What a B2B page has to look like to get quoted

The page-level tactics that survive scrutiny are the ones that were good writing anyway. Cite your sources, use real numbers, state things directly, and keep one question per URL with the answer near the top. Most of the rest of what gets sold as page-level GEO has been tested and found ineffective, or harmful to your ranking.

Desses sells an AEO audit that covers content structure, which gives me an obvious reason to tell you your pages need rewriting. Most of this article argues against the version of that work everyone is selling, including the parts I'd be paid for.

Generative engine optimisation, usually shortened to GEO, is the practice of changing a page's content so AI systems quote it in their answers. Page-level GEO is the subset that touches copy, headings and structure. Nearly all the published advice lives there, and it has the weakest evidence behind it of anything in this field.

Where the evidence converges

One paper carries almost this entire category. Aggarwal et al. published GEO at KDD 2024: GEO-bench, 10,000 queries, measuring what content edits did to visibility. The headline claim is a gain of up to 40%.

The per-tactic percentages everyone quotes are printed nowhere in that paper. They come out of Table 1's Position-Adjusted Word Count column against a 19.3 baseline: quotations 40.9%, statistics 30.6%, fluency 28.0%, citing sources 27.5%, an authoritative tone 10.4%, keyword stuffing −8.3%. Every guide reciting those to one decimal place is reciting arithmetic somebody else did, this one included.

In February 2026 Microsoft wrote GEO guidance into the Bing Webmaster Guidelines, and every item is about the form of the content: state facts directly, use clear and consistent entity names, keep one topic per URL, put the essential information near the top. Bing also writes down what no vendor puts in a deck, that GEO guarantees a citation about as much as SEO guarantees a ranking. That page sits behind a sign-in, so you're reading Search Engine Journal's account of it.

Where it falls apart

Read the paper's methods and the 40.9% changes shape. The main experiments ran on a custom pipeline feeding the top five Google results to GPT-3.5-turbo, never against ChatGPT, Perplexity or AI Overviews. The one test against a live product was a secondary experiment on Perplexity, and its 200 samples were file uploads, not live web fetches. All of it is 2023 and 2024 vintage.

A critical survey published in July 2026 says it flatly. The famous 40% "describes a relative visibility gain in a simulator in which five documents have already been placed in context". The 40.9% is real. It measures what happens after you've already been fetched.

The one attempt to reproduce this at scale landed badly. C-SEO Bench, from Puerto et al., ran 1,921 queries across 16,360 documents and found "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking". Three of 54 method-domain combinations came out significantly positive. Fifty-one didn't.

The same authors measured the effect as zero-sum: "as we increase the number of C-SEO adopters, the overall gains decrease". Whatever lift exists is a lift over the people who haven't done it yet.

The page tactics, scored

The folklore, named so you can refuse it

Two claims circulate as fact in page-structure guides and neither has a primary source at any provider.

That AI systems chop your page into passages of a specific token count. Nobody at OpenAI, Anthropic, Google, Perplexity or Microsoft has published a chunk size. Google has confirmed something adjacent, that it ranks passages inside a document, described by Martin Splitt as one document with multiple annotations scored independently in February 2021. A token count quoted as fact is the quickest way to spot someone who read three blog posts and none of the sources.

That there's an ideal answer paragraph, usually given as 40 to 60 words. No provider states a length. The figure shows up in one guide, gets quoted by the next, and has no origin.

Two studies, one question, and what separates them

You'll see two numbers quoted for how much AI citation overlaps with search ranking, and they don't agree.

Ahrefs looked at 1.9 million citations from 1 million AI Overviews and found 76.1% of cited pages rank in the top 10. seoClarity looked at 5.1 million citations across 362,000 AI Overviews on 12 October 2025 and found 56% came from that query's own top 20, which works out at roughly 41% from the top 10.

76.1% against 41%. The gap is a sampling decision sitting in the methods of both papers. Ahrefs took only the top three most visible citations in each Overview. seoClarity took all 5.1 million, about fourteen per Overview. Pull the three most prominent out of fourteen and you've selected for the pages most likely to rank. Ahrefs describes the front of the citation list.

The first draft of this page had that wrong. I quoted a 32% from seoClarity, a per-keyword overlap rate sitting under a different heading in the same study, and built an explanation on it that didn't hold. Fact-checking caught it. The note stays because an article about what gets claimed versus what gets measured doesn't get to quietly swap its own numbers out.

Ranking well is the strongest available proxy for getting retrieved, and it isn't a requirement.

What's left after all that

Five things survive contact with the evidence. Cite your sources, with the publisher and the date inside the sentence. Use real numbers where you'd otherwise reach for an adjective. State facts directly, so a sentence lifted out of the page still means what you meant it to mean. One question per URL. The answer near the top.

That last one has a thin measurement behind it. CXL read 100 pages cited in AI Overviews and reported that "55% of citations came from the first 30% of content, while only 21% came from the bottom 40%", published 6 March 2026. A hundred pages, undisclosed methodology, unreplicated. Kevin Indig's count of 18,012 verified ChatGPT citations came out at 44.2% from the first 30%, the only corroboration on record.

That's the list, and it's a description of good writing. It was good writing in 2015. The measured lift attached to it is real, small, and shrinking as adoption rises.

I sell this work. I'd rather you knew what you were buying.

When page work is the wrong thing to spend money on

Two conditions disqualify every tactic above, and both sit underneath the page.

The first is crawlability and rendering. If your text arrives through JavaScript, most AI crawlers never see any of it, and if you're blocking OAI-SearchBot you've been removed from ChatGPT search entirely, which no amount of quotation density repairs. The crawler article covers both.

The second is retrieval. Bing Webmaster Tools added an AI Performance report in beta on 10 February 2026 surfacing grounding queries, the phrases that pulled your content in. It's the only place any provider shows what was asked before you were cited.

Then the flat one. If your buyers don't use AI search, none of this is your problem, and the channel is smaller than the pitch. Ask five customers how they found you first.

What we'd do to a page

Fetch it first, without JavaScript, and read what comes back. On Framer you can append ?md to any URL or send an Accept: text/markdown header, and the page comes back as markdown. If your answer isn't in that output, stop reading this section and go fix the rendering.

Then one pass, top to bottom. Move the answer into the first paragraph and write it so it survives being lifted out. Give every number its publisher and date in the same sentence. Cut the page to one question and give the others their own URLs. Then leave it alone.

That's an afternoon per page. The work that costs real money sits underneath: rendering, internal links, headings that don't skip levels, a CMS that doesn't quietly break slugs. That part is our website design work, from $8,000 over three to four weeks.

Send me your URL and I'll send back three things I'd change on the page. If the answer is nothing, I'll tell you that instead.