Why GEO is decisive now: search stopped sending the click

What actually changed on the results page

For twenty-five years search ran on one pattern: you typed a query, you got a list of links, you clicked through it. The whole discipline of SEO, search engine optimization, existed to move you higher in that list. That pattern is now the exception in a growing share of queries. Google's AI Overviews reach more than two billion people a month across 200-plus countries and 40 languages, and Pew Research, watching the real browsing of 900 US adults across roughly 69,000 Google searches, found that when an AI summary appeared people clicked a traditional result in 8% of visits instead of 15%, and clicked a link inside the summary itself in 1%. Google disputes that methodology. Ahrefs, working from Search Console data on 300,000 keywords, landed in the same direction: a 58% lower click-through rate for the number one organic result when an AI Overview is present.

This is a change in who decides, not a change in layout. The user used to pick which page got the visit. Now a language model decides which retrieved pages make it into the answer and which names get said out loud. That selection is not the old one wearing a new coat. Ahrefs found only about 6 to 8% of the URLs ChatGPT cites also sit in Google's top ten for the same query, and roughly 80% do not rank in Google's top 100 at all. Position one is no longer the stage that matters most. That is the gap GEO is meant to close: earning the citation inside a generated answer, not the rank inside a list.

Take a concrete case. Someone looking for a tax advisor used to type in their city and open six tabs. Now the same person asks an assistant what to check before signing an engagement letter and which firms nearby have a solid reputation, and gets one condensed answer with two or three names in it. Whether one of those names is yours is settled before anyone loads your homepage.

Why now and not in two years

Fair objection: tools come and go, so why move now? Because the adoption already happened. Google places AI Overviews above the organic results, the Gemini app passed a billion monthly active users in August 2026, Microsoft 365 Copilot passed 30 million paid seats, and tools like ChatGPT and Perplexity are default research surfaces now rather than curiosities, with ChatGPT alone reporting 900 million weekly active users in February 2026. What was a side channel two years ago is the first point of contact, and often the only one, for a large slice of your market.

The second reason is compounding. What an engine says about you is assembled from what the web already says about you, and that inventory is slow to turn over. Ahrefs, looking at 1.4 million real ChatGPT prompts and 23.4 million retrieved URLs, found the pages that actually get cited have a median age of around 500 days. What you publish this quarter is stocking the citation pool for next year, not next week. Starting once competitors are already inside the answers means serving out that lag from behind.

The honest framing: GEO does not replace SEO, it sits on top of it. Google says its generative features are rooted in the same core Search ranking and quality systems, so a page still has to be indexed and eligible to show a snippet before it can be quoted at all. But eligibility is where the overlap ends. Only 37.9% of URLs cited in AI Overviews still rank in the traditional top ten, down from roughly 76% in mid-2025. Both channels need serving, and they are drifting apart.

The quiet loss: cited but never visited

The awkward part of this shift is how late it shows up in your dashboard. Rankings can hold steady while sessions drift down, because the answer arrived before the click was needed. That is zero-click search: the question gets resolved and nobody visits a page. Pew found sessions ended immediately after the search 26% of the time when an AI summary was present, against 16% when it was not. Your knowledge gets used. Your name may or may not travel with it.

What that costs depends on what you sell. A shop loses comparison traffic when the model weighs the specs itself. A practice or a trades business loses inquiries when the local recommendation lists three other names. A software vendor loses trials by being absent from the best-tools answer in its category. Across US Google searches, SparkToro and Similarweb counted roughly 276 clicks reaching the open web per 1,000 searches in early 2026, down from 374 in 2024. The loss is real, it just arrives as a slope instead of a cliff.

That is exactly what makes it dangerous. A crash forces a response; a slow bleed gets explained away for quarters. So change the question you ask of your reporting: not only how many people reached the site, but whether you are named in the answers being given about you and your category. Being named is worth something measurable. Ahrefs found pages cited inside an AI Overview earn roughly 35% more organic clicks than pages that rank without being cited.

How answer engines choose the sources they name

None of the major engines answer business and vendor questions from pretrained memory alone; they retrieve first, then write. So the practical question is what survives retrieval and makes it into the sentence. Content that answers one question directly, in plain declarative sentences, is easy to lift. A brand paragraph that circles the point for eighty words is not. None of OpenAI, Perplexity or Anthropic publishes a scoring formula for which retrieved pages get cited, so treat any confident list of weighted signals as third-party inference. Extractability, though, is not in dispute.

Structure a machine can parse matters for the same reason. Unambiguous headings, real question-and-answer blocks, facts stated once and stated exactly. Schema.org markup helps here: it makes prices, hours and locations reliably extractable and cuts down on invented details. It is not a ranking factor, and Google has said so since 2018, so treat it as legibility work rather than leverage. Note what Google explicitly warns against, too: producing a separate body of content aimed only at AI, which its scaled-content-abuse policies target directly. No special markup, schema or llms.txt file is required to appear in AI Overviews.

The third factor is corroboration. A model builds its picture of you out of everything it retrieves, so agreement across sources does real work. If your positioning, your numbers and your basic facts match wherever they appear, an engine has an easy time treating you as settled. If your own pages contradict a directory listing, or almost nobody outside your domain describes you at all, there is nothing to corroborate, and the model reaches for a source that states it more plainly.

Mentions correlate better than backlinks

Classic SEO ran on links pointing at you. AI citation appears to run on being talked about, link or no link. Across roughly 75,000 brands, Ahrefs found the frequency of brand web mentions correlated with AI citation rate at about 0.66, roughly three times the correlation for backlinks at 0.22. That is the strongest evidence-backed lever anyone has published, and most of it lives off your own domain: expert write-ups, industry directories, comparison lists, forum threads, reviews.

Which changes where the effort goes. Polishing your own site is necessary and no longer sufficient. You want to be present, and described correctly, wherever your category actually gets discussed. A regional energy consultant gains from local portals and trade publications. A B2B vendor gains from case studies specific enough to quote and from appearing in the comparison round-ups buyers ask about. The model stitches those fragments into one impression of you, whether or not the fragments agree.

Four questions worth answering honestly:

• Am I described anywhere outside my own domain, and is the description accurate?
• Do those mentions say plainly what I do and who I am the right choice for?
• Is there an independent source a model could use to corroborate my claims?
• Are my core facts, pricing, coverage, hours, credentials, identical everywhere, or are old versions still live?

This is not only a big-brand problem

The reflex that GEO is for enterprises with a content team gets it backwards. The GEO paper from Princeton and IIT Delhi, presented at ACM SIGKDD 2024, found its optimization tactics helped lower-ranked sources most and actually hurt the source already sitting at rank one. Its main experiment ran on a simulated engine, GPT-3.5 writing an answer from the top five Google results, so read the numbers as directional rather than proven in live search. But the asymmetry is intuitive: the incumbent has less to gain from being easier to quote than the challenger does.

The split between winners and losers rarely follows industry lines. It follows posture. Pages written as self-presentation give a model very little it can lift. Pages that take a real customer question and answer it, including the inconvenient parts, hand the model exactly what it needs to build a paragraph. That holds for a shop, a law firm, a clinic and a fabrication business alike.

Picture two vendors with near-identical offerings. One has a handsome site full of superlatives and no specifics. The other answers its ten most common customer questions on the page, with numbers, named constraints and clear statements. To a person browsing, the two look comparable. To a system that has to assemble a defensible answer and attribute it, only the second is usable. Worth noting from the same research: keyword stuffing produced no meaningful gain. The old lever is simply dead here.

Where to start, in order

Start by measuring, not building. Ask the engines the questions your buyers ask, in the words they would use, and read what comes back. Are you named? Is what is said about you true? Is a competitor standing in your place? Do this before you commission anything, because the failure mode is not only absence. Columbia Journalism Review's Tow Center ran 1,600 source-identification queries across eight AI search tools and found more than 60% of answers incorrect; ChatGPT misidentified 134 of 200 articles while signalling uncertainty just 15 times. Confident and wrong about your business is a real outcome.

Then do the substance work. Take the questions you actually get asked and answer them on your own pages, concretely, without the marketing register. Make your core facts identical everywhere and machine-readable. Work on being described coherently off your own domain, because that is where the measured correlation lives. Skip the shortcuts: John Mueller confirmed in 2025 that no Google Search system reads llms.txt, and an Ahrefs review of about 137,000 sites publishing one found roughly 97% got no measurable referral traffic from it. Cheap hygiene at best.

Stay honest about the ceiling. GEO is not a switch, nobody can guarantee a citation, and no platform publishes its rules. Perplexity at least runs a named crawler that respects robots.txt, so check you have not blocked yourself out of its index. What you control is whether your traces on the web are clear, consistent and quotable. That compounds slowly and is genuinely hard to catch up on. The shift already happened. The only open question is whether you are in the answer or watching it.