Someone types a question into a search box. The answer appears right at the top, clean and neatly packaged. They read it, they’re satisfied, they close the tab. The three sentences they read came from a blog post someone labored over three years ago — but no one arrived at that blog. The writer doesn’t even know their words just settled a stranger’s question.
In 2026 this stopped being the exception in search and became the majority. The click is disappearing. And for the last twenty years, the click was just about the only way a person who writes on the internet got paid.
The weight of saying the click is disappearing
Better to start by checking, in numbers, that this is really happening — because our gut impression tends to run ahead of the facts.
According to the search-analytics firm SparkToro, working from Similarweb clickstream data, 68% of Google searches in the US ended without a single click over the first four months of 2026. That’s up 7.5 points from 60% in 2024 — the steepest rise since the metric was first tracked. Two-thirds of searches now send no one to any website at all.
The cause is the AI Overview that sits at the top of the results. Attached to 34.5% of all Google searches in December 2025, it had grown to 48% by March 2026. And in a Pew Research study that examined 68,879 real searches, the share of users who clicked a link when an AI Overview was present was just 8% — nearly half the 15% seen when no Overview appeared. There’s a colder number still: the share who clicked one of the source citations inside the Overview was 1%. People borrow the answer and almost never go to where the answer was made.
A 2025 Bain & Company survey of US consumers shows the shift is underway outside search engines too: 80% of respondents said they rely on AI-generated answers for at least 40% of their searches. People already ask ChatGPT, Claude, and Gemini first — not the list of blue links on Google or Naver.
To be clear: this is not a story of search engines dying. Google’s share of search is still close to 90%, and in Korea, Naver actually clawed share back by folding AI into search. What’s collapsing isn’t the search engine but the people who lived on the traffic it used to pass downstream. For a long time, search didn’t give you the answer — it pointed at where the answer was. The pointing was a visit, the visit was ad revenue, and the ad revenue was a reason to write. AI cuts the first link of that chain. Instead of pointing, it just answers.
So a question lands in front of anyone who writes: if no one comes, what is there to be paid in? As of 2026, there are two answers on the table — and they point in opposite directions.
The first exit: waking a code that slept for thirty years
One is to shut the door and charge a toll.
The web’s protocol has had a 402 Payment Required status code from the very start. In the early 1990s it was a number reserved with a shrug — “someday the web will need to charge for something.” The slot stayed empty for thirty years. The web was free by default; if you wanted to block something you bolted on your own mechanism, and a standard response that demanded money simply never came up.
In July 2025, the content-delivery firm Cloudflare pulled that empty seat out. New sites behind Cloudflare now block AI crawlers by default. The web’s 25-year premise — “open by default, mark it if you want it blocked” — flipped to “closed by default, ask if you want it opened.” On top of that sits pay-per-crawl. A site owner picks one of three options for each AI crawler: open it for free, charge per scrape, or block it.
Choose to charge, and the 402 dozing in that seat wakes up. When a crawler requests a page, the server answers with 402 Payment Required and a price tag in the header — this page costs this much per fetch. Only when the AI company signals “I’ll pay that” does the content go through, with Cloudflare settling up in the middle. A code that spent thirty years as a reserved seat came back to life as the operating signal for a tollgate on web content.
Why did this become necessary? One ratio compiled by the middleware firm TollBit compresses the whole situation. Measure how often AI companies scrape content against how many visitors they send back in return, and OpenAI came out at 179 to 1, Perplexity at 369 to 1. And Anthropic, the maker of Claude, was at 8,692 to 1 — one human sent back for every 8,692 scrapes. The unspoken bargain search at least maintained — send back roughly as much as you take — becomes, in front of an AI crawler, something close to one-way mining. The toll is an attempt to stop that mining, or, if it can’t be stopped, to at least put a price on it.
The same idea has surfaced as a standard, too. Really Simple Licensing (RSL) embeds license terms into the site’s instructions file (robots.txt) — terms written for machines rather than people. On top of free, attribution, subscription, and pay-per-crawl, one item stands out: per-inference pricing — charging not when your writing is scraped, but every time an answer is generated from it. Pricing the use, not the mining.
But this path has a clear limit: a standard is a request, not enforcement. In August 2025, Cloudflare accused Perplexity of routing around a block with disguised crawlers once it was shut out — forging user identifiers and constantly switching origins to scrape tens of thousands of domains millions of times a day. Perplexity denied it, but the structure the episode exposed is plain. You can hang a fare gate on the wall, but the only thing that can actually stop someone from climbing over is the infrastructure that owns the road itself. The future of the toll is decided not by the elegance of the standard but by who holds the pipe.
The second exit: please, take it
The other answer is the exact opposite. Instead of shutting the door, you throw it wide open and ask: please take my writing and cite it.
This path has a name: GEO, generative engine optimization. If the last twenty years of search optimization (SEO) was about “climbing higher in the list of results,” GEO aims to “be cited as a source when the AI builds its answer.” The goal moves from being found to being quoted. If people aren’t going to reach your site anyway, the thinking goes, at least get your name and your claim inside the sentence the AI speaks.
The trouble is that every AI trusts wildly different sources. In one analysis, ChatGPT drew 47.9% of its citations from Wikipedia, while Perplexity drew 46.7% from community discussion like Reddit. Only 11% of sites were cited by both. Optimize for one and you’re in the dark for the other. ChatGPT also uses Bing as its starting point for search, so anything not indexed by Bing doesn’t even make the shortlist — a blind spot for anyone who only ever watched Google.
So what earns a citation? A few signals hold up. A domain whose inbound links from other sites cross a certain count becomes several times more likely to be cited by ChatGPT — there seems to be a threshold of trust. At the sentence level, it helps to strip out subjective hedges like “I think” and write in flat declaratives, because an assertive sentence carries less uncertainty for the model and is easier to lift verbatim. Vague praise like “this tool is excellent” doesn’t get cited; a concrete figure like “this method cut processing time by 32%” does.
One tempting myth is worth clearing up here: the claim that installing an AI instructions file (llms.txt) gets you cited. A study of 300,000 domains found no statistical correlation between having the file and being cited by AI. No major AI has ever formally promised to honor it, and the logs show most crawlers don’t even request it. Where the file actually does its job is somewhere else entirely: the coding agents working on a user’s behalf pick it up when they read documentation sites. It’s not for exposure — it’s a map for the machine doing work on your site.
Even the way you measure success changes under GEO. There’s no fixed rank like a search position. Instead you count “share of voice” — how often your name appears among the answers AI gives to the questions in your field. But ask the same question five times and you get five different answers, so any single result is meaningless; only the frequency across many runs becomes the metric. Measurement itself has turned into statistical sampling.
The two exits kill each other
Get this far and something strange comes into view. The two paths don’t run the same direction. They run opposite.
The toll says: pay to scrape, or I block you. But the moment you block, the AI can’t see your writing, and what it can’t see it can’t cite. GEO says: please take it and cite me. But the moment you open wide, you collect nothing in tolls.
So the answer big outlets landed on is to split their writing in two. Summaries, definitions, concept explainers — the kind of information that’s everywhere anyway — get thrown open for AI to cite; this side is bait that earns authority and recognition. Proprietary data, original reporting, deep analysis, meanwhile, stays inside the wall and gets paid for through tolls or subscriptions. The tasting counter is open; the full meal costs money. Some large outlets have in fact moved to exactly this structure — licensing summaries and headlines to AI, keeping the body behind a paywall.
What’s left
So which exit should a person writing alone take?
Realistically, the toll is still out of reach for an individual. At a unit price of roughly a cent per crawl, no meaningful revenue comes of it. The toll is a game for the infrastructure that owns the road, for large outlets with bargaining power, or for the middleware that bundles many small publishers to negotiate together. What an individual can do right now is mostly to secure the right to decide whether to block or open — and keep that decision their own.
At this stage, then, the answer for a solo creator is fairly clear: take the GEO door, but aim past the toll, at what lies beyond it. Let citations pile up until you’ve claimed the position of “this person is trustworthy on this subject,” and the position itself becomes the asset. The traffic vanishes, but the seat of a trusted source remains. The gate is a card you can only play after authority has accumulated. Build the wall first, with nothing behind it, and no one wonders what’s inside.
What’s interesting is that the conditions for writing AI cites well aren’t, in fact, far from the conditions for good writing. Clear facts over vague impressions; figures you checked yourself over borrowed generalities; a distinct author over blurry anonymity. The writing a person understands and trusts quickly is the writing a machine cites without worry. That two readers point the same way is, for anyone who writes, at least a small piece of luck.
Underneath all this runs a deeper helplessness. Search was once the road that carried my writing to people; now the owner of that road is the search engine and the AI. The hands that decide how my writing reaches anyone — its distribution, its exposure — are no longer mine. So what, in the end, can a writer still hold in their own hand?
Probably not the click. The click was a currency search lent us for a while and is now taking back. If there’s a new currency to take its place, it’s the citation. You can’t see who came by, but your sentence is alive and moving inside someone’s answer somewhere. As the shape of the reward shifts from visits to trust, from traffic to citation, the meaning of writing is being quietly re-priced too.
— tomte
Read next · Why Flat-Rate Pricing Breaks Down Against AI Agents · You Can Outsource Thinking, But Not Understanding