Every couple of years, somebody publishes SEO’s obituary. This time the argument looks stronger than usual: people really do ask ChatGPT things they used to type into Google, and Google itself answers half the questions on the results page. “SEO is dead” is trending again.
The data, however, says the opposite of what the obituary writers want. Search didn’t die — the click died, and it was replaced by something that needs SEO more, not less: the cited source.
In short
- 68% of Google searches now end without a click — SparkToro and Similarweb measured 68.01% in early 2026, up from 60.45% in 2024.
- AI referrals are the fastest‑growing channel on the web: +357% year over year, reaching 1.13 billion visits in June 2025 (Similarweb).
- An AI Overview cuts position‑1 clicks by ~58% (Ahrefs, 2026) — so ranking first buys less than it used to.
- The strongest correlate of brand visibility in AI answers is branded web mentions (0.664) — backlinks score just 0.218 (Ahrefs, 75,000 brands).
- GEO methods boosted AI visibility by up to 40% in the Princeton study — quotable writing is a measurable lever.
- Shortcuts don’t work: 97% of llms.txt files are never fetched. Fundamentals do.
What the numbers actually say
Start with the scariest number. SparkToro and Similarweb’s 2026 study found that 68.01% of US Google searches ended without a click in the first four months of the year, up from 60.45% in 2024. Two‑thirds of searches now resolve on the results page itself.
Zero‑click is not a new phenomenon — it has been the majority outcome for years, because so many queries are answered by a map, a calculator or a knowledge panel. What changed is the rate: AI Overviews pushed it to a record, at what SparkToro calls the fastest two‑year acceleration they have measured.
Now the counterweight. While classic clicks shrink, AI platforms send 1.13 billion referral visits to the top 1,000 websites monthly — up 357% year over year as of June 2025. That channel is still small next to Google’s volume, but it compounds at a rate nothing else on the web does. And the visitors it sends are different: Semrush’s study found the average AI‑referred visitor converts at 4.4× the rate of a traditional organic visitor. Fewer doors, but the people who walk through them came with intent.
What actually changed: from ranking pages to being the cited source
Here is the mechanism behind both trends, and it is simpler than the panic suggests.
AI answers are not magic. An AI Overview or a ChatGPT answer is assembled from sources the system could crawl, render and parse cheaply — and it cites them. When an AI Overview does appear above the results, Ahrefs measured that the #1 organic result loses ~58% of its click‑through rate compared with the pre‑AI baseline (December 2023 versus December 2025, 300,000 keywords). Ahrefs’ earlier 2025 study put the same effect at −34.5%; the direction is one‑way and the number keeps growing.
So the top blue link has become a pay‑cut position. But look at who does get traffic and visibility: the sources the AI quotes. Being cited is the new being ranked. That reframes the whole discipline — you are no longer optimizing to win a click auction; you are optimizing to be the fact a machine repeats with your name attached.
One thing worth stating plainly, because it anchors everything else: search is not a vibe — it is information retrieval, an exact discipline with some seventy years of mathematics behind it. Every system in this story, classic or generative, stands on the same floor: an inverted index that maps words to documents, a retrieval stage that fetches candidates, a ranking stage that orders them. The LLM era added a generation step on top — but a generator still needs something retrieved to ground itself on, and it grounds itself on what was found. That is why crawlability, indexation and clean structure did not become old‑fashioned; they became the entry ticket to being citable at all.
Five IR concepts that explain AI search
Words like “embeddings” and “RAG” sound like an initiation ritual, but there are concrete mechanisms behind them. Five are enough to read any “AI SEO” news and separate mechanics from mantras.
Vector search and embeddings. Models can turn text into a vector — a point in a high‑dimensional space where proximity means semantic similarity, with cosine similarity as the usual measure of distance. One consequence reshapes half the familiar advice: a page can answer a query that shares none of its words. Exact‑match keywords are still there — they are just no longer the only door in.
Information gain. A measure of how much a document adds to what the system already knows. A retelling of common knowledge yields a near‑zero gain, and in generative search that is fatal: the answer is assembled from passages, and a duplicate adds nothing to it. Bing’s engineers describe information gain as the criterion for entering the generative index: what gets in is what the index doesn’t know yet, not another version of the same text.
RAG — retrieval‑augmented generation. The working scheme behind ChatGPT with search, Perplexity and AI Overviews: first retrieve candidates, then generate an answer from what was found. The generator is fed not whole pages but passages cut out of them. That sets a direct requirement on how you structure meaning: one idea per section, self‑contained paragraphs, zero filler. Well‑structured text is chunked into pieces that can be quoted whole; poorly structured text is cut into scraps the machine has to guess around.
Knowledge graphs and entity SEO. A large share of queries are about entities: a company, a product, a person. Search engines have spent decades building a graph that connects a brand to its industry, geography, authors and products. Schema.org markup and unambiguous wording feed that graph, and mentions across the web corroborate its edges from the outside. A brand wired into the graph is, to a machine, an established fact; a brand living only on its own website is an unsupported claim.
BM25 and transformers — together, not in a “which one wins” mode. BM25, a lexical ranker from the nineties, still performs the first pass over billions of documents: it is cheap and precise on word matches. Transformers re‑rank what survived — this time on semantic closeness. That is hybrid search as it runs in production: a sparse index catches exact terms, a dense index catches paraphrases, and a model assembles the final ordering. The “which one wins” question can be retired; both layers work at once.
None of the five hands you a new checklist. They all converge on one point: machines search among documents that have something to say, and cite what is easy for them to parse.
What GEO adds — and what it doesn’t
GEO (Generative Engine Optimization) — also sold as AEO, “answer engine optimization” — is the practice of making your content easy for AI systems to lift. It sounds like hype, but it has real evidence behind it. The Princeton‑led GEO paper (GEO: Generative Engine Optimization, Aggarwal et al.) tested content changes across thousands of queries and found that adding quotations, statistics and clearly attributed sources improved visibility in generative engines by up to 40%. Not bad for rewriting how you state facts.
The second lever is off‑site. Ahrefs studied 75,000 brands’ visibility in AI Overviews and found that branded web mentions — your name appearing across the web — correlate with AI visibility at 0.664, while backlinks manage just 0.218. AI systems trust consensus, and consensus is built by being talked about: directories, press, lists, communities. Link building is not dead, but it is no longer the strongest signal in the room.
And the cautionary tale, because every new channel breeds shortcuts: llms.txt was pitched as a magic file that tells AI systems what your site is about. Ahrefs checked 137,000 websites and found that 97% of llms.txt files received zero requests in May 2026. AI bots don’t go looking for it. The shortcut did nothing; the underlying content infrastructure does everything. (We take apart the whole file layer — robots.txt, sitemaps, llms.txt — in a separate guide.)
What still works — and works better than before
Here is the part the obituary writers miss: almost everything AI search consumes is what classic technical SEO already produced.
- Server‑rendered, fast HTML. LLM crawlers spend serious compute generating answers and many won’t wait for your JavaScript to hydrate. If the text isn’t in the first HTML response, for many AI systems you don’t exist. This has been true for Googlebot for a decade; AI just raised the stakes.
- Structured data. Schema.org markup turns probabilistic guesswork into facts a model can lift verbatim — prices, specs, authors, FAQ answers.
- Clean semantic structure. One idea per section, descriptive headings, no keyword confetti. Models parse structure; they don’t reward noise.
- Quotable statements. Clear sentences with numbers, dates and named sources — the exact features the GEO paper found worth up to 40%.
- Off‑site presence. Mentions in directories, articles and communities, which the correlation data says outweigh backlinks for AI visibility.
- Speed, still. Every crawler — human or robotic — rewards a fast site. Our whole architecture exists because of this (see why speed matters).
We covered the technical foundation in detail in what technical SEO actually buys you. Nothing in the AI era invalidated it — the AI era made it load‑bearing. And for the behavioral half of the story — what the courts proved about NavBoost, and where click signals are heading as the click pool shrinks — see NavBoost and user interaction signals.
Where this leaves a business in 2026
The false choice is “Google or AI”. There is no fork in the road: one content infrastructure feeds both. A fast, server‑rendered, structured, factually rich site ranks in classic search and gets cited in AI answers. A slow client‑side app with marketing fluff loses both races at once.
That is why we build SEO and AI visibility in from the first commit rather than bolting it on after launch — it is one and the same work now. If you want that approach applied to your site, that is exactly what our SEO & AI Search service covers, from technical audits to GEO‑ready content.
Summary
SEO didn’t die; the click died. Attention moved from ten blue links to answers — and answers need sources. The businesses that win the next few years aren’t the ones that abandoned SEO for AI, but the ones that noticed both run on the same rails: clean, fast, structured, quotable content, plus a brand people mention.
Build for the machines, write for the humans, and let both channels feed you. If you’d rather have this handled end to end, start at our SEO & AI Search service — and if you’re choosing a platform for a rebuild, see our notes on headless CMS first.
Yuri Osipov
Founder, SEO & analytics
Runs Take the Space. Specialises in technical SEO, measurable analytics setups and AI search visibility.