AI Search Is Repricing the Open Web

AI search is draining the web traffic that publishers (news sites, blogs and other media that live on advertising) depend on, while quietly helping the businesses that sell things. That split is what this fortnight's consensus misses. After Google made AI Mode the default search experience for everyone at its developer conference on May 19, the traffic sent to publishers kept falling, while the traffic that turns into sales grew fast and was worth more. The web that is dying is the ad-funded web; the web that is growing belongs to the businesses an AI can read and recommend.

Give the consensus its strongest case first, because it is a strong one. At I/O, Google's annual developer conference, on May 19, the company rebuilt search around AI Mode and made Gemini 3.5 Flash its default, the largest change to the product in 25 years by Google's own description. It said AI Mode had crossed a billion monthly users, with queries more than doubling each quarter, and previewed AI-built results that assemble custom tables, charts and simulations directly on the results page, alongside background agents that read across the web so the user need never open anything. Time and most of the trade press drew the same conclusion: your link is now raw material the AI reads and summarises before the reader ever decides whether to click.

The casualty data is real and specific. Chartbeat, measuring more than 2,500 publisher sites, found Google search traffic to publishers fell 33% globally in the year to November 2025. The damage is uneven: in Chartbeat figures reported by Axios in March 2026, small publishers running 1,000 to 10,000 daily page views lost 60% of their search referral traffic over two years, medium publishers lost 47%, and large publishers 22%. The mechanism is documented. A Pew Research Center study published in July 2025 found that when an AI summary appears, users click a traditional result in 8% of searches, against 15% when there is no summary, and click a link inside the summary just 1% of the time. AI answers have only spread since. The Reuters Institute's 2026 survey of 280 media leaders found most expect search referrals to fall by more than 40% over the next three years. This is the hollowing we traced in our piece on defensibility in the AI era, where Stack Overflow's monthly questions collapsed from about 200,000 to under 7,000, its level at launch in 2008. If your business rents attention by the click, the consensus has the story right: this is a direct hit to your revenue.

Now the half the panic leaves out. Adobe's Digital Insights team, in its quarterly AI traffic report published in April 2026, found that traffic from AI sources to United States retail sites rose 393% year over year in the first quarter of 2026, after a 693% jump across the 2025 holiday season. The quality matters more than the volume. In March 2026 that AI-sourced traffic converted 42% better than non-AI traffic, a record, while those visitors spent 48% longer on site, viewed 13% more pages, and generated 37% more revenue per visit. Adobe is candid that AI is still a small share of total retail visits, so it is not yet refilling the volume publishers lost. But the same person a publisher counts as a lost click, a retailer counts as the best-qualified buyer on the site.

Both data sets are true because the word traffic means two different things. A publisher sells attention and is paid per impression, so it needs raw volume, and raw volume is exactly what an AI summary removes. An operator sells a product or a service and needs qualified buyers, which is exactly what an AI assistant delivers, because the assistant has already done the comparison and hands over the user at the moment intent is highest. What AI actually killed is the cheap middle: the undifferentiated page that ranked for a keyword, harvested a click, and monetised it with a banner before the visitor bounced. That page was always a tax on the reader. Removing it reprices the web around something more valuable.

The lens is the one we used in Jevons Won't Save the Pipeline You Just Broke, turned from labour toward attention. When asking a machine a question becomes cheaper and more conversational, people ask more of them, and Google's own claim that AI Mode queries are doubling each quarter is that effect in action. So three things move at once: total query volume rises, the raw link click falls, and the slice of high commercial intent that lands on a site the model can actually read grows. Treating the falling click as the whole story is the error.

That sets a specific playbook. In the same April report, Adobe found that roughly a quarter of the content on retailers' homepages was not machine-readable for large language models, and product and category pages were worse. That gap defines the discipline now replacing search-engine optimisation (SEO): making your pages legible and quotable to the model that sits between you and the buyer. Structure your product data, make your claims extractable, and treat brand as the trust shortcut a buyer reaches for when they never see your page before the model describes it, the brand-as-trust dynamic we mapped in the defensibility piece. The people who should be uncomfortable are ad-funded publishers, agencies still selling volume-based SEO retainers, and anyone whose moat was a ranking. The ones who can hold their position are operators with a real product, a name a model has reason to trust, and pages a model can read.

So read the I/O announcements as a repricing of the web, and respond to it as one. The traffic AI strips out was the low-intent, ad-monetised kind that was never worth much to anyone selling something real. The traffic it adds converts 42% better and stays longer once it arrives. The losers are businesses whose only asset was ranking for a query; the winners are businesses an AI can read and has a reason to recommend. Lost search traffic is not coming back in its old shape, so winning it back is the wrong goal for the next two quarters. The real question is simpler and harder: when a buyer asks an assistant about your category, are you the answer it can quote and the product it suggests? If you cannot answer that with confidence, that uncertainty is the most useful thing this fortnight handed you.