How AI Retrieval Selects Your Page: The Six Gates

Quick answer

Your page passes through six gates before it can be cited in an AI answer, and it can be eliminated at any of them. The system decides whether to search at all, rewrites the user’s question into its own queries, retrieves candidates, reranks them by whether they actually answer, fits a handful into a limited context budget, and finally resolves citations during rendering. Most content work targets gate three. The two that eliminate the most pages are gate one, which you cannot influence, and gate four, which almost nobody optimises for.

I have read a great many pieces on AI visibility and almost none of them describe the mechanism. The best-ranking ones talk about making content “interpretable, verifiable and reusable by large models” — language that sounds technical and specifies nothing you could act on. One of the more widely cited overviews of AI visibility platforms explains how to track visibility across ten tools and never explains how a source gets chosen.

So here is the pipeline. Not a metaphor — the sequence of decisions, what eliminates you at each, and which of them you can actually influence.

The six gates

user question
  │
  ├─ GATE 1  Does this need retrieval at all?
  │            → if no: answered from weights. You never existed.
  │
  ├─ GATE 2  Query formulation
  │            the system writes its OWN search queries
  │
  ├─ GATE 3  Candidate retrieval
  │            hybrid vector + lexical, top 50-200 chunks
  │
  ├─ GATE 4  Reranking
  │            cross-encoder: does this chunk ANSWER the question?
  │
  ├─ GATE 5  Context assembly
  │            3-10 chunks survive. Budget is finite.
  │
  └─ GATE 6  Generation + citation resolution
               model may use a chunk without attributing it

Six gates. Content strategy in this field overwhelmingly addresses gate three. That is one gate out of six, and not the one eliminating most candidates.

Gate 1: whether it searches at all

Before anything else, the system decides whether this question requires external information. Definitional, conceptual and general-knowledge questions are frequently answered from parameters. No search runs. No source is consulted. There is no competition to enter.

This is the single largest filter and the one nobody measures.

Query characteristic	Retrieval likely?	Implication
Asks for current pricing, availability, dates	Yes	Freshness demands grounding
Names a specific product, company or version	Usually	Entity specificity triggers lookup
Asks “best” or “compare” with a live market	Usually	Commercially the most valuable class
Asks what a concept means	Often not	Answered from weights. Being cited is unlikely
Asks for an explanation of a stable idea	Rarely	You are competing with training data, not other pages
Asks for code, maths, reasoning	Rarely	Generated, not retrieved

Directional rather than absolute — routing rules are undisclosed and change. The pattern of freshness and specificity triggering retrieval is stable.

The uncomfortable part

A large share of classic informational content — the “what is X” explainer that has been the backbone of content marketing for fifteen years — sits in query classes that frequently do not trigger retrieval. You can write the best definitional page on the internet and be structurally excluded from the answer, because the system never looked. No amount of schema, structure or authority changes a routing decision. This is the part of AI visibility that is genuinely outside your control, and I think the industry avoids saying so because it is not a sellable problem.

Gate 2: the system rewrites the question

If retrieval is triggered, the user’s question is not what gets searched. A model rewrites it — frequently into several queries — and those are what hit the index.

User asks:
  "is it worth switching from mailchimp for a small list"

System might issue:
  → "Mailchimp alternatives small business"
  → "Mailchimp pricing contact tiers 2026"
  → "email platform migration cost small list"
  → "Mailchimp send limit contact tier"

Your page is matched against THOSE, not the original.

Consequence: optimising for how a human phrases a question
is optimising for a string the index never sees.

This quietly invalidates a chunk of prompt-level AEO advice. Targeting conversational phrasing assumes the conversational phrasing is the query. It is the input to a query generator whose output you cannot observe.

What survives rewriting is entity names, product names, technical terms and numbers — the tokens a rewriter preserves because they carry the specificity. Which is an argument for naming things precisely rather than writing around them.

Gate 3: candidate retrieval

The gate everyone optimises for, and the least selective of the six. Hybrid retrieval pulls perhaps fifty to two hundred chunks: a vector path catching paraphrase, a lexical path catching exact strings, results fused.

Two things to hold onto. First, retrieval operates on chunks, so what competes is a fragment of your page rather than your page. Second, this stage optimises for recall — it is deliberately generous. Getting retrieved here is necessary and close to meaningless on its own. A hundred candidates enter; three to ten survive.

Gate 4: reranking — where it is actually decided

This is the gate that matters most and receives the least attention.

Gate three used bi-encoders: your chunk and the query were embedded separately and compared. Fast, and the two were never examined together. Gate four uses a cross-encoder that processes the query and the chunk jointly and scores whether that passage answers that question.

The reordering is aggressive. Chunks that ranked first on similarity routinely get demoted because topical proximity is not responsiveness.

Chunk characteristic	Similarity (gate 3)	Reranker (gate 4)
Discusses the topic at length, answers nothing	Scores well	Demoted hard
States the answer in one sentence, then explains	Scores moderately	Promoted
Answer split across the chunk boundary	Scores well	Demoted — half an answer reads as no answer
Answers a related but different question	Scores well	Demoted
Requires prior context to make sense	Scores well	Demoted — reads as incoherent alone
Contains hedging without resolution	Neutral	Demoted — “it depends” is not an answer

Rows three and five are the same underlying failure: chunk boundaries falling in the wrong place. That is a structural property of your document, not a content-quality property.

My position

The last row is my favourite and the one I would argue hardest for. Writers hedge because hedging is honest and because committing is risky. Rerankers penalise it, because a passage that carefully declines to conclude does not answer the question. There is a real tension here between intellectual honesty and retrievability, and I do not think it resolves cleanly. My compromise: state the answer, then qualify it. The commitment comes first so the passage is responsive; the nuance follows so it is honest. Hedge-first construction loses both.

Gate 5: the context budget

Surviving reranking is still not enough. A finite number of chunks — usually a handful — are placed into the prompt alongside system instructions, conversation history and memory.

Two effects worth knowing. Long chunks consume more of a fixed budget, so an extremely long passage can be dropped in favour of two shorter ones that between them cover more ground. And position within the assembled context is not neutral — material at the beginning and end of a long context is attended to more reliably than material in the middle, a pattern documented in the literature on long-context behaviour. Where your chunk lands in the assembly affects whether it is used, and you have no influence over that whatsoever.

Gate 6: used but not credited

The final failure is the most galling. Your chunk made it into context. The model read it. The answer reflects it. And no citation to you appears.

The model synthesised across sources and attributed the combined statement to the one it considered most authoritative.
The fact appeared in several retrieved chunks, so it was treated as common knowledge requiring no attribution.
Citation resolution failed during rendering — the link did not map cleanly back to a source.
Citation count was capped and yours fell outside the limit.

The second bullet has a strategic implication I find genuinely uncomfortable: publishing the same widely-known facts as everyone else makes you uncitable. If a claim appears in six retrieved sources, no single source gets credit. Being the only place a specific number, framing or piece of data exists is not merely a differentiation tactic — it is a mechanical precondition for attribution.

What this means for what you do

Gate	Can you influence it?	How
1. Search triggered	Barely	Target query classes that require freshness or specificity
2. Query rewriting	Indirectly	Name entities, products and versions explicitly — those survive rewriting
3. Retrieval	Yes	Hybrid coverage: exact terminology plus natural paraphrase
4. Reranking	Yes, most of all	Self-contained chunks that answer directly in the first sentence
5. Context budget	Somewhat	Moderate chunk length; do not bury the answer in a wall
6. Citation	Yes	Publish claims that exist nowhere else. Uniqueness enables attribution

Gates four and six are where the leverage is, and they receive the least attention in published AEO advice.

What I would stop doing

Writing pages targeting conversational phrasings. Gate two rewrites them. You are optimising for a string that never reaches the index.
Treating retrieval as the finish line. It is the most permissive gate. Being retrieved and being cited are separated by a reranker that eliminates most candidates.
Adding length for comprehensiveness. Longer chunks compete worse for a fixed context budget, and cosine cannot see depth anyway.
Republishing consensus. A fact present in six sources is attributable to none. This is the mechanism behind “add original insight” — it was always true and now it is structural.
Assuming a ranking loss is a content problem. Gates one, five and six can eliminate you for reasons entirely internal to the vendor.

The limit of all this

I am describing a general architecture. No vendor publishes their routing thresholds, rewriting prompts, chunking parameters, reranker or citation policy, and all of it changes without notice. Anyone claiming to know the specific configuration of a given assistant is selling something.

What is durable is the shape: multiple sequential filters, each with a different objective, most of them invisible. The correct response is not to reverse-engineer a configuration that will change next month. It is to write documents that pass gates three, four and six under any plausible implementation — self-contained, directly answering, precisely named, and containing something that exists nowhere else.

That is a demanding standard. It is also just a description of writing something worth citing, which is the conclusion I keep arriving at from every technical direction in this field.

The infrastructure underneath gates three and four is in vector databases, embeddings and semantic search. Why gate one exists at all follows from the architecture in ChatGPT is not an LLM, and the distinction between surviving this pipeline and ranking in classic search is in ranking and being cited are two different processes.

Frequently asked questions

How does an AI system decide which pages to cite?

Through six sequential gates: whether to retrieve at all, rewriting the question into its own search queries, retrieving candidate chunks via hybrid vector and lexical search, reranking those by whether they actually answer the question, fitting a handful into a finite context budget, and resolving citations during rendering. Elimination at any gate ends your chance.

Why is my content never cited in AI answers?

The most common reason is that the query class never triggers retrieval — definitional and conceptual questions are frequently answered from model parameters with no search performed. The second is reranking: your chunk was retrieved but a cross-encoder judged it did not directly answer the question. Both are invisible in visibility tools.

Does the AI search for exactly what the user typed?

No. A model rewrites the user’s question into its own search queries, often several, and those hit the index. Your content is matched against the rewritten queries rather than the original phrasing. What survives rewriting is entity names, product names, technical terms and numbers, which is an argument for naming things precisely.

What is reranking and why does it matter most?

After initial retrieval returns 50 to 200 candidates using fast bi-encoders that never see query and document together, a cross-encoder processes each pair jointly and scores whether that passage answers that question. It reorders aggressively, demoting chunks that are topically similar but not responsive. It is the most selective gate and the least discussed.

Why does hedging hurt AI citation?

Rerankers score whether a passage answers the question, and a passage that carefully declines to conclude does not. There is genuine tension between intellectual honesty and retrievability here. The workable compromise is to state the answer first and qualify it afterwards, so the passage is responsive at its opening and honest by its end.

Can my page be used in an AI answer without being cited?

Yes, and it is common. The model may synthesise across sources and attribute to the one it considers most authoritative, treat a fact appearing in several retrieved chunks as common knowledge needing no attribution, fail to resolve the citation during rendering, or hit a cap on how many citations it displays.

Why does publishing well-known facts make you uncitable?

Because if a claim appears in six retrieved sources, no single source receives attribution — it reads as common knowledge. Being the only place a specific number, framing or dataset exists is therefore not just a differentiation tactic but a mechanical precondition for being credited at all.

Which gates can I actually influence?

Gate four, reranking, offers the most leverage: self-contained chunks that answer directly in the first sentence. Gate six, citation, is influenced by publishing claims that exist nowhere else. Gate three responds to hybrid coverage of exact terminology plus paraphrase. Gate one, whether search runs at all, is effectively outside your control.