In February 2024, Gartner made the prediction that launched a thousand agency pitch decks: traditional search engine volume would drop 25% by 2026 as users defect to AI chatbots and virtual agents. The deadline is now, and the scorecard is more interesting than the prediction. Search volume did not fall off a cliff. What is falling — measurably, quarter after quarter — is the thing search used to deliver to businesses: the click.
That distinction matters, because the two failure modes call for different responses. If search were dying, you would diversify away from it. Instead, search is mutating into an answer layer that sits between your content and your buyer. When a managing director types "which quoting tools integrate with our ERP, and who implements them?" into ChatGPT or Google's AI Mode, the assistant composes a shortlist from what it has read and retrieved. Either you are in that answer or you are not. Nobody scrolls to page two of a paragraph.
This post covers what the 2025–26 data actually shows, what generative engine optimization (GEO) is once you strip out the snake oil, and the playbook we run for clients — including the parts we would not spend money on.
Search didn't collapse — the click did
Start with the most careful dataset available. In 2025, Pew Research Center analyzed the real browsing behavior of 900 US adults and found that when a Google search produced an AI summary, users clicked a traditional result on 8% of visits — versus 15% when no summary appeared. Users clicked a source link inside the AI summary on just 1% of visits. And they were markedly more likely to end their browsing session entirely after seeing a summary (26% vs. 16%). In Pew's sample, 18% of all Google searches already triggered one — in March 2025, before AI Mode went mainstream.
Ahrefs got the same signal from the other direction. Comparing Search Console data across 300,000 keywords before and after the AI Overviews rollout, they found the presence of an AI Overview cut click-through on the #1 organic result by roughly 34.5%. Being the best-ranked answer to a question and losing a third of its clicks anyway is the new normal for informational queries.
Zoom out to the channel level and the picture gets almost paradoxical. Semrush analyzed billions of web visits across 17 industries through 2025 and found that referral traffic from AI tools grew 66% — while still amounting to less than 0.15% of total visits. In the same study, organic search declined in 13 of the 17 industries, with drops around 25–30% in healthcare, education, and banking. Total web traffic was essentially flat. The demand for answers hasn't shrunk at all; the visits that answers used to generate are being absorbed upstream.
So Gartner's number was arguably wrong and its direction right — which, as Search Engine Land noted at the time, is about what one should expect from analyst predictions. Treat the 25% as a rounding error on the real lesson: the interface changed, and interfaces decide who gets seen.
The B2B twist: fewer visitors, warmer visitors
If AI referrals are a rounding error in aggregate, why should a B2B company care? Because averages hide where the shift is concentrated — and B2B buying research is precisely the query type assistants are good at: comparative, multi-constraint, question-shaped. "CRM for a 40-person industrial services firm, self-hosted, German data residency" was always an awkward Google query. It is a perfectly ordinary ChatGPT prompt.
The referral data reflects that. Semrush's clickstream analysis of ChatGPT — over a billion rows of US browsing data — found outbound referral traffic from ChatGPT to the web grew 206% year over year between January 2025 and January 2026, and the set of domains receiving that traffic expanded from roughly 71,000 a month in late 2024 to a peak around 260,000. ChatGPT is becoming a doorway, not a destination — a small one, growing fast, and unevenly.
What we see in client work matches the "warmer" thesis: leads that mention an assistant in the first call arrive with the comparison already done. They have read a synthesized version of your positioning, your competitors' positioning, and someone's opinion of both. The top of your funnel didn't shrink so much as move somewhere you can't see it. That's the practical meaning of "AI search visibility": the funnel stage you used to observe in analytics now happens inside a model's context window.
One opinionated observation from that work: the companies that panic about lost traffic are usually measuring the wrong asset. A services firm doesn't need 10,000 informational visits; it needs to be one of the three names the assistant offers when the question has budget behind it. Those are different optimization targets, and most SEO dashboards only show the first.
GEO is real — but it's mostly rigor, not tricks
The term comes from an actual research paper. Aggarwal et al. introduced "GEO: Generative Engine Optimization" (KDD 2024), tested nine content modifications against generative engines, and found the winners could boost a source's visibility in generated answers by up to 40%. The details are the useful part: the top-performing tactics were adding quotations, adding statistics, and citing sources — while keyword stuffing, the classic dark art, performed worse than doing nothing. Engines that write answers reward content that reads like evidence.
Hold that against Google's official position, which sounds like a contradiction and isn't. Google Search Central states plainly that there are no additional requirements to appear in AI Overviews or AI Mode — no special files, no AI-specific markup, no new schema; AI features run on the same core Search systems, and their traffic shows up in Search Console under the ordinary "Web" type. Both things are true at once: the retrieval layer is still search infrastructure (crawlability, indexing, relevance), while the synthesis layer selects passages the way the GEO paper describes — quotable, specific, attributed. GEO, honestly defined, is classic findability plus content that survives being paraphrased by a machine.
The playbook we actually run
The order matters; each step compounds the next.
1. Baseline yourself in the engines
Before touching anything, ask ChatGPT, Perplexity, Gemini, and AI Mode the questions your buyers ask — including the unbranded ones ("who builds X for companies like Y in Germany?"). Log the answers, the citations, and who gets recommended instead of you. Repeat on a fixed cadence. This is unglamorous and it is the single highest-information exercise available, because there is no Search Console for ChatGPT.
2. Open the gates — deliberately
Answer engines can only cite what they can read. OpenAI documents this explicitly: sites that block OAI-SearchBot will not be shown in ChatGPT search answers, and it recommends allowing it in robots.txt. Note the separation of concerns in the same document: GPTBot governs model training, OAI-SearchBot governs search visibility — you can refuse one and welcome the other. While you're in there, make sure your money pages render as server-side HTML. A pricing table that only exists after client-side JavaScript executes is invisible to most retrieval pipelines.
3. Write pages that survive synthesis
Apply the paper's findings literally. Every important page should carry concrete numbers, named sources, dated claims, and at least one sentence a model could lift verbatim as a complete answer. "Leading provider of innovative solutions" compresses to nothing in a synthesized answer; "we migrated a 200-user ERP in 11 weeks, here's the architecture" is quotable. The first screen of a service page should answer the buyer's actual question, not clear its throat.
4. Make your entity unambiguous
Models resolve entities, not keywords. One canonical company name, description, address, and founding story — consistent across your site, imprint, LinkedIn, and registries — plus Organization and Person structured data that matches the visible text (Google's guidance makes that match explicit). Real authors with verifiable roles beat "admin" bylines here for the same reason they do in Google's quality frameworks: attribution is evidence.
5. Measure like it's 2026
Track three things monthly: assistant referral sessions (they're small; their conversion behavior is what matters), your citation share in the baseline queries from step 1, and branded search volume as a proxy for answers you never see. Accept fuzziness. Anyone selling you a precise "AI rank" for a probabilistic system that changes with every model release is selling confidence, not measurement.
What we wouldn't spend money on
Three honest anti-recommendations. First, the llms.txt cargo cult: Google states outright that you don't need to create AI text files or new machine-readable files to appear in its AI features, and no major engine has committed to consuming the format — it costs little, but treat it as a lottery ticket, not a strategy. Second, guaranteed-placement offers: retrieval favors content that is consistently cited across the open web, which takes months and cannot be bought on a deadline. Third, chasing raw traffic recovery: the Pew and Ahrefs numbers above describe a structural repricing of the click, not a dip you can optimize your way back from.
The encouraging part is how boring the winning move is. Almost everything the evidence rewards — specificity, citations, verifiable authorship, machine-readable structure — is what credible content looked like before anyone abbreviated anything to GEO. The engines changed. The bar, mostly, didn't. It just started being enforced.