Inside the shopping-AI extractor: why the attributes behind your photos, not the photos themselves, decide the citation

Discussions among ecommerce and search-optimization teams this week keep circling the same puzzle: why a store with professional product photography still loses the citation to a plainer rival. The running answer, echoed across community threads and agency write-ups, is that the "assistant" shopping for your product is an extractor over text and markup, not a viewer of pixels.

Practitioners describe the AI shopping loop in three steps. First the assistant finds candidate pages, then it extracts citable facts from the text and structured data it can read, and finally it cross-checks those claims. A product photographed beautifully but described thinly rarely survives step two, because a rendered image is not a retrievable attribute. Sizes, weights, materials, and ratings only help when they appear as searchable sentences or as fields an engine can verify.

The pattern holds across the format differences people call out between ChatGPT, Perplexity, and option-driven shopping agents: whatever the interface, the ground layer is the same machine-readable text underneath the page.

Across posts and audits circulating this week, the advice being repeated to image-first stores is consistent:

• State every photographed feature in text, ideally above the fold next to the image
• Keep dimension, material, and availability in a parseable spec block and in the product markup
• Treat product feeds and structured catalogs as a citation channel in their own right
• Verify with a scraper-eye tool what an assistant can actually read on a given URL

The common thread is that winning references depends less on how a page looks to a person and more on how cleanly its facts can be pulled and confirmed by a machine.

The litmus test is simple: open a product URL in an AI-style reader and ask whether each selling point you care about is visible as text. If a fact exists only inside an image, it is currently uncitable. A quick way to surface those gaps is the free scan at geovisora.com/audit, which reads a page the way an assistant does and points to the exact spots where your specs are stuck in pixels instead of copy. Photography still wins the click from a human; the text and schema underneath are what win the citation.