How AI Platforms Decide What to Cite

The earlier version of this lesson listed seven "ranking factors" for AI citations: content structure, E-E-A-T, freshness, schema markup, quotability, Domain Authority and brand entity strength. Most GEO courses still teach a list like it. When researchers and the engines themselves put those factors to the test, most of them failed or turned out never to have been tested at all.

The engines say very little. OpenAI's entire public account of how ChatGPT ranks search results is two sentences: it uses "multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed." Everything else on a factor list comes from studies, and the studies are where the trouble starts.

This lesson replaces the list with an evidence ledger. Each factor gets a stage, a confidence grade and a plain answer to one question: can you pull it? Some factors are strong but give you nothing to pull. Some popular ones turned out to hurt. A few survive. They are less exciting than the old list, and they are the ones that hold up.

Why most GEO effect sizes are Stage 3 results sold as Stage 1 outcomes

Recall the four stages from the previous lesson: parametric memory, access, retrieval and generation. A publisher wants to win the early stages. They want to be found. Almost every GEO experiment measures the last one.

The typical design: take a handful of documents, place them directly in a model's context, rewrite one of them (add statistics, add quotations, change the tone), and measure whether the model cites it more. No search engine is called. No index is touched. The rewritten page is already in the room. The original GEO paper (KDD 2024), C-SEO Bench (NeurIPS 2025) and a 252,000-trial factorial study (arXiv:2605.25517) all work this way.

Those studies can tell you what a model does with a page it already has. They cannot tell you whether a rewrite gets your page retrieved in the first place. Sold as "this tactic gets you cited by ChatGPT," a Stage 3 result becomes a Stage 1 promise it never tested. A survey of 45 GEO studies (arXiv:2607.14035, July 2026) reached the same verdict: retrieved content can be nudged, but no technique has shown a stable effect on being discovered.

One more problem runs through the correlational studies: brand size. Cited pages belong disproportionately to large, well-known brands, and no study controls for that properly, because the usual "controls" such as Domain Authority are themselves built from backlinks and traffic. Any finding of the form "cited pages tend to have X" may be measuring "big brands tend to have X."

The evidence ledger

Here is every factor that commonly appears in GEO advice, graded against the verified evidence this course is built on.

Factor	Stage	Evidence	Can you pull it?
Access gates: robots.txt per bot, firewall and CDN, server rendering, index and snippet eligibility	1	Strong, documented by the engines	Yes. Subtractive: you can only fail them
Topical relevance to the retrieval query	2 and 3	Moderate	Partly. It is topical fit, not keyword coverage
Position in the model's context	2	Strong effect	No. The retrieval system sets it
Brand prominence	All	Strong association	No. It is a confounder, not a lever
Organic rank	2	Moderate, decaying proxy on Google	Indirectly
Bing index presence	1	Documented upstream, effect unquantified	Cheap hygiene with no promised effect
Freshness	2	Exposed to engines, not shown to be weighted	Not as a lever
Content rewriting for citation	3	Strong evidence of no effect or harm	Don't
Schema markup	3	No lift in the one controlled test	Not for citations
An E-E-A-T "score"	None	No such score exists	No
Domain Authority	None	Third-party metric; no engine publishes one	No
llms.txt	None	No engine documents reading it	No

What survives

Access gates: strong, and yours

The highest-confidence factors in the whole evidence base are the boring ones. The retrieval bot is allowed in robots.txt. The firewall and CDN let that bot through. Load-bearing content is in the server-rendered HTML. The page is indexed and eligible for a snippet. Each is documented by the operator of the engine, each is checkable from outside, and failing any one of them removes you from that engine's search-backed answers. Passing them buys eligibility, not preference. That is still the best return on an hour of GEO work, because one failed gate wastes every other effort.

Topical relevance to the retrieval query: moderate

Two sources independently name it the most reproducible driver of citation among retrieved documents. The 45-study survey concludes that "topical relevance and context position are the most reproducible levers," and the 252,000-trial factorial found that "topical relevance and list position are the biggest drivers of being cited first." Discount both. The survey is single-author and not peer-reviewed, and the factorial came from a vendor team that ships its own GEO checklist and ran on a synthetic corpus with two candidates per trial.

Relevance is not keyword coverage. Google's AI optimization guide (July 2026) says you don't need to worry about "long-tail" keywords or every variation of a query, and the original GEO paper found keyword stuffing lowered its own visibility metric. Relevance means you have a page whose subject is the thing the buyer is asking about, stated plainly. Note the word retrieval, too: the engine rewrites the question first, so relevance is judged against a query you never see.

Position in the context: strong, no lever

Where a document sits in the model's context has a large effect on whether it gets used. C-SEO Bench found that moving a document up in the context was significantly more effective than the rewrites it tested. But in that study, "moving it up" meant the researchers placed it first by hand. No index was queried and no ranking was earned. You cannot do that. Position explains why content edits underperform. It is not a service anyone can sell you.

Brand prominence: strong association, not a lever

Well-known sources take most citations. A study of AI Search Arena data (arXiv:2507.05301, March to May 2025) found the top 20 most frequent news sources accounted for 67.3% of citations for OpenAI models. Semrush found only 36 brands visible across all four AI surfaces it tracks every month, among them YouTube, Amazon, Walmart and Disney. That is real, you cannot pull it this quarter, and it contaminates almost every correlational GEO study. Use it to pick fair comparisons: benchmark yourself against brands your size.

Organic rank: a moderate, decaying proxy

On Google, ranking still helps. Ahrefs found 37.9% of URLs cited in AI Overviews also ranked in the top 10 (March 2026). BrightEdge reported a very different figure with a proprietary method and no published sample size, so the two do not corroborate each other. Rank is a proxy for getting into Google's candidate set, not a guarantee, and it tells you nothing about engines that search other indexes.

What failed or is unproven

Content rewriting for citation. The best-designed test, C-SEO Bench, found most rewriting methods ineffective and often harmful. The 252,000-trial factorial found that "formatting-only edits have little impact." The 45-study survey warns that "citation-oriented rewrites can impair retrieval." This is a within-context result. It does not prove content quality is irrelevant. It shows that rewriting a page to look more citable is not a strategy.

Schema markup. Ahrefs' matched-control study (May 2026) followed 1,885 pages that added schema, compared them with matched control pages, and found "no major uplift in citations on any platform," with a statistically significant 4.6% decline on AI Overviews. Google says structured data isn't required for generative AI search. One honest gap: in Ahrefs' study every treated page already had 100 or more AI Overview citations, so whether schema helps a page get in the first time is untested, not refuted. Google also retired FAQ rich results in May 2026.

An E-E-A-T score. Google states E-E-A-T "isn't a specific ranking factor," and its Search Quality Rater Guidelines (September 2025) say using rater scores to position results "would not be feasible." There is no number to raise. Named authors and checkable claims are good practice for readers. Any tool that shows you an E-E-A-T score made it up.

Domain Authority. Moz's Domain Authority and Ahrefs' Domain Rating are vendor estimates built from backlinks and traffic. Google publishes no site-level authority score and does not use the term. Raising DA is not a GEO goal.

llms.txt. Google's AI optimization guide (July 2026) says Google Search ignores AI text files, and that adding one "will neither harm nor help" your visibility there. No other engine documents reading one. It is harmless, and nobody has shown it does anything.

Freshness as a lever. Anthropic's search tool exposes a page's age to the model, so at least one engine can see the signal. Nobody has shown it is weighted. Seer found 75% of cited pages had been updated within a year (July 2026), but it never compared them with pages that were not cited, so the figure may describe the web rather than a preference. One factorial result did find a recent timestamp helped. Treat freshness as unresolved: keep your dates honest, and never bump a date to game a signal nobody has shown exists.

Worked example: sorting a vendor's checklist

Common mistakes

• Reading a Stage 3 study as a Stage 1 promise. Before you trust an effect size, ask whether the study ever called a live search engine. Almost none did.
• Copying what big brands have. Big brands share many traits. Brand size, not the trait, is the likeliest explanation for their citations.
• Ripping out existing schema in a panic. Google says existing markup causes no harm. Removing it is as unsupported as adding it for citations. Keep it for what it still does in Google Search, and make sure it agrees with your visible text.
• Bumping dates. A refreshed date on an unchanged page games your own measurement and tells an engine nothing true.
• Concluding that nothing works. Access gates, fact consistency and off-site corrections all survive. The evidence narrows what you do. It does not tell you to do nothing.