The 10 gates: how AI search engines actually decide what to cite.

Why “the page” is the wrong unit

For twenty years, the unit of search optimization was the page. SEO tools scored pages, ranked pages, recommended page-level fixes. The algorithm rewarded pages with strong overall signals.

AI search engines do not work this way. They work at the level of the passage — the discrete 40–150 word block of text that answers a specific question. A single page may contain twenty extractable passages, each competing independently for citation. A page that ranks well in Google AI Mode may have only one of its passages cited; the other nineteen contribute nothing to the eventual answer.

The shift from page-level to passage-level optimization is the most under-discussed structural change in the discipline. Most GEO advice still operates at the page level because that’s where SEO operated and where most practitioners’ mental models live. Operators who internalize the passage-level reality — who write content as a sequence of self-contained answer blocks rather than a continuous narrative — capture disproportionate citation share.

To understand why, you need to look inside the engine.

The 10 gates

Search Engine Land published the cleanest framework for the AI engine pipeline in 2025: every piece of content must pass through ten distinct gates before it can be cited in an AI response. Most operators optimize for one or two of them. The brands that earn citation share understand all ten.

Most SEO advice still focuses on gates 2–5. Most GEO marketing focuses on gates 9–10. The biggest untapped opportunity is gate 6 — annotation — where structured data, entity markup, and semantic clarity decide whether the AI engine understands what your content is about.

Search Engine Land’s framing was direct: gate 6 is “the biggest untapped opportunity in search, assistive, and agential optimization right now.”

“Most SEO advice operates at gates 2–5. Most GEO advice operates at gates 9–10. The biggest opportunity is gate 6.”

RAG explained operationally

Behind every gate is the same core technical pattern: Retrieval-Augmented Generation, or RAG. RAG is the engineering reality that determines what AI engines can and cannot cite, and how content needs to be structured to participate.

The RAG pipeline, in operational terms:

1. Query understanding. The user’s prompt is parsed for intent. AI search prompts average 23 words versus 2–3 for traditional search — the engine has more to work with.
2. Query fan-out. The single user prompt is broken into multiple sub-queries. A question about “best CRM for a SaaS startup” might fan out into queries about pricing, integration depth, ease of use, and customer support, each retrieved separately.
3. Retrieval. Each sub-query hits the search index. Candidate documents are fetched.
4. Chunking. Retrieved documents are broken into passages. The engine doesn’t reason over your whole page — it reasons over the slices it can extract.
5. Embedding. Each passage is converted into a vector embedding — a numerical representation of its meaning in high-dimensional space.
6. Scoring. The user’s prompt is also embedded. Cosine similarity between query embedding and passage embeddings determines which passages get pulled into the model’s context window.
7. Synthesis. The LLM generates the final answer from the top-scoring passages, attaching citations to the source documents.
8. Display. The user sees the synthesized answer with inline citations.

The implication for content: if your content can’t be cleanly chunked into 40–150 word passages with self-contained meaning, it never makes it past step 4. The engine throws away most of your page in favor of the passages it can extract.

The Wu et al. citation accuracy crisis

The most consequential under-cited finding in the AI search literature comes from Wu et al., published in Nature Communications in April 2025. The researchers tested whether the citations AI engines attach to their answers actually support the claims being made.

The result: 50–90% of LLM citations don’t fully support the claims they’re attached to.

Read that twice. The citation footnotes in AI responses are frequently fictional in their support relationship. The cited source exists, but doesn’t actually back up the specific claim it’s attached to. The user sees an authoritative-looking citation; the underlying source is at best tangentially related.

A complementary finding from Venkit et al. (arXiv 2024): citation accuracy across major AI search platforms ranges from ~66% (best) to under 50% (worst). Even the best-performing platforms get the citation relationship wrong a third of the time.

The downstream effect: users see citations and trust the answer. 82% of users are skeptical of AI results in principle, but only 1% click into the cited source to verify (Pew). The citation acts as a credibility marker even when it doesn’t do the work of credibility. AI engines have, accidentally or not, evolved a system where the appearance of sourcing substitutes for actual sourcing.

The crawl-to-cite ratio

Cloudflare’s 2025 telemetry on AI bot behavior surfaced a number that reframes how operators should think about visibility. ClaudeBot’s crawl-to-cite ratio is approximately 38,065:1. For every page Anthropic’s crawler reads, it cites approximately one in 38,000 in an actual user-facing response.

The other major AI crawlers operate at similar orders of magnitude. GPTBot grew 305% in crawling volume from May 2024 to May 2025, becoming the dominant AI crawler at 30% of all AI bot traffic. The volume of crawling is enormous; the volume of citation is comparatively tiny.

The implication: your content being crawled is not visibility. The crawl is the entry ticket; everything between the crawl and the citation is where the real competition happens. Brands obsessed with bot-traffic dashboards are watching the wrong number. Citation share — how often your content is the one in 38,000 that gets surfaced — is the real metric.

This also explains why analytics-only GEO platforms see so much traffic data without obvious correlation to outcomes: there’s an enormous funnel between the crawl event and the user-facing citation, and most of the funnel is invisible to the brand.

Per-platform architecture differences

Each major AI engine implements RAG with different architectural choices. The differences matter because they determine which optimization moves work on which platforms.

The architectural divergence is meaningful for optimization strategy. Perplexity’s document-and-sub-document scoring means individual passages within a page compete independently — passage-level structure matters more here than anywhere else. ChatGPT’s heavy Bing dependency means that classical SEO investment in Bing visibility carries through directly to ChatGPT visibility. Google AI Overviews’ 93.67% correlation with organic top-10 means SEO foundations are the prerequisite, not a bonus.

The cross-platform implication is the same conclusion the published research keeps surfacing: only 11% of domains receive citations from both ChatGPT and Perplexity. The same content, run through two different engines, lands in two different citation pools. A unified GEO strategy has to anticipate the architectural differences, not assume them away.

Why query fan-out changes content design

Query fan-out — the practice of decomposing a single user prompt into multiple sub-queries — is the architectural detail with the largest practical implication for how content should be written.

When a user asks ChatGPT for “the best project management tool for remote design teams,” the engine doesn’t run that exact query. It fans out into a series of sub-queries:

• Best project management tools 2026
• Project management for remote teams
• Project management for design teams
• [Specific tool] reviews (for each tool surfaced in the first three queries)
• [Specific tool] vs [specific tool] (for comparative framing)
• Project management tool pricing

The synthesized answer is built from passages retrieved across all of these sub-queries. A page optimized for the literal user prompt is competing in only one of the seven retrieval rounds.

The operational shift: content needs to satisfy multiple sub-queries simultaneously. A pricing-focused passage, a use-case-focused passage, a comparison passage, and a feature passage on the same product page each compete in different sub-queries. Pages that bundle these dimensions earn citation share across more retrieval rounds. Pages that focus narrowly compete in one round and are absent from the rest.

The annotation gap

Returning to gate 6 — the annotation gate Search Engine Land called the “biggest untapped opportunity” — the practical work has three components.

Structured data. JSON-LD schema markup that gives AI engines a pre-digested summary of what each page is about. Organization schema. Product schema. FAQ schema (with the caveat that direct citation lift from FAQ schema is debated; the indirect benefit through entity grounding is real). Article schema with author and publication date. Most pages have minimal or generic schema; pages with comprehensive, accurate schema are interpretable in ways unstructured pages aren’t.

Entity markup. Explicitly defining the entities your content is about — brands, products, people, places — with consistent identifiers across owned properties and to external entity databases (Wikidata, Google Knowledge Graph). The Onely study found that LLMs grounded in knowledge graphs achieve 300% higher accuracy than those working with unstructured text alone.

Semantic clarity. Heading hierarchy that maps logically to the content’s argument structure. Passage breaks that align with discrete sub-topics. Internal linking that explicitly states relationships between related concepts. None of this is new SEO advice. The difference is that AI engines reward it more directly than ranking algorithms ever did.

The Reddit and Wikipedia gravitational pull

The source distribution data tells a structural story about where AI engines pull from when they synthesize answers. Two domains dominate disproportionately:

• Reddit: 40.1% of LLM citations across all platforms (Averi AI). For Perplexity specifically, 46.7% of top sources come from Reddit (Digital Bloom).
• Wikipedia: 26.3% of LLM citations across platforms. For ChatGPT specifically, 47.9% of top sources come from Wikipedia.

Neither was a primary SEO target before AI search. Both are now first-class citation real estate. The mechanism is structural: Reddit is conversational, multi-perspective, written in the language users use; Wikipedia is encyclopedically structured with explicit entity definitions. Both are exactly the kind of content RAG architectures process most cleanly.

The operational implication for brands: presence on these platforms isn’t optional for serious AI visibility programs. Not in the form of astroturfing — community platforms detect and punish that — but as authentic participation, third-party coverage, and entity grounding. Brands that are authentically discussed on Reddit get cited by Perplexity. Brands that have well-maintained Wikipedia entries get cited by ChatGPT. Brands that have neither compete uphill regardless of how good their owned content is.

What the engineering means for operators

The architectural details add up to four operational shifts every brand serious about AI visibility should internalize:

1. Optimize for the passage, not the page. Write content as a sequence of 40–150 word self-contained answer blocks, each tied to a specific sub-query the engine might fan out into. Pages that read well as continuous narrative get cited less than pages structured as extractable answer units.
2. Invest in gate 6 — annotation. Schema markup, entity grounding, knowledge graph submissions. This is where the largest under-claimed competitive opportunity lives. Most brands have done little here; the brands that invest pull ahead.
3. Track citation share, not crawl traffic. The 38,065:1 ClaudeBot ratio is a reminder that crawl events are not visibility events. The metric that matters is how often your content is the one passage in 38,000 that the engine actually surfaces in a user-facing response.
4. Build presence where AI engines pull from. Authentic Reddit participation, well-maintained Wikipedia/Wikidata entries, third-party editorial coverage on the publications each engine over-indexes on. The off-site signal layer is doing more of the work than most operators realize.

The brands that internalize the engineering reality — that AI engines are RAG pipelines with specific architectural constraints, not magic boxes — will out-execute the brands operating from a 2018 SEO mental model. The gap is the technical literacy. The work is the application.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.