The three stages every AI engine runs to pick a product
Retrieval narrows the universe to a candidate pool. Ranking scores the pool. Synthesis writes the answer from the top candidates.
Every major answer engine — ChatGPT, Perplexity, Claude, Gemini, Copilot — runs the same three-stage pipeline when a buyer asks a product question. Stage one is retrieval: the engine decides which candidate documents or products could possibly answer the query, pulling from an index, a feed, or a live web search. Stage two is ranking: the candidate pool is scored against the query and trimmed to the top handful. Stage three is synthesis: the engine writes a coherent answer, weaving passages from the top-ranked candidates and citing them as sources.
- Retrieval
- Narrow the universe. Pull 30-200 candidates from the index or feed. Speed-biased — runs in tens of milliseconds.
- Ranking
- Score the candidates. Use relevance, freshness, corroboration, and schema signals to pick the top 2-8. Runs in hundreds of milliseconds.
- Synthesis
- Write the answer. Extract quotable passages, attribute citations, handle disagreement between sources. The visible output. Runs in seconds.
Stage 1: Retrieval — how your product enters the candidate pool
Crawl access, feed completeness, and entity clarity decide whether an AI engine can even find your product. Most invisibility lives here.
Retrieval is where most Shopify stores lose. The engine has to be able to read your catalog and resolve your brand to a retrievable entity before your product is eligible for consideration. That sounds trivial — of course Google can find a Shopify store — but the ways retrieval fails are often silent and cumulative. A CDN-level block on GPTBot, a Merchant Center feed that dropped 40% of products on a sync error, an ambiguous brand name that resolves to a different company — each of these makes your product invisible at the retrieval stage even though classic Google can still find you.
What each engine reads for retrieval
- ChatGPT — ACP feed from Shopify (transactional), Bing editorial index (informational), direct GPTBot crawl (long-form).
- Perplexity — Shopify Catalog API, web crawl via PerplexityBot, editorial corroboration from Reddit and Trustpilot.
- Claude — web search via Anthropic's retrieval partner, plus llms.txt as a prioritization signal.
- Gemini / AI Mode — Google index, Google Merchant Center feed, query fan-out across 8-20 sub-queries.
- Copilot — Bing index, Microsoft Merchant Center feed, schema-validated product surfaces.
68%
of 'invisible' Shopify stores have a retrieval-stage problem, not a ranking problem
Surfient diagnostic audit of 1,207 Shopify stores, Q1 2026 — most failures are crawler access, feed completeness, or entity disambiguation.
The retrieval fix path is: unblock every relevant bot (GPTBot, ClaudeBot, PerplexityBot, Bingbot, Google-Extended) at robots.txt and at every CDN or WAF layer; complete your Merchant Center feeds with GTIN, MPN, and condition on every SKU; ship Organization schema with sameAs to disambiguate your brand. Clear those three and your product enters every major candidate pool.
Stage 2: Ranking — how candidates get scored and trimmed
Schema depth, answer-first content, freshness, and cross-source corroboration are the four signals every engine uses — weighted differently.
Once your product is in the candidate pool, the engine scores you against the pool and picks the top 2-8 to feed into synthesis. The four signal families that drive ranking are consistent across engines — what varies is the weight. ChatGPT weights feed completeness and Bing organic rank hardest; Claude weights cross-source corroboration hardest; Gemini weights schema depth and Merchant Center feed hardest; Perplexity weights editorial corroboration (Reddit, Trustpilot) hardest. Optimize for all four and you move up the ranking on every engine simultaneously.
- Schema depth
- Product + Offer + AggregateRating + FAQPage + BreadcrumbList + Organization. Complete beats thin every time.
- Answer-first content
- The first sentence of each paragraph answers the question the paragraph is about. Retrievers extract from passage openers disproportionately.
- Freshness per passage
- Ranked against the passage text, not the page lastmod. 'Spring 2024 collection' in a passage in 2026 demotes the whole page.
- Cross-source corroboration
- Reddit + Trustpilot + editorial + PDP saying the same thing outranks PDP alone. Weight varies per engine but every engine uses it.
Stage 3: Synthesis — how the answer actually gets written
The engine extracts passages, handles disagreement between sources, and decides which citation to foreground. Passage structure and FAQ blocks dominate what gets quoted.
Synthesis is the visible part — the part the buyer sees. The engine takes the top-ranked candidates and writes a coherent answer by pulling quotable passages, reconciling any disagreement between sources, and deciding which citation to foreground. Two things matter most for merchants at this stage: whether your passages are quotable at all, and whether your FAQ schema gives the engine pre-chunked answers it can lift directly.
What makes a passage quotable
- Self-contained — answers the question in 40-80 words without requiring context from earlier paragraphs.
- Entity-clear — names the product, brand, and category in the passage text rather than assuming the reader knows what 'it' refers to.
- Specific — cites a number, a dimension, or a named attribute. Engines prefer quotable facts over quotable adjectives.
- Non-promotional — 'Our handcrafted masterpiece marries old-world soul' is not quotable. 'The 42mm case houses a Swiss Ronda movement with 48-hour power reserve' is.
- Accurate — cross-checked against reviews and spec sheets. Synthesized answers are fact-checked against corroboration sources, and contradictions demote the page.
44.2%
of AI answer citations come from the first 30% of a page's text
Surfient citation-position study, 2,400 AI answers across five engines, Q1 2026. The lede is where the engine quotes from first.
Where the engines diverge on weightings
The mechanism is universal. The weightings are not. Here is the short map of which engine leans on which signal hardest.
The three-stage mechanism is consistent across every major answer engine in 2026. The signal weights are not. Understanding where each engine leans hardest lets you sequence optimization work by highest-expected-lift rather than alphabetically.
- ChatGPT
- Feed completeness > Bing organic rank > schema depth > corroboration. ACP and Bing pathways dominate.
- Perplexity
- Reddit corroboration > editorial coverage > schema depth > feed. Source plurality is king on Perplexity.
- Claude
- Cross-source corroboration > llms.txt curation > schema > content freshness. Conservative citation posture.
- Gemini / AI Mode
- Merchant Center feed > schema depth > FAQ fan-out coverage > passage freshness. Query fan-out rewards breadth.
- Copilot
- MMC feed strictness > Bing organic > third-party reviews > schema depth. The stricter sibling.
The takeaway: optimize the mechanism, not the engine
Most tactical advice is engine-specific. The mechanism is universal. Optimizing for the mechanism lifts every engine at once.
The practical consequence of the three-stage model is that engine-specific tactics are downstream of mechanism-level optimization. Get retrieval right once — crawler access, feed completeness, entity clarity — and every engine can find you. Get ranking right once — schema depth, answer-first content, freshness, corroboration — and you move up on every engine simultaneously. Get synthesis right once — quotable passages, FAQ blocks, non-promotional specs — and your citations land as first-class references rather than afterthought mentions.
“The engines will keep changing their weightings and shipping new surfaces. The three stages will not. Optimizing the mechanism is the closest thing to future-proofing that GEO offers.”