How AI shopping visibility actually works

01How AI answers are assembled

Every AI answer is built in three stages: retrieval (a search index finds candidate pages), grounding (the model reads them at answer time), and citation (a handful of sources get named, linked, or recommended). You can lose at any stage — blocked from retrieval, unreadable at grounding, or outcompeted at citation. And each engine runs its own crawler and its own index: being visible to one says nothing about the others.

02The crawlers that matter (and the ones that don't)

Never block these three — each is the documented gate for its surface: OAI-SearchBot (OpenAI's docs: opted-out sites “will not be shown in ChatGPT search answers”), PerplexityBot, and Googlebot (Search, AI Overviews, and the crawl side of the Shopping Graph behind Gemini).

These are training-only — blocking them costs zero visibility: GPTBot and Google-Extended (Google: “does not impact a site's inclusion in Google Search nor is it used as a ranking signal”). And ChatGPT-User / Perplexity-User fetch on behalf of live users and largely ignore robots.txt either way. Full write-ups: GPTBot vs. OAI-SearchBot and Google-Extended's actual scope.

robots.txt isn't the whole story: WAF/CDN “block AI bots” toggles silently override it. Ground truth is your server logs checked against each vendor's published IP-range JSON. Changes propagate slowly (OpenAI cites ~24 hours), and a 5xx on robots.txt can read as “disallow everything.”

03Structured data

Ship Product/Offer JSON-LD on every product page to Google's merchant-listing spec: required name, image, and an Offer with price + priceCurrency; strongly recommended availability, brand, sku, gtin/mpn, shippingDetails, hasMerchantReturnPolicy. Use exact schema.org enums (https://schema.org/InStock, not “available”).

Honesty check: OpenAI publishes no JSON-LD requirement — markup matters through Google's rules and Perplexity's crawl-fed index. And consistency is itself a signal: when markup, feed, and visible page disagree on price or availability, the data reads as stale and products get suppressed.

04Product feeds

ChatGPT Shopping: OpenAI documents two parser profiles — the native schema (twelve required fields, gtin optional) and a Google-compatible profile where gtin-or-mpn is required unless identifier_exists: no. The parser is selected per file, never per row; XML is not accepted. Direct feeds are gated to approved partners, but Shopify merchants are auto-integrated via Shopify Catalog. Details: the GTIN parser-mode rules.

Google/Gemini: the Shopping Graph is fed by Merchant Center feeds plus the open-web crawl. A wrong GTIN causes disapproval — validate GS1 checksums, never guess. Perplexity: no public feed spec; ingestion is the Merchant Program, the Shopify integration, or the crawl — so crawlability and markup carry the weight. Everywhere: refresh at least daily.

05Content that gets cited

Answer-shaped content wins: question-form headings, the direct answer in the first sentence, supporting detail after — engines quote passages, not pages. Original data (studies, benchmarks, methodology) is the strongest citation magnet: models prefer the primary source over the tenth summary of it. Specificity beats volume — exact numbers, named entities, dates, verbatim quotes give a model something quotable. Comparison and “best X for Y” pages map directly onto buyer-intent questions. Keep visible dates fresh, and publish an llms.txt — an emerging convention, not a documented ranking factor, but cheap and legible to the machines you're courting.

06Brand & entity signals

Models learn brand associations from the wider corpus — reviews, forums, listicles, press. Being mentioned where models look influences recommendations in ways your own site can't. Keep entity data consistent across your site, profiles, and directories, and maintain legitimate third-party review presence.

07Technical hygiene

Server-render what matters — most AI crawlers execute little or no JavaScript, so client-rendered product data may simply not exist to them. Clean canonicals, working sitemaps referenced from robots.txt, fast responses, no soft-404s, no walls in front of key pages. Retrieval layers are search indexes underneath; the old fundamentals still apply.

08Measurement

Track AI referral traffic (chatgpt.com and perplexity.ai referrers convert unusually well — the intent is pre-qualified). Check whether you're actually cited for buyer-intent queries on each surface, on a schedule — the ToS-safe way at scale is the vendors' official APIs (OpenAI's web-search tool, Perplexity Sonar, Gemini grounding), which expose the same retrieval layer with citations. Track who gets cited instead — the AI competitor set often differs from the SERP set. And re-verify against vendor docs on a cadence: specs change without announcement.

09Myths that waste budgets

1. “Block GPTBot to protect content” — fine, but it neither costs nor saves any shopping visibility; the gate is OAI-SearchBot.
2. “Blocking Google-Extended affects AI Overviews or Shopping” — Google says outright it doesn't.
3. “OpenAI requires schema markup” — no such requirement exists in any OpenAI document.
4. “GTIN is required in OpenAI's feed” — depends entirely on which parser profile your feed hits.
5. “One blanket AI-bots robots rule covers everything” — the bots have different jobs; blanket rules over-block or under-protect, always.

10Priority order

1. Unblock the three gating crawlers — robots.txt and WAF, verified in logs.
2. Fix product JSON-LD to merchant-listing spec; eliminate feed↔page drift.
3. Make the feed conformant: right parser profile, valid identifiers, daily refresh.
4. Reshape key pages into answer-shaped content; add FAQ/Article markup.
5. Publish original data worth citing.
6. Set up citation monitoring and AI-referral tracking; review monthly.
7. Build off-site mentions where models look.