AI search
How AI answer engines choose which stores to mention
AI Overviews, AI Mode, ChatGPT, Gemini and Perplexity retrieve differently but reward the same things: crawlable pages, complete product data and clear answers.
By CartKernel · Published
A growing share of product research now happens as a conversation. Someone asks which robot vacuum handles pet hair on rugs, or which standing desk fits a 120-centimetre wall, and an assistant answers with two or three recommendations and a handful of sources. For an ecommerce brand the question is straightforward: how do we become one of the sources? The honest answer is that nobody outside the engines can guarantee it. But enough is documented, and enough is consistent across engines, to act on with confidence. This is a summary of what we rely on.
What is documented
Google AI Overviews and AI Mode run on Search. Google’s own guidance states that its AI features use the same core systems as classic Search, and that there are no special optimisations beyond standard SEO best practice: crawlable, indexable pages with helpful content. Google also lists things it calls myths, such as special AI files, content chunking and AI-specific rewriting. If a page cannot rank, it will not be used in an AI Overview either.
Google’s shopping features draw on Merchant Center. Product information in AI-driven shopping surfaces comes from the same product data that powers Shopping ads and free listings. A product that is missing, disapproved or thinly described in Merchant Center is not available to those experiences. This is the most concrete lever an ecommerce brand has, and the one most often ignored.
ChatGPT search uses its own crawler and index. OpenAI documents OAI-SearchBot as the crawler for its search features and describes how to allow it in robots.txt. Blocking it removes a site from ChatGPT search results. A separate crawler, GPTBot, is used for training, and it can be blocked independently.
Perplexity crawls and cites pages directly. PerplexityBot is documented, along with how to allow or block it. Perplexity answers are built from retrieved pages and show their sources prominently, which makes it the most transparent engine to study.
Anthropic’s crawler is documented too. Claude’s web access uses identified crawlers that respect robots.txt, per Anthropic’s support documentation.
What is inferred, consistently, from observation
Across engines, the pages that get cited share characteristics that follow from how retrieval-augmented systems work.
- They answer the sub-question directly. Engines break a query into related sub-queries, retrieve pages for each, and compose an answer from passages. A page that opens with a clear, self-contained answer to a specific question is easier to quote than one that buries the answer in paragraph six.
- They are structured for extraction. Descriptive headings, short paragraphs, comparison tables and lists give the engine clean units to lift. This is also good for shoppers, which is why it works.
- Their entities are unambiguous. Brand, product, organisation and offer data that agree across the page, the schema, the feed and third-party profiles. Contradictions, such as a different price in the feed and on the page or a brand name spelled three ways, are the fastest route to being left out.
- They are reachable as HTML. Content rendered only by client-side JavaScript, hidden behind tabs that are not in the DOM, or blocked by robots rules is not retrievable. Server-rendered pages with the full content in the source win by default.
- They are trusted. Engines lean on the same authority signals as search: links, brand mentions, reviews, and the absence of falsifiable errors. Buying guides that are genuinely expert, with the recommended products linked, tend to appear more than product pages alone.
What we do not know, and will not claim
We cannot see the ranking functions, and they change. We cannot promise that a store will be cited for a query. Monitoring shows that citations shift week to week even for stable pages. Anyone selling guaranteed AI visibility is selling something they cannot deliver. What can be delivered is the removal of every reason to be skipped, followed by measurement.
How we work on it
- Baseline. Run the store’s priority queries across AI Overviews, AI Mode, ChatGPT, Gemini and Perplexity. Record which stores and sources appear and inspect why: what those pages contain that yours do not.
- Access. Confirm robots.txt allows Googlebot, Google-Extended if you want to be in AI training-related features, OAI-SearchBot, PerplexityBot and ClaudeBot. Verify the important pages render their content server-side.
- Data. Fix Merchant Center eligibility, identifiers and attributes. Align Product, Offer, Organization and Brand schema with the feed and the page.
- Pages. Restructure priority category pages and buying guides answer-first, with comparison tables and specifications, and link them to the products they recommend.
- Measure. Re-run the baseline monthly, log the changes, and pick the next batch of pages from where the gap is largest.
The work overlaps almost entirely with strong ecommerce SEO and clean product data, which is why we deliver it as part of the same system rather than as a separate service with a separate invoice. Stores that do the fundamentals well show up more often. Stores that chase tricks do not.