords that gained an AI Overview between May and October 2025. On those exact keywords, searches ending with no click went slightly down, from about thirty four percent to thirty one and a half. Their reading: AI Overviews land on questions that already got few clicks, so part of the drop we blame on AI was there before it arrived.
So who is right? All of them, about different things. A panel follows people across every search they make. A click study follows keywords, often informational, often on desktop. A zero-click study counts any click at all, ads included. The mistake is borrowing one study's number to answer another study's question.
Put them on one timeline with dates and you get something you can plan with. AI answers cut clicks on a lot of informational searches, anywhere from a small dip to more than half, depending on the query and the month. Anyone quoting you one number without a date is selling something. Usually a tool. Occasionally a panic.
So next time a scary click stat lands in your inbox, ask it three questions before you forward it. What exactly was measured, people or keywords? When was the data collected? And on which kind of queries? If the post can't answer all three, it is a vibe with a percent sign on it.
Cited vs not cited
Now the number that matters if you want to win. That fifty three brand study also split informational queries by whether the brand was cited in the AI Overview. Cited brands averaged two point zero seven percent click rate across 2025. Brands left out got zero point nine four. Queries with no AI Overview at all still beat both, at three point three five.
Before you put "twice the clicks" on a slide, read the study's footnote. One big account makes up almost half of the not-cited group, with a tiny click rate. Take it out, and that group rises to about one point six. So the fair gap is closer to a third more clicks than double. Still a clear win for being cited, just not a miracle.
The paid side is a handy sanity check, and it tells the same story. Ads for cited brands got about fifteen point seven percent click rate. Ads for brands left out got about eleven point two. No AI Overview, about nineteen point eight. Same order every time.
Two takeaways. If an AI answer shows on your query, being inside it beats being left out. And being inside it still does not take you back to the old world with no AI answer. One caveat we will actually make a fuss about today: this is correlation. Brands that get cited may simply be stronger brands that people click anyway.
So how do you get in? Google's own documentation is almost boring about it. The page has to be indexed and eligible to show with a snippet, and in Google's words there are no additional technical requirements. Both features may run several related searches behind the scenes to build one answer. The road into the AI answer still runs through ordinary search.
How much traffic AI sends
Flip the question around. How much traffic do AI assistants actually send to websites? Contentsquare looked at six and a half thousand sites and found zero point two percent of visits in late 2025. Conductor, measuring enterprise sites in mid 2025, found about one percent. So the honest range is roughly zero point two to one percent of a normal site's traffic.
If that sounds like a rounding error, two things. By Contentsquare's count it grew six hundred and thirty two percent during 2025, which is more than seven times bigger in a year. And clicks from Google's own AI Overviews and AI Mode never show up as AI traffic. Google reports them as normal search traffic. So the real share is small, growing fast, and partly hiding inside your search numbers.
Answers get absorbed, tools get clicked
Why are the clicks so small? Let's watch it happen. Same two questions, five places: Google's AI Overview, AI Mode, Claude, ChatGPT and Perplexity. Question one is a plain fact: how long is the film Oppenheimer. Question two needs something live: where can I legally watch Oppenheimer in the UK this week.
Question one. Every engine answers it in a single line, and nobody needs to click anything. Whichever site supplied that running time got used and got nothing back. Now multiply that by every definition, every date and every five-step how-to on the web. That is where most of the lost clicks went.
Question two behaves differently. What is streaming changes by country and by week, and the person still has a job to do: pick a service and press play. The answer can point, but it cannot finish the job. That is when the links inside an answer start earning their keep.
We checked this on an education site we run, using its server logs, over two weeks in September 2026. Topic guides got zero point eight visits for every hundred times an AI assistant fetched them. An interactive test got twenty three point five. Same direction every study above points: answers get read in the chat, tools get clicked.
So here is the rule we work by. Answers get absorbed, tools get clicked. If your page's whole value fits in a paragraph, an AI answer can carry that paragraph away. If your page lets someone check something live or decide with their own numbers, the answer has to send them to you.
Take ReelRadar, our fictional film site. Its page on a film's running time is pure fact, so an AI answer will happily carry it off. Its where-to-watch checker, which looks up your country and the services you already pay for, is a do-something page. Same site, same film, very different exposure to AI answers.
What do you change on Monday? Find the pages that only answer a fact, and ask what a reader could do next that a chat window can't. Add that where it genuinely helps. Keep the fact itself clear and early, because being the cited source still beats being skipped. And stop judging those pages on clicks alone.
Three claims, three verdicts
Time for three claims you will hear this week. Claim one: AI killed SEO, so stop investing in search. It gets said with great confidence, usually in a post that is itself trying to rank.
In July 2026, Google told investors that AI Mode had passed a billion monthly users and that its AI features send billions of clicks to websites every week. That is Google grading its own homework, so weigh it. Pew's panel points the same way, though: people still clicked normal results with a summary on the page. And the way into the AI answer is the ordinary index. SEO changed shape. It did not die.
Claim two: AI traffic converts twenty three times better than search, so it is worth more than everything else. That one shows up in a lot of pitch decks.
The number is real. It comes from one software company measuring its own website in June 2025, where AI search was half a percent of traffic and about twelve percent of signups. One site, selling a marketing tool, to people who ask chatbots about marketing tools. Not exactly a random sample.
Broader data shows a smaller gap. Adobe's US retail figures for July 2026 had AI visits converting sixty percent better than other traffic. Better, often. Twenty three times, not a number to build a budget on.
Claim three: get cited by AI, and the traffic you lost will come back. This is the hopeful one, and it is popular with people who sell citation tracking.
Links inside Google's AI summary got clicked on one percent of visits. Assistants send somewhere between zero point two and one percent of site traffic. A citation is exposure, and it beats being left out. It will not refill a traffic chart on its own, so report citations as brand visibility and clicks as clicks.
Measure your own exposure
Instead of arguing about other people's studies, open your own. Since the thirty first of August 2026, every Search Console property has generative AI performance reports. They show how often your pages appeared in AI Overviews, AI Mode and Discover's AI features, by page, country, device and date. Impressions only, no clicks and no queries. It tells you where you are visible, not what that is worth.
The useful move is to put that page list next to your normal clicks for the same pages. A page whose AI impressions keep rising while its clicks stay flat is a page whose answer is being read inside someone else's box. That is your shortlist for the do-something layer, and episode four builds the full tracker around it.
There is also a new off switch under Settings, called the Search generative AI control. It pulls your links and content out of AI Overviews, AI Mode and Discover's AI features. Google says it is not a ranking signal anywhere else in Search, and it does not touch AI training, which is Google-Extended's job.
Think hard before you flip it. If the answer itself is your product, it might make sense. If you need to be recommended, it almost never does, and the next episode covers the other switches people confuse it with.
Then run the worksheet. Export your top queries from Search Console and sort each one into a class: a plain fact, a how-to, a comparison, or a do-something query, where the person needs a tool, live data or a purchase. Weight the classes by clicks. Facts and how-tos are your exposure. Do-something queries are the part an answer has to send to you.
Sorting a few thousand queries by hand gets old fast. We hand the export to Claude Code, which reads the file, writes the class column back and prints the split. Any capable assistant can do the sorting. Whichever you use, check a random handful of labels yourself, because one wrong class moves the whole answer.
Wrap up
The episode on one screen. AI answers cut clicks, by an amount that depends on the query and the study. Being cited beats being left out, and still sits below no AI answer at all. Assistants send a small but fast-growing share of visits. And the pages that keep getting clicked are the ones that let people do something.
The click studies will change every quarter, and so will Google's reports, which is why every number here carries its date. The pattern underneath has held across all of them. The query-class worksheet is free on the episode page at irongoo.com. Run it on your own Search Console before you believe anyone's headline, ours included.
