68% of searches now end without a click. Here's what that actually changes about building a website.

Sometime in the last year, zero-click search stopped being a niche SEO talking point and became the default experience of using a search engine. Across Google generally, roughly two-thirds of searches now end without a click to any website, ours or anyone else's. On a search where Google's AI Overview shows up, that number gets worse — over 80% end on the results page itself, because the answer already got read to the person asking. Google's own newer AI Mode is more aggressive still: north of 90% of those sessions never reach an external site at all.

The instinct this creates in a lot of teams is panic, or a kind of "why bother" fatalism about content and SEO work generally. Neither reaction is right. What's actually happened is that the job of a page changed. It used to be enough to rank; ranking got you the click, and everything that mattered happened on your site after that. Now, for a large and growing share of searches, an AI system reads your page, decides whether it trusts what's on it enough to cite or summarize, and the person asking may never visit at all — but whether you get cited, recommended, or mentioned by name in that AI answer still depends entirely on what's actually on the page.

That means the technical foundation that always mattered for classic SEO — fast load times, clean semantic markup, real structured data — hasn't stopped mattering. If anything it matters more, because an AI system deciding whether to trust and cite your page is running the same kind of technical checks a search crawler always did, just as a gate before it'll consider citing you at all. What's different is what has to sit on top of that foundation: a direct, extractable answer near the top of the page, in plain language, rather than three paragraphs of scene-setting before you get to the point. AI systems summarizing a page tend to grab the first clear, well-formed answer they find and mostly ignore the throat-clearing around it.

In practice, this changes how we structure a page from the first draft, not just the meta description. The direct answer to the page's core question goes in the first two or three sentences after the heading, in plain declarative language an AI system can lift as a quotable snippet without needing to interpret tone or infer meaning from context. The elaboration, caveats, and supporting detail come after, for the reader who does click through and wants the full picture. Headings are written as the actual questions a reader — or a model summarizing the page — would ask, not as clever or branded phrasing that reads well but answers nothing on its own. And a page answering more than one distinct question gets that structure repeated per question, in an FAQ block with real FAQPage structured data behind it, rather than one long undifferentiated block of prose covering several topics at once.

The harder practical problem is measurement, because a citation in an AI answer doesn't show up in Google Search Console the way a ranked, clicked result does. What we track instead: branded search volume for the company name (a proxy for how often people are hearing about you from somewhere, AI answers included, even when that mention doesn't generate a click), direct traffic that arrives already knowing what it's looking for rather than searching generically, and — where it's available — manually checking how a page's content shows up when the same question is asked directly of ChatGPT, Perplexity, and Google's AI Overview. None of these are as clean as a keyword ranking used to be. That's the actual state of measurement in this transition, not a solved problem we're choosing to describe simply.

The question we get most often is whether it's still worth investing in classic SEO at all, given how much traffic now ends without a click regardless. The answer is yes, for a reason that isn't obvious until you look at where AI citations actually come from: brands that perform well in AI-generated answers are overwhelmingly the same brands that already rank well in classic search, because the two systems are evaluating overlapping signals — authority, structure, freshness, technical soundness — even though they present the result differently. Treating GEO as a replacement for SEO misreads the relationship. It's closer to an additional layer built on the same foundation, which means neglecting the foundation to chase the new layer tends to hurt both at once.

The other real shift is consistency. An AI system trying to decide what your company actually is, and whether to trust a claim on your site, is often cross-referencing multiple pages and multiple mentions of you elsewhere — which means an entity described one way on the homepage and a slightly different way in a blog post is a small trust problem in a way it never used to be when a human reader just read whichever page they landed on. Structured data — the same Organization and Service markup search engines have used for years — is what lets an AI system confirm what you actually are instead of piecing it together from inconsistent prose, and freshness signals, a page that's genuinely been reviewed recently rather than just re-dated, carry real weight in whether that system trusts it enough to cite.

None of this is a reason to stop caring whether people click through. It's a reason to stop assuming they will. We now build every page assuming an AI system, not a person, might be the first reader — which mostly means writing the direct answer first and the elaboration after, rather than the other way around, and treating structured data as load-bearing content rather than an afterthought bolted on before launch.

There's a reasonable worry buried in all of this about giving away expertise for free — if an AI system can just read your page and answer the question without sending anyone to your site, what's the incentive to write the genuinely good, detailed content that makes a page worth citing in the first place? We don't think that worry holds up in practice. The pages getting cited most consistently in the AI answers we track are still the ones doing real, specific, checkable work — a documented process, a real number, a genuine answer to a question a generic response can't match — rather than thin content optimized purely for extractability. Being the source an AI system trusts enough to cite by name is still a form of visibility and credibility that a purely defensive, give-nothing-away page doesn't earn either.