You want a recipe for lentil soup. What you get is a 1,400-word memoir about the author’s grandmother, a farmhouse in Umbria, and the precise quality of afternoon light in a kitchen you will never visit, and somewhere below the sixth photograph of a bowl, the actual instructions. This is the most-mocked genre on the internet, and the joke usually lands on the writer — the narcissism of food bloggers, the audacity of making you scroll. But the blogger didn’t choose the memoir. The memoir was chosen for her, by a ranking system that rewarded long pages, fresh pages, pages stuffed with the words people type. She was filling out a form. The childhood anecdote is what the form looks like when a human being has to live inside it.
That distinction matters, because the standard story about search’s decline is a moral one — greedy SEOs, lazy Google, content farms ruining a good thing — and the real story is structural. Nobody designed the recipe-memoir. It precipitated out of a feedback loop: the algorithm rewarded a format, the web’s economics made producing that format the rational move, and so the web produced it industrially, at a scale no ranking tweak could keep ahead of. The results page filled with pages built to be retrieved rather than read. Search didn’t stop working. It started working perfectly on a web that had reorganized itself around being searched.
The loop ran for roughly fifteen years, and by the early 2020s its output was the web we all learned to route around — appending “reddit” to queries, treating the first page of results as a toll plaza to pass through quickly. Then AI answers arrived, sold explicitly as the fix for the clutter. And they are a fix, in the narrow sense that they work. What’s less visible is that they complete the loop rather than break it.
The mechanism is worth stating plainly. The answer on top of the results page is synthesized from pages the user never visits. The visit was the payment. Search engines and the open web had an unwritten arrangement: Google indexed your pages, and in exchange sent you readers, whose attention you could sell. The AI summary keeps the first half of the deal and quietly cancels the second.
The numbers, where they exist, are stark. When Pew Research Center analyzed the actual browsing of 900 US adults in March 2025 — nearly 69,000 real searches — it found that users shown an AI summary clicked a traditional result in just 8% of visits, versus 15% when no summary appeared. Links cited inside the summaries were clicked in 1% of visits. More than a quarter of sessions that produced an AI summary simply ended there, answer received, web unvisited. A year later, clickstream data assembled by Rand Fishkin’s SparkToro and reported by Search Engine Journal suggested that for every 1,000 US Google searches in early 2026, only 232 clicks reached the open web at all — down from 360 per 1,000 in his 2024 baseline — with 68% of searches ending in no click whatsoever. Similarweb estimates put worldwide search referrals down roughly 15% year over year, and the Wall Street Journal, citing Semrush data, reported in July 2026 that some publications had lost more than 40% of their Google traffic in twelve months.
These figures describe the pipe narrowing, and the pipe was the business model. Nilay Patel coined the term “Google Zero” back in 2024 for the moment the referrals stop, and what the data since then shows is not a cliff but a gradient, steepening. The Independent Publishers Alliance filed an antitrust complaint with the European Commission over AI Overviews in July 2025; Cloudflare, which says bots now account for more than half of all web traffic, moved to block multi-purpose crawlers by default in September 2026. These are the maneuvers of an ecosystem renegotiating a contract that was never written down.
It’s tempting to cast this as extraction — the platform eating the garden it tended — but that framing smuggles in a villain the structure doesn’t need. Google’s AI answers are a rational response to a results page its own ranking incentives had already degraded; the publishers’ dependence on search referrals was a rational response to where the readers were; the recipe memoir was a rational response to what ranked. Every actor in the loop behaved sensibly, and the sum of their sensibility is a web where the answer layer consumes the page layer that feeds it. Nobody chose this. That’s precisely the problem — systems that nobody chose are systems nobody can simply choose to stop.
Which leaves a question that has no comfortable answer, and shouldn’t be given one. The open web’s corpus — the recipes, the explainers, the forum threads, the local news — exists because producing a page used to pay, in traffic if nothing else. If the visit disappears, the payment disappears, and the index the AI answers are synthesized from becomes a record of an economy that no longer operates. New pages will still be made, of course. But made by whom, and for what return? The likeliest candidates are the actors who never needed readers in the first place: companies writing for the answer engine itself, models trained to please models, a web that is read mostly by machines and quoted to humans secondhand.
The recipe blogger’s grandmother, at least, was real. The open question is what fills the index when there is no longer any reason for her to be.