How AI Search Engines Decide Which Brands to Recommend

Modern generative search engines all run roughly the same five-stage pipeline.

Stage 1: Query classification

The engine first classifies the prompt. Is it commercial intent? Informational? Time-sensitive? Localized? Does it require a list of named entities, or a single recommendation? Is the prompt safe to answer at all? This stage is invisible to the user but it determines everything downstream — the retrieval corpus, the citation rules, and the response format.

A "best project management tool for design agencies" prompt is classified commercial-intent, list-format, lightly time-sensitive (the answer should reflect the current product landscape, but not last week's news), and B2B-software adjacent. That classification will route the engine to a specific corpus slice and a specific citation rule set.

Stage 2: Retrieval

The engine pulls candidate sources. There are three major retrieval modes:

• Pre-trained synthesis. The engine answers from its parametric knowledge — what the model already learned during training. ChatGPT does this for prompts where a live retrieval call is not necessary (or the prompt is "stale-ok").
• Live web retrieval. The engine issues real web requests, parses returned pages, and conditions the answer on what it just read. Perplexity does this aggressively. ChatGPT search does it on commercial-intent prompts. Google AI Overviews effectively does it because Overviews are anchored in the live Google index.
• Curated index retrieval. Some engines maintain a private index of vetted sources and pull from that index in addition to the live web. Anthropic and OpenAI both invest in this layer for higher-stakes domains like medical and finance.

Most commercial-intent prompts in 2026 use a hybrid: pre-trained synthesis as scaffolding, plus live retrieval to anchor specific named recommendations.

Stage 3: Candidate selection

From the retrieval set, the engine selects which sources actually inform the answer. Sources are ranked. Weak sources are dropped. Sources that contradict each other are reconciled (or one is discarded). The set of named brands that survive this stage is the set the user will see.

Stage 4: Citation ranking

The engine decides which subset of the selected sources to cite explicitly in the rendered answer. Some sources inform the answer without being cited — they are upstream entity-disambiguation signals only. Others are cited as evidence. ChatGPT typically cites three to eight sources per answer; Perplexity cites more.

Stage 5: Synthesis

The engine writes the response, attributing claims to citations as configured. The response is what the user reads.

A brand can be filtered out at any stage. A brand wins by clearing all five for the prompts that matter.

The single biggest predictor of whether your brand is recommended on a category prompt is how many credible third-party sources cite your brand by name on closely related prompts.

This is the AI-engine analog of the link graph that powered classical SEO, but with crucial differences:

• It is denser. A "citation" in the GEO sense includes any contextual mention — bylines, comparison tables, methodology footnotes, "as recommended by" quotes — not just hyperlinks. Engines parse text aggressively and learn entity adjacency.
• It is weighted by source quality, not just volume. A single mention in a high-trust editorial source (mainstream tech press, university research, a recognized industry analyst) outweighs dozens of mentions on syndication mills. Engines learn which sources to trust during training and reinforce those weights during fine-tuning.
• It is topic-specific. A brand cited heavily for "remote project management" may be invisible for "agency operations." Authority does not transfer cleanly across topics; it has to be earned per-topic.
• It includes negative signal. Engines penalize brands cited in scam roundups, lawsuit coverage, or low-quality SEO spam — even if the volume of mentions is high.

What this means in practice: the strongest GEO move on a new category is to seed citations into authoritative third-party sources before optimizing your own pages. A single review on G2, a Forbes mention, or a TechCrunch comparison will move the citation graph more than ten optimized blog posts on your own domain.

MentionAudit records, on every audit, the full set of cited URLs per engine per prompt. By comparing your cited domains to your competitors' cited domains, you can read the citation graph for your category off the audit directly — which is the fastest known way to discover where to work.

When an AI engine retrieves a page, the first thing it does is disambiguate. Is this page about a product, a company, an article, a person, or an event? Who is the author? When was it published? What entities are mentioned and how are they related?

Engines answer those questions by parsing structured data. Pages with valid Article, FAQPage, Organization, Product, and Person schema get unambiguous answers handed to them. Pages without schema force the engine to infer from prose, which is slower, less reliable, and produces lower-confidence parses that get filtered earlier.

The practical effect:

• A product page with valid Product schema, with aggregateRating, brand, and offers, is parseable as a product. An engine asked "best CRM for SaaS startups" can confidently include or exclude it from a list of products.
• An article with valid Article schema and author linkage is parseable as authored content. The author's identity becomes part of the citation signal.
• An Organization schema with sameAs links to LinkedIn, Crunchbase, and Wikipedia lets the engine merge your brand into a single entity. Without that linkage, your brand may be split across three or four entity records and your aggregate authority drops fractionally per-entity.

Schema is not a ranking signal in the SEO sense. It is an eligibility signal in the GEO sense. Pages without schema can still be cited, but they have to fight uphill against parsing ambiguity. In our audits, the brands that show up most often in AI Overviews almost always have clean schema on the pages being cited. The brands that do not, almost always do not.

After candidate selection, the engine has to decide which of your pages to cite as evidence. Engines preferentially cite pages that:

• Make explicit claims. "Tool X is recommended for design agencies because…" — a sentence that hands the engine a defensible pull-quote.
• Cite their own sources. A page that links to a primary source (a study, a methodology, a vendor doc) gets cited more often than a page that asserts the same claim without sourcing.
• Use comparison structure. Tables, lists, and side-by-side feature compares are over-cited relative to flat prose because the engine can extract a single row as a citation unit.
• Have crawlable canonical content. Server-rendered HTML, lang declared, no required JS for the primary content. Pages that hide content behind client-side rendering or paywalls are systematically under-cited.
• Have stable, semantic URLs. Engines remember URLs. A URL that changes every six months loses citation memory each time.

The mistake we see most often is brands that have invested heavily in SEO content (long, keyword-rich, soft-pitched) and assume that work transfers. It transfers partially. The classical-SEO tendency to bury claims behind hedged language ("many companies find that…") is the opposite of what an AI engine wants. Direct, specific, dated, sourced claims get cited. Soft claims get summarized away.

Freshness behaves differently per prompt class.

• Evergreen prompts ("what is photosynthesis"): freshness is essentially zero-weight.
• Slowly-evolving commercial prompts ("best project management tool for design agencies"): freshness is moderate. A 2023 review still counts; a 2018 review is filtered. Engines prefer published-date stamps within the last 12-24 months for commercial recommendations.
• Time-sensitive prompts ("latest GPT-5 release notes"): freshness is dominant. Pages older than days or weeks are dropped.
• News-class prompts ("did X company get acquired"): freshness is binary. Only fresh sources are cited.

The implication for GEO is that you cannot rely on a single 2022 cornerstone post to keep earning citations. The audit signal will tell you when a citation slot starts trending older — that is the cue to refresh.

In practice, a brand that gets recommended on a target prompt has cleared all four bars:

1. The citation graph already mentions the brand on the topic, in sources the engine trusts.
2. The brand's relevant pages have parseable schema that lets the engine confirm "yes, this is a Product/Article/Organization in this category."
3. The pages themselves contain crawlable, claim-rich, source-linked, comparison-friendly content that is easy to cite as evidence.
4. The cited content is fresh enough to clear the engine's freshness bar for that prompt class.

A brand that fails any one of those falls out of the recommendation set. The point of a GEO audit is to tell you, per prompt and per engine, exactly which bar you failed and on which page — and that is the spine of how MentionAudit builds recommendations.

The four-class framework holds across engines, but the weighting differs:

• ChatGPT weights pre-trained citation graph heavily. Newer brands struggle even with strong on-page work because the engine "doesn't know about you yet."
• Perplexity weights live retrieval heavily. A well-structured page published last week can earn citations on Perplexity that ChatGPT will not give for months.
• Google AI Overviews anchor in the classical SERP. If you do not rank in the top 10-20 organic positions, you are unlikely to be in the Overview citation set, regardless of how good your schema is.
• Claude is the most conservative. It declines to recommend brands when the citation graph is sparse or contested, leaving an empty answer rather than a low-confidence one.
• Gemini sits between ChatGPT and Perplexity, with a strong bias toward Google-indexed sources.

This is why a multi-engine GEO program is not optional. The same brand can be #1 on Perplexity, absent from ChatGPT, and middling on Google AI Overviews — and the optimization moves to fix each are different.

Recommendation in AI search is not a black box. It is a layered ranking pipeline, with four dominant signal classes that have predictable behaviour. A brand can be measured on each one, and a measurement gap can be mapped to a specific intervention. That is the entire spine of an effective GEO program.

If you have not seen your brand's per-engine, per-signal-class data, the gap is almost certainly larger than you think.