Say Hello

HOW AI ENGINES
CHOOSE WHAT TO CITE.

Most articles on AI citation answer the question with a single factor (Bing index, schema, freshness). The reality is five mechanisms working in parallel, weighted differently across the four major surfaces. This is the framework.

AI Surfaces
4
Citation Factors
5
Sites Studied
47
Methodology
Open

WHY THIS QUESTION IS HARD TO ANSWER.

Ask ten SEO blogs how AI engines decide what to cite and you will get ten different one-line answers. "Bing index." "FAQ schema." "Recency." "Authority." Each is partially right and individually misleading. The truth is that five distinct mechanisms operate in parallel, weighted differently across the four major surfaces, and no single factor dominates.

This article is the framework I use in audits. It is not a list of tactics. It is the map you need before tactics make sense. Once you know which mechanisms are at play, you can stop optimising for one factor at the expense of the four others.

Caveat up front: AI engines are opaque. Nothing here is leaked from a vendor; everything is observed at the network level across 47 instrumented sites and confirmed against citation-tracking runs over six months. The mechanisms are real. The relative weights are estimates.

Finding 01.
The discovery layer

MECHANISM 01: BING INDEX DEPENDENCY.

ChatGPT's web retrieval is heavily Bing-dependent. Claude's web search uses a mix of Bing and Brave. Perplexity has its own crawler but cross-references Bing for fallback. Google AI Overviews uses Google's index by definition.

The implication: if you are not in the Bing index, three of the four major AI surfaces have a structural reason to skip you. This is the easiest factor to fix and the most commonly broken. Bing Webmaster Tools is free and takes 30 minutes to set up.

What this does NOT mean: Bing-indexed sites are automatically cited. Indexing is a precondition, not a guarantee. The other four mechanisms decide what gets surfaced from the index.

Finding 02.
The extraction layer

MECHANISM 02: CONTENT EXTRACTABILITY.

AI engines fetch HTML, parse it, and look for content that can be lifted into an answer. If your content is rendered client-side via JavaScript and is not present in the initial HTML response, most AI fetchers see an empty page. If your content is in <div> soup with no semantic structure, the parser struggles to identify what is body, what is sidebar, what is ad.

The signals that help extraction: server-rendered HTML, semantic tags (<article>, <section>, <header>), clean heading hierarchy (one H1, descriptive H2s), short paragraphs near the top of the page that answer the question directly.

The signals that hurt extraction: SPA frameworks without SSR, infinite scroll, content gated behind interactions (clicks, hovers), schema that contradicts the visible page, machine-translated content with awkward phrasing.

Finding 03.
The grounding layer

MECHANISM 03: AUTHORITY AND ENTITY GROUNDING.

AI engines need to anchor citations to real entities: a real person, a real organisation, a real product. The cleaner your entity grounding, the more confidently the engine cites you. This is where schema markup, named authorship, and consistent identity references actually pay off.

The minimal entity stack: Organization at site root with stable @id, Person for each named author with stable @id, sameAs links to social profiles and Wikidata where applicable, consistent NAP (name/address/phone) data, schema that mirrors visible content.

Wikipedia citations are the gold standard here, which is why Wikipedia gets cited disproportionately - its entity grounding is essentially perfect. Most other sites can close the gap by treating their Organization and Person schema as foundational, not decorative.

Finding 04.
The freshness layer

MECHANISM 04: RECENCY WEIGHTING.

Different surfaces weight recency differently. ChatGPT and Perplexity surface recent content aggressively for time-sensitive queries ("latest", "2026", "yesterday"). Claude is more conservative, often citing older sources for stable factual queries. Google AI Overviews follows traditional Google freshness signals.

The mechanism: recency comes from article:published_time meta tags, datePublished in schema, lastmod in sitemap.xml, and feed timestamps. When these disagree (which they often do on poorly-maintained sites), the engine picks one and may surface stale content as fresh or vice versa.

Stale content disappears. Sites that go silent for 60+ days see their citation rate drop to near-baseline within roughly two weeks. Freshness is the dimension that compounds the most over time.

Finding 05.
The specificity layer

MECHANISM 05: TOPICAL SPECIFICITY.

AI engines surface specific, well-scoped content over generic content. A page titled "How to set up MX records for Cloudflare Email Routing in 2026" will out-cite a page titled "Email setup guide" for the same query, even if the second page is technically longer and better-linked.

The mechanism: specificity gives the engine confidence that the page actually answers the query. Generic pages create ambiguity ("is this really about MX records?") and engines hedge by citing more confident sources.

This is also why long-tail queries return more diverse citations than head-term queries. Head-term queries ("AI search", "SEO") are saturated with generic content; engines fall back to Wikipedia. Long-tail queries ("OAI-SearchBot vs GPTBot rate limits") have fewer generic pages competing, so specific authoritative pages win.

FOUR ENGINES. FIVE MECHANISMS.
NO SINGLE FACTOR DOMINATES.

HOW THE FOUR SURFACES DIFFER.

The five mechanisms are constants. The weights are not. Each surface skews differently:

WHAT TO PRIORITISE.

If you have to optimise in order, the priority sequence that has held up across 47 sites and six months of citation tracking is:

THE BOTTOM LINE.

There is no single factor. There are five, weighted by surface. Optimising for one at the expense of the others is the single most common mistake practitioners make, and it is what produces sites that score 8 on schema and still earn no citations because they failed extraction.

Audit on all five. Fix in priority order. Re-measure quarterly. The framework is open; the discipline of applying it consistently is what separates sites that get cited from sites that wonder why they don't.

Stop Guessing What AI Sees

MEASURE THE LEVERS
THAT ACTUALLY EXIST.

If you want this methodology applied to your specific site - your real logs, your real citation data, your real fix list - the audit is the productized way to do it.