How AI engines pick which products to cite.

AI-engine citation feels like a black box until you look closely. It is not. The flow is roughly three steps — retrieve, ground, rank — and each step runs on signals you can understand and mostly control. The specifics differ across ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude, and Copilot, but the shape is shared, and once you know the shape, the PDP work worth prioritizing becomes obvious.

Retrieval: query to documents

When a buyer types a shopping query, the engine first decides which web documents to pull into a working set — a mix of keyword matching, embedding-based semantic search, and, in some engines, signals from an underlying search partner. Retrieval is fast, looks at many candidates, and typically keeps 20 to 200 pages for the next step. Some engines also expand the query first — query fan-out — which is why a PDP can be cited on a question the buyer never literally asked.

If your PDP is not retrieved, nothing downstream matters. Retrievability means a clean URL, crawlable content (no JavaScript-only rendering for the parts that matter), an accurate title and H1, and language that matches what buyers actually type.

Grounding: what makes a page citeable

From the working set, the engine grounds its answer — pairing every claim it wants to make ("this jacket retails for $239 and weighs 8 ounces") with a source page that asserts those facts. Structured data makes grounding trivial. A page that declares price, SKU, brand, material, and weight in Product JSON-LD is much cheaper to cite than one that buries the same facts in marketing prose. As of mid-2026 the engines vary in strictness: Perplexity is the strictest about grounding, ChatGPT the loosest, Gemini between.

Ranking: who makes the final list

Among grounded candidates, the engine picks what to cite. Observable behavior is consistent with a mix of: how directly the page addresses the query, source authority, recency, how well the page's entity matches the one the query resolves to, and how diverse the final list should be. The practical reading: your PDP competes against review pages and marketplace listings in a contest that favors directness, authority, and freshness.

Source-trust signals

Authority is the signal brands struggle with most because it is slow to build. Roughly, the engines look at: domain reputation inherited from traditional web ranking, schema-level cues (Organization with verified sameAs, Product with GTIN), historical citation patterns, and user feedback on the engine itself. A new brand starts thin; a ten-year-old brand starts thick. Authority is not the only input, but it is the one that takes quarters to move — citation analysis shows how it decomposes inside the citation score.

Entity clarity

Entity clarity is the engine connecting your page to a real-world product. A PDP titled "Patagonia R1 Air Full-Zip Hoody" with matching JSON-LD, a canonical URL, and a consistent brand value is unambiguous. A PDP titled "New Arrival" with a vague description is an identification problem. Engines can still cite the ambiguous page — but often to the wrong query, or not at all. When grounding fails badly enough, the engine asserts things about your product that are not true; that failure mode has its own name, hallucination detection, and its own monitoring.

Why review coverage amplifies everything

Review pages are structurally built to answer comparison queries. When an engine retrieves a roundup that includes your product, it often adopts you as a recommended option from that page — then optionally fetches your PDP as a secondary source. One review-site inclusion can lift a SKU from invisible to cited across dozens of related queries. That is why review pitching is disproportionately cheap as a GEO motion: an amplifier on top of your own PDP work.

What this means in practice

Three moves fall out of the mechanism. Make retrieval easy: clean titles, crawlable content, buyer language. Make grounding trivial: complete Product JSON-LD, FAQ schema, unambiguous identification — the schema.org Product spec and Google's Product structured data reference are both worth a team read. Extend authority: review coverage, consistent brand signals, real aggregateRating data. Nothing exotic; everything rewarded by consistency. The same three moves raise both scores eCommerce Insights assigns each SKU — citation and agent-readability — because an engine that can ground your page and an agent that can parse it are reading the same fields. The SKU-level AEO guide turns the mechanism into a checklist.

What is still unknown

The honest hedge: the exact weights each engine applies, and how fast they shift with model updates. A retrieval change at ChatGPT in February can reshape citations for a quarter. Public documentation is patchy — Google's is clearest, Perplexity's help center has useful pieces, ChatGPT's working model is largely inferred from behavior. Treat the three-step shape as stable and the formula as moving; that is the case for weekly measurement rather than assumptions.

Key takeaways

• Engines run retrieve → ground → rank. Each step uses signals you can shape.
• Retrieval rewards crawlable pages in buyer language. Grounding rewards structured data. Ranking rewards directness, authority, recency.
• Entity clarity connects your PDP to the right query; ambiguity costs citations or invites hallucination.
• Review-site coverage is the cheapest amplifier in the system.
• Weights shift per engine update. Weekly cadence catches the drift.