How LLMs read and interpret product pages

When ChatGPT, Perplexity, Gemini, or Claude answers "what's a good linen shirt brand that ships to Denmark," something different from classic search happens. There's no page of ten blue links to rank on. The system retrieves candidate pages, reads them, extracts facts, and composes an answer that mentions some brands and not others. Understanding that reading process is the foundation of everything in this phase.

Retrieval first: the old rules still gate the new game

AI assistants with live web access typically run searches under the hood and read what comes back - which means classic search visibility is the qualifying round. Pages that don't rank don't get retrieved; pages that don't get retrieved can't be read; pages that can't be read don't get cited. Everything from Phases 4 and 5 - indexable, crawlable, ranking pages - isn't obsoleted by AI search. It's the entry ticket. (Some systems also lean on training data for well-known entities, which is why brand presence across the web matters too - a later lesson.)

Then reading: extraction, not ranking

Here's where the game changes. A retrieved page isn't scored against others - it's parsed for usable facts. Can the system determine what this product is, what it's made of, what it costs, whether it ships to the user's country, what makes it distinct? Pages that state facts clearly, in extractable structures, become sources. Pages where the facts are implied, scattered, or locked in mood copy get skimmed and skipped - not penalized, just unusable.

This is why the same properties keep winning: explicit attributes rather than vibes ("100% European linen, relaxed fit, 165 g/m²" over "effortlessly breezy"), structured data as the no-interpretation-needed fact layer (the schema from Phase 5, now with a second consumer), clean headings and scannable structure (models parse structure much like skimming humans), and self-contained statements - a fact that requires three other page sections as context extracts badly.

What "winning" even means here

In classic search you win a position. In AI search you win a mention - being the brand named in the answer, the product cited with a link, the source the assistant trusted. There are fewer mentions per answer than links per results page, which makes this more winner-take-most than classic SEO. And the deciding factor, more nakedly than ever, is whether your pages contain clear, complete, extractable facts about your products.

You can see where this is going: the catalog work from Phase 2 was never just about filters and feeds. It was, it turns out, preparation for being readable by the systems your customers increasingly ask first. The rest of this phase builds on that foundation.