AI search for B2B websites: what the evidence actually shows

Across 2025, a B2B page cited in an AI Overview got 2.07% organic click-through. Uncited, 0.94%. On the same informational queries with no AI Overview, 3.35%. Seer Interactive measured that across 5,471,127 queries. Being cited was worth roughly double the clicks of being left out. AEO is damage control before it is growth.

That 0.94% is a full-year average across 53 brands. The same Seer page tracks the recovery: click-through bottomed out at 1.3% in December 2025 and reached 2.4% by February 2026. Winning the citation is a recovery play, climbing back toward a number you had for free.

Desses builds B2B websites, mostly on Framer, and we sell an AEO audit that can turn into a retainer. I have a commercial reason to tell you this channel is urgent. The evidence below doesn't support urgency. We're a Framer Partner, so Framer shows up here more than any other builder. Named, so you can discount it.

What AEO and GEO mean

Answer engine optimization (AEO) is the practice of making a website's content retrievable and quotable by systems that answer in prose where a search engine used to return ten links: Google's AI Overviews and AI Mode, ChatGPT, Perplexity, Claude and Copilot.

Generative engine optimization (GEO) is the same goal named from the model's side: shaping a page so that a large language model composing an answer selects it, cites it and reproduces its claims accurately.

The two get used interchangeably. Bing became the first search engine to formally recognise GEO, in its February 2026 Webmaster Guidelines, with no markup requirement anywhere in them. That page sits behind a sign-in, so it reaches you through Search Engine Journal.

What changed on the results page

Pew Research watched 900 US adults across 68,879 searches in July 2025. With an AI summary present, 8% clicked a traditional result against 15% without, 26% ended the session there against 16%, and 1% clicked a link inside the summary. Google went at the method, calling it "a flawed methodology and skewed queryset that is not representative of Search traffic" and saying nothing about the finding.

Ahrefs, 300,000 keywords, February 2026: position-one content loses 58% of its average click-through when an AI Overview is present. That's a revision of the same study's 34.5% in April 2025. Anyone citing both as two findings doubles their own evidence.

How much traffic AI search actually sends

Semrush, 50,000-plus websites, 27 April 2026: AI referral traffic grew 66% during 2025, from 462 million to 767 million monthly visits, and sits at 0.14% of all web visits. Ahrefs, 81,947 sites, June 2025: roughly 10x growth in twelve months, averaging 0.25% of a site's traffic, with Google still sending about 210 times more than ChatGPT, Perplexity and Copilot combined.

The counterweight is Semrush's July 2025 finding that the AI search visitor is 4.4 times as valuable as the organic one, based on conversion rate. Scope was 500-plus marketing and SEO topics, so it's a number about SEO blogs until someone runs it in your category.

In Semrush's 2025 industry breakdown, computer software and development was the only declining sector, down 26%. Every other sector grew. If you sell software or development to other businesses, the channel everyone is selling you shrank in your category last year.

Four surfaces, four different problems

Ranking gets you into the pool. The citation is a separate fight. Semrush found only 6 to 27% of an industry's most-mentioned brands also ranking as top cited sources. A model can talk about you all day without linking to you.

Query fan-out, and what it does to a page

Query fan-out is Google's named mechanism for AI Overviews and AI Mode: issuing multiple related searches across subtopics and data sources, then assembling one answer. Robby Stein, VP of Product for Search, put it plainly in July 2025. "It may think of a bunch of questions... and it'll start Googling basically."

Your page gets evaluated against questions your visitor never typed. A page built for one query answers one branch of the fan and drops out of the rest, which is where one topic per URL earns its keep.

What holds up, and what's being sold to you

The GEO paper (Aggarwal et al., KDD 2024) tested nine content edits against GEO-bench, 10,000 queries. Quotations lifted visibility 40.9%, statistics 30.6%, keyword stuffing cost 8.3%. Those one-decimal figures appear nowhere in the paper. They're derived from its Table 1 against a 19.3 baseline, and the paper's own headline is "up to 40%".

The caveats almost never travel with them. All of it is 2023 to 2024 vintage, and a critical survey published July 2026 states that the 40% figure "describes a relative visibility gain in a simulator in which five documents have already been placed in context." It was measured after the retrieval fight was won.

Benchmarks since have gone the other way. C-SEO Bench (Puerto et al., 1,921 queries) found "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking," with 3 of 54 method-domain combinations significantly positive and the effect zero-sum.

Two things get sold hard with the evidence against them. Ahrefs checked 137,000 domains in June 2026: 28% publish an llms.txt, and 97% of the roughly 38,000 carrying a valid file got zero requests for it in May 2026. Ahrefs' May 2026 study of 1,885 pages that added JSON-LD, against roughly 4,000 controls, found AI Overviews citations down 4.6% and statistically significant. Does llms.txt do anything and does schema markup get you cited by AI take those apart.

Who shouldn't spend money on this yet

If AI referrals are under a quarter of a percent of your traffic and your pipeline is healthy, run the free checks and stop. Server logs, a curl of your top three pages as OAI-SearchBot, a robots.txt read. That's an afternoon. A retainer on top buys a report.

If your site is client-rendered React with no server rendering, you have a rendering problem sitting in front of an AEO problem. AI crawlers still do not run JavaScript covers that one.

If you have no measurement baseline, the work is an unfalsifiable invoice. That's what how to actually measure AI search visibility exists for.

Anyone selling AEO as a traffic channel today is selling a 2028 argument at 2026 prices. Anyone ignoring it is planning to arrive late to a channel growing 66% a year off a small base. Both are true at once.

What we'd do, in order

Read the server logs first. It's the only method that shows whether OAI-SearchBot, Claude-SearchBot and PerplexityBot reach your content. Verify by published IP range, because spoofing is common. Anthropic lists its ranges at claude.com/crawling/bots.json.

Then curl your three most important pages as a retrieval bot. If the H1, the body text and the canonical link aren't in the initial HTML, fix that first. Vercel established in December 2024 that no major AI crawler executes JavaScript, and searchVIU tested five systems in December 2025 with Gemini the only one rendering JS on a direct live fetch. Framer clears this by default, pre-rendering every page server-side.

Get the robots.txt distinctions right. GPTBot and ClaudeBot build training corpora, and blocking them costs nothing in AI search. OAI-SearchBot and Claude-SearchBot feed the indexes those products cite from, so blocking either takes you out of that product. Then put the answer in visible body text near the top, one question per URL, in the words a buyer would use.

Turn on the free measurement before you change anything. Search Console, and Bing Webmaster Tools' AI Performance report, the only place a provider shows you the queries that retrieved your page.

These calls get made during design, before a line of copy exists. That's where they sit for us, inside how we approach B2B website design.

Where to read next

The two files everyone asks about: llms.txt and schema markup. The check that decides the rest: AI crawlers and JavaScript, then design decisions and page structure. Platforms: Framer's Markdown endpoint, the whole Framer build, Next.js, the three compared. Then measurement, redesigning, cost and whether to hire anyone.

Send me your URL and I'll tell you what OAI-SearchBot gets back from your homepage. Takes about twenty minutes and there's nothing to buy at the end. If it's clean, I'll say so.