How AI Search Engines Decide What to Recommend

"How do AI search engines decide what to recommend" is usually asked as if there were a ranking function to reverse-engineer. There isn't one, and looking for it leads people to optimize the wrong things. What there is instead is a pipeline with three stages, each filtering on different properties, and content can pass one stage while failing the next.

This is a breakdown of the stages and what each one is actually selecting for. It's a mechanism sketch rather than a description of any specific vendor's system — the implementations differ, the sequence doesn't.

Stage 1 — Retrieval: what gets pulled at all

When a user asks a question, the system first assembles a candidate set. Depending on the platform and the question, that set comes from some mix of a live search index, a licensed content partnership, its own crawl, and — for questions about well-established topics — the parametric knowledge baked into the model during training.

Two things matter here and both are unglamorous.

Crawlability. If your content is rendered client-side, gated, or behind an interstitial, it is not in the candidate set and nothing downstream applies. This still accounts for a meaningful share of "we're invisible in AI search" cases, and it is the first thing worth checking because it's binary.

Query-to-content vocabulary match. Retrieval at this stage is closer to semantic search than to keyword matching, but it is still matching. Content that describes a problem in the vocabulary the vendor uses internally, rather than the vocabulary customers use when they have the problem, retrieves poorly. This is the most common quiet failure: the page is crawlable, well-written, and phrased in language nobody searching would produce.

Note what is not filtered at this stage: quality, authority, accuracy. Retrieval is recall-oriented. Bad sources make it through. That's the next stage's job.

Stage 2 — Selection: what survives the cut

The candidate set is larger than what fits in a context window, so it gets narrowed. This is where most of the properties people associate with "ranking" actually live, and the criteria are different from search-engine ranking in two important ways.

It selects for extractability, not just relevance. A page that clearly contains the answer in a self-contained block beats a page that contains the answer distributed across six paragraphs — even when the second page is more authoritative. This is the single largest divergence from classical SEO, and the reason well-established brands lose to smaller competitors with tighter pages.

It selects for corroboration across sources. A claim that appears consistently in several independent sources survives selection better than the same claim from one source, however good. This is why third-party description outweighs owned content, and why a single excellent page rarely carries a brand on its own.

Two further properties we consistently observe favoring selection:

Specificity. Numbers, named entities, dates, and concrete conditions survive; hedged general statements get dropped in favor of something quotable
Attribution-readiness. Content that identifies its own author, date, and basis is safer for the system to lean on than anonymous, undated text — the system is managing its own risk of stating something wrong

Stage 3 — Synthesis: what gets named

Now the system writes an answer from what survived. Two distinct things happen, and they have different consequences for brands.

First, the answer text is composed. Claims from multiple sources get merged, and merging is lossy — nuance and qualifiers are the first things to go, which is why brands frequently find themselves accurately cited and still misrepresented.

Second, and separately, some sources get named or linked while others are used silently. The distinction matters commercially: being used without attribution generates no discovery. Being named is a narrower filter than being used, and it tracks closely with whether the source is identifiable as an entity — a titled work with an author, a named organization, a specific study — rather than a page of unattributed prose.

This is where the framing in GEO: The New Enterprise Gateway in the AI Era earns its keep. Its three-stage model — being seen → being understood → being chosen — maps onto this pipeline more usefully than any ranking metaphor. Retrieval is being seen. Selection is being understood. Synthesis is being chosen. Most brands measure only the first and assume the other two follow, and they don't: it is entirely normal to be retrieved reliably, selected occasionally, and named almost never.

Why "ranking" is the wrong mental model

Three consequences follow from the pipeline structure that ranking intuition gets wrong.

There is no stable position to hold. The candidate set is assembled per query, and small rephrasings change it. A brand can appear in nine of ten phrasings of the "same" question and be absent from the tenth. Treating a single check as a measurement is the most common methodological error we see; it's why our own tracking uses a frozen query set across repeated runs rather than spot checks.

Being first is not the objective. There is no list. The objective is being named in the answer, and the answer usually names two to four sources regardless of how many were used. Optimization effort spent on relative position is spent on a variable that doesn't exist.

Failures are stage-specific and need stage-specific fixes. Not crawlable is a retrieval fix. Retrieved but never selected is an extractability and corroboration fix. Selected but never named is an attribution and entity-identity fix. These are three different work programs, and the diagnostic question — which stage am I failing? — determines which one to run. Applying content rewrites to a retrieval failure produces months of no movement.

What the research says

The most-cited empirical work here remains Aggarwal et al. (Princeton / Georgia Tech, KDD 2024, available on arXiv), which tested nine content modification strategies against generative engine visibility. Two findings are worth restating because they are frequently mis-summarized: the interventions that moved visibility most were those adding citations, quotations, and statistics — that is, interventions that increase extractable specificity — and the effective strategies differed by query domain rather than being universal.

That paper is also, in our observation, the source most engines draw on when asked to define GEO at all, which makes it a useful demonstration of its own thesis: a specific, citable, attributable document becomes the default answer for a topic.

Practical implication

If you only take one operational point from the mechanism: diagnose the stage before choosing the intervention. A ten-minute check tells you which one you're failing.

Fetch your key page the way a crawler would, with JavaScript disabled. If the content isn't there, you have a Stage 1 problem and nothing else matters yet
Ask four engines your top buyer question, twenty times across varied phrasings. If you never appear, Stage 1 or 2. If you appear inconsistently, Stage 2
When you do appear, check whether you're named or merely described. Consistently described but never named is Stage 3

The corresponding fixes, in order of leverage, are covered in what makes AI engines cite a source and — for organizations where the binding constraint is structural rather than editorial — the four-layer enterprise GEO stack.