How AI assistants pick products (observed behavior)
No one outside these companies has the ranking algorithm. But you can watch what assistants actually do — and the pattern is consistent enough to design for.
Start with the disclaimer
This is field observation, not a leaked algorithm. We watch how assistants answer real buying prompts across catalogs we operate, and we log what gets cited, what gets bought, and what gets ignored. Everything below is a pattern we can reproduce — not a secret we were handed.
Assistants are already a primary recommender
The behavior is worth designing for because the traffic is real. AI is now one of the most popular sources of product recommendations — trailing only physical stores, and ahead of social media, friends and family, and brands’ own sites, per Accenture.¹ Between May 2024 and June 2025, 2.1% of all ChatGPT queries were shopping-related, often phrased like “recommend a laptop under $1,000.”²
What we see the models reward
First, structured, machine-readable product data — assistants quote specs, prices and availability they can parse cleanly, and skip catalogs they can’t. Second, corroboration across independent sources: a claim that appears on the retailer, in reviews and in editorial coverage survives; a claim that lives only on your own product page does not. Third, recency and specificity — exact model numbers, current prices, concrete use-cases.
And critically, the assistant is rarely the last step. In AI shopping sessions, nearly 80% of people still visit a retailer or marketplace to validate before buying.³ The model narrows the set; the retailer closes. If your product isn’t cleanly present in both places, you lose in the gap.
The model narrows the consideration set. The retailer closes the sale. You have to win both.
How we design for it
We build AI-ready catalogs, seed corroborating signals off-site, and make sure the retailer page an assistant hands off to confirms exactly what the assistant promised. That’s the GigaCommerce stack — Brand Agents, Copilot Checkout and structured catalogs — pointed at the moment a machine is deciding what to recommend.