Guide · AI commerce data
AI commerce product feed data guide
What AI shopping surfaces actually need from product data — indexable pages, feed data that matches the page, and separate crawler decisions for OAI-SearchBot, GPTBot, PerplexityBot and Perplexity-User — plus an honest list of what no tool or vendor can promise.
64 versioned rules across 3 rulesets · ruleset 1.0.0 · rule data verified 11 November 2026 · no account needed for a local run · no AI-visibility score exists
Applies to
Public product pages and merchant feed data
Extra AI-only files
None required by Google’s published guidance
Crawler policies
Documented per crawler, never inferred
Last verified
11 November 2026
1. Start with the boring truth: there is no AI shortcut
Google’s current guidance on AI features in Search says the ordinary requirements still apply: a page must be crawlable, indexable, and eligible to appear with a snippet before it can be used as a supporting link. There is no AI-specific file, no special schema, and no trick that makes a page eligible. Anything that promises an “AI optimisation” layer on top of an unindexable page is selling a step that does not exist.
For a merchant, that means the work is unglamorous: a product page that returns HTTP 200 to a crawler, states the price and availability the feed claims, and carries machine-readable data that agrees with both.
2. The three layers a merchant actually controls
Feed readiness
The product data you submit: identifiers, price with an explicit currency, availability, image and URL, modelled for the destination schema you are submitting to. Google and OpenAI have different schemas; never apply one platform’s enum to the other’s field.
Product page evidence
What the public page says when read as static HTML: status, redirects, structured product data, and whether the price, currency and availability contradict the feed. A page that only renders with JavaScript is a limitation to report, not a verdict.
Crawler access
Whether the relevant search crawler can reach the page at all, according to your public robots.txt. Advisory only: a robots allow rule is not an approval, and a CDN or WAF can still block a crawler that robots.txt permits.
These layers are independent. A perfect feed does not make a blocked page discoverable, and a crawlable page does not fix a feed that contradicts it. Treating them as one thing is how merchants end up paying for feed tooling that cannot help and page tooling that cannot either.
3. Crawler policies, kept separate on purpose
The most common technical mistake is collapsing crawler identities. OpenAI’s OAI-SearchBot exists for search discovery; GPTBot is a separate, potentially training-related crawl with an independent policy. Perplexity documents PerplexityBot for search and Perplexity-User for user-triggered retrieval, and the two do not behave the same way. Deciding to allow one is not a decision about the other.
Googlebot
Google · Search crawling
Allowed on public indexable pages. No AI-specific access is needed for AI Overviews / AI Mode.
Googlebot-Image
Google · Image crawling
Allowed on public marketing images; product images are served from merchant hosts, not ours.
OAI-SearchBot
OpenAI · ChatGPT search discovery
A separate control from GPTBot. We document the choice explicitly rather than inferring one from the other.
GPTBot
OpenAI · Potentially model-training related crawling
An independent owner decision. We do not infer this policy from the OAI-SearchBot decision.
ChatGPT-User
OpenAI · User-triggered fetches
Not a search crawler. It is not the right token for expressing a search-visibility policy.
PerplexityBot
Perplexity · Perplexity search discovery
Documented separately, with its own choice.
Perplexity-User
Perplexity · User-triggered retrieval
Behaves differently by design; a robots rule for PerplexityBot says nothing about it.
bingbot
Microsoft · Bing search
Allowed on public indexable pages; Bing Webmaster Tools reports citation aggregates for supported AI experiences.
This site’s own policy is published in its robots.txt, and it does not invent a training-crawler decision on the owner’s behalf. User-agent names and IP ranges change; when an allow list is needed, verify against the operator’s current published IP-range endpoint rather than a hard-coded list.
4. What machine-readable data has to be true
- Structured product data must describe the visible page. A price in JSON-LD that differs from the rendered price is worse than no markup, because it is a contradiction a machine will read confidently.
- Do not invent ratings, reviews, availability or shipping promises to become eligible for a rich result. There is no display guarantee, and a fabricated aggregate rating is a policy violation.
- Keep one canonical URL per product and make it agree with the sitemap; conflicting canonicals waste the crawl budget you need for the pages that matter.
- Serve the substantive content in the initial HTML. A product page whose price only appears after a client-side fetch is a page a crawler may read as incomplete.
5. What the feed has to be true about
- The identifier is stable and unique per sellable variant; a parent handle is not an item ID.
- The price carries an explicit ISO 4217 currency, and a sale price is a real lower price in the same currency.
- Availability uses the destination’s own enum. Google and OpenAI spell and allow different sets, and the OpenAI Stable schema has its own five values.
- The image URL serves image bytes publicly, at or above the applicable minimum, in an accepted format.
- For OpenAI, the nine required basic fields are populated: item_id, title, description, url, brand, seller_name, image_url, availability, price. Optional enrichment such as seller_url is never a blocker.
6. What nobody can promise you
- That a product will appear in an AI answer. Citations vary by query, market, session, time and model version; a manual check is an observation, not a ranking.
- That a feed will be ingested or that an account will be approved. Both are platform-side decisions with onboarding gates.
- That a robots.txt allow rule means a crawler will reach you. CDN and WAF configuration sit in front of it, and a spoofed User-Agent means an allow rule is not an identity check.
- That a citation count in Bing Webmaster Tools AI Performance is traffic, rank or authority. It is an aggregate citation metric for supported experiences.
- That ChatGPT Business Hub accepts the same schema as the Stable file-upload path. They are different surfaces; do not assume one contract applies to the other.
7. A measurement plan that survives contact with reality
- Verify the property in Google Search Console and in Bing Webmaster Tools, and submit the sitemap. Search Console AI-feature exposure is reported inside the ordinary web performance report, not as a separate citation score.
- Instrument first-party events you control: import started, import parsed, mapping confirmed, validation started, validation completed, remote scan opted in, report exported, revalidation completed. Never include product values, identifiers or URLs in analytics events.
- Record referrals from known AI platforms where the referrer is present, and treat missing referrers as normal rather than as zero interest.
- Keep a dated manual observation sheet: query, market, platform, cited URL, context. Label it as a sample. It is not a ranking measurement and should not be presented as one.
- Re-check robots and WAF behaviour after any CDN rule change, because that is the most common way a previously crawlable page becomes invisible.
Ruleset version 1.0.0; last verified 11 November 2026.
Frequently asked questions
Do I need a special file or schema to appear in AI search results?
No. Google’s published guidance on AI features in Search says the ordinary requirements still apply: a page must be crawlable, indexable, and eligible to appear with a snippet before it can be used as a supporting link. There is no AI-specific file, no special schema, and no trick that replaces an indexable page.
Does allowing OAI-SearchBot also allow GPTBot?
No. OAI-SearchBot exists for ChatGPT search discovery and GPTBot is a separate, potentially training-related crawl with an independent policy — deciding to allow one is not a decision about the other. Perplexity documents PerplexityBot for search and Perplexity-User for user-triggered retrieval, and those two do not behave the same way either.
Do I need a perfect feed to be discoverable?
Neither layer substitutes for the other. A perfect feed does not make a blocked page discoverable, and a crawlable page does not fix a feed that contradicts it. The three layers — feed readiness, product page evidence and crawler access — are independent, and treating them as one thing is how merchants end up paying for tooling that cannot help.
What must the machine-readable product data agree with?
The visible page. A price in structured data that differs from the rendered price is worse than no markup, because it is a contradiction a machine will read confidently. Ratings, reviews, availability or shipping promises must never be invented to become eligible for a rich result, and each product keeps one canonical URL that agrees with the sitemap.
What do the nine required OpenAI fields have to say?
item_id, title, description, url, brand, seller_name, image_url, availability and price must all be populated, with an explicit ISO 4217 currency on the price and the destination’s own availability enum — Google and OpenAI allow different sets, so one platform’s enum must never be applied to the other’s field. Optional enrichment such as seller_url is never a blocker.
Can any tool guarantee an AI answer will cite my product?
No. Citations vary by query, market, session, time and model version, so a manual check is an observation and not a ranking. Feed ingestion and account approval are platform-side decisions with onboarding gates, a robots.txt allow rule is not an identity check, and a citation count in Bing Webmaster Tools AI Performance is an aggregate citation metric rather than traffic, rank or authority.