What Actually Gets a Page Retrieved, Ranked and Cited

Most of what circulates about "AI SEO" does not survive contact with the evidence. A review of roughly 350 primary sources, including Google and Bing's own documentation, the 2024 Google antitrust trial record, and every major peer-reviewed study on generative-engine citation, finds that retrieval is the real bottleneck, not clever phrasing, and that a handful of well-evidenced practices matter far more than the popular ones. Every claim below is graded A to U for evidence quality: A is a primary document or peer-reviewed source, C is a large vendor-published correlational study, U is a widely repeated claim that does not check out.

The five-minute version

• There is no separate discipline for "AI search". Google's May 2026 guidance says its AI features run on the same core ranking systems, need no special files or markup, and that Google Search ignores llms.txt entirely.
• Being retrieved is the bottleneck. Peer-reviewed research finds that rewriting a page to sound more "citable" is mostly ineffective, and can even reduce a page's odds of being retrieved at all.
• Once a page is retrieved, close relevance to the literal question and extractable evidence (a number with a source, a named quote, a plain definition) are what get it used.
• User satisfaction, not links, is the best-evidenced organic ranking input. Sworn testimony in the Google antitrust case confirms a system called Navboost tracks click and return-to-results behaviour over 13 months and was not replaced by AI signals.
• A single technical defect, an incomplete or truncated page title or description, is one of the cheapest, most testable things to check on your own site before spending anything on content or authority work.

What actually moves organic rankings

Two documents changed what can be said with confidence here: the 2024 leak of Google's Content Warehouse API reference, and the findings of fact in the US v. Google antitrust trial. Neither reveals exact ranking weights. Both reveal what Google stores, and what a federal court found it actually relies on.

User interaction is the best-evidenced input. The trial record states plainly that a system called Navboost "pairs queries and documents through memorizing user click data" over a 13-month window, and that "the more recent LLM signals did not replace Navboost." The practical target is what search engineers call the last long click: the visit after which a searcher does not go back to the results page. [A]

Links matter, but less than commonly assumed. Google removed the word "important" from its own description of links back in March 2024, and in several UK search results checked for this research, pages with zero referring domains ranked in the top ten for competitive, buyer-intent search terms. Meanwhile brand mentions, even unlinked ones, correlate with AI visibility roughly three times as strongly as backlinks do across a 75,000-brand dataset. [A] [C]

Genuine folklore to retire: content length as a ranking factor (Google has stated there is no preferred word count), publishing frequency, "link velocity", and third-party domain authority scores as if they were one of Google's own signals. Google has explicitly said third-party authority metrics "don't correspond to any of Google's own signals." [A]

What actually moves AI citation

Citation in a tool like ChatGPT, Perplexity or Google's AI Overviews is the last stage of a five-stage pipeline: search, crawl, retrieve, survive re-ranking into context, and only then get cited. Almost every published claim about "generative engine optimisation" measures only that final stage, with the first four held fixed, which is why so many of the popular tactics fail to replicate.

What holds up: close relevance between a page's headings and the literal question asked (the single strongest page-level signal one large study found), extractable evidence such as a sourced statistic or a named quote, and appearing in the first third of a page, where roughly 44 percent of verified citations originate. [A] [C]

What is contradicted: the widely quoted "30 to 40 percent" citation uplift from rewriting content for AI only held inside a fixed, artificial lab context. A 2025 academic benchmark found the technique worked in just 3 of 54 real-world test combinations. Adding schema markup specifically for AI citation measured a small negative effect in a matched study of 1,885 pages. And two independent server-log studies found that llms.txt files receive essentially zero requests from any frontier AI crawler. [A] [B]

Perhaps the most important finding for anyone running one KPI dashboard: any two AI engines share only about 17 percent of the domains they cite, and roughly two-thirds of an engine's cited sources change from one day to the next. A single-day "are we cited" check is noise. Measuring this properly requires a fixed set of prompts run repeatedly, over weeks, per engine.

What buyers and journalists actually respond to

Three-quarters of B2B decision-makers say they have researched a company they were not otherwise considering, purely because of a piece of thought leadership they read (Edelman and LinkedIn, roughly 3,500 respondents across seven countries). But fewer than half rated what they had read as good, and only 15 percent rated it very good, which is exactly the gap an organisation willing to publish real research and named expertise can close.

Journalists behave the same way. A 2026 study of 1,899 journalists found 66 percent rely on material supplied by communications teams for story ideas, but 86 percent say they immediately reject anything off-beat for their coverage area. Original research and data consistently earn more editorial links than commentary: one analysis found annual survey-style reports earn roughly ten times the linking domains of comparable non-data content.

How to write a page that gets retrieved, ranked and cited

Put the direct answer in the first 100 to 150 words, in plain, definitional language, with one concrete number and its source named inline. Use a heading structure where each subheading is phrased as the specific sub-question a reader (or an AI system's own query expansion) would actually ask, and answer it in the first sentence beneath. Cover the sub-questions that genuinely belong to this page's topic rather than every adjacent one: pages covering roughly a quarter to a half of the relevant sub-questions are cited more than pages that try to cover everything.

Include a byline with real credentials, cite every statistic with its source and date, and add a short line whenever you meaningfully update a page (Google explicitly lists bumping a visible date without real change as a bad practice it can detect). Above all, include something genuinely original: your own data with a stated method, a named expert's considered position, or a specific case none of your competitors can also publish. That is precisely the category Google's own guidance singles out as what it wants to see, and precisely what the buyer research above says decision-makers are starved of.

Practices worth actively dropping: writing to a fixed word count, rewriting headings purely into keyword form (the one technique that measured below baseline in peer-reviewed testing), bolting a generic FAQ block onto a page that already answers its own question, and treating any single engine's citation pattern as the template for every other engine.

A five-step diagnostic you can run on your own site today

1. Pull your analytics and Search Console data for your commercial pages: clicks, impressions, click-through rate and average position over the last 90 days.
2. Flag any page where click-through rate sits well below the published curve for its position (roughly 30 to 35 percent at position one, low single digits by position ten, and often halved again wherever an AI summary appears above it). A well-ranked page with a starved click-through rate has a snippet or relevance problem, not a ranking problem.
3. Fetch your own key pages using a Googlebot or GPTBot user agent, and read the title and description exactly as served, not as they appear in your content editor.
4. Check specifically for truncation: does any title or description end mid-sentence, or with a literal ellipsis, in the raw page source? Google's documentation names incomplete or boilerplate descriptions as a specific, stated reason it discards your snippet for one it generates itself. This is one of the cheapest, most testable fixes available to any organisation.
5. Compare your domain's authority metrics against two or three real competitors, alongside whatever AI-citation count your tools provide. Authority without genuinely useful, indexable content produces neither strong rankings nor AI citations, and the reverse holds too.

The bottom line

Most organisations chasing "AI visibility" are optimising the wrong layer. The evidence says technical eligibility, being indexed, server-rendered and free of snippet defects, is a gate you either pass or fail, not a lever you can pull harder on. Above that gate, the two things that reliably move both rankings and citation are the same two things that always mattered: content genuinely worth citing, and a reputation, earned through real coverage and real expertise, that a search engine or a language model can find corroborated somewhere other than your own website.