Ask Alexa for Shopping, Amazon’s AI assistant, for its best recommendation on a product, and there’s a good chance it reaches for something your Amazon SEO would never have flagged as a competitor. A large-scale study of 1,963 non-branded queries and 12,810 recommendations, captured in May and June 2026 by AI-optimization vendor Autopilotbrand.com, found that 63.9% of the assistant’s picks fell outside the organic top 10 for the matched search term, and 40.9% never appeared on the visible search results page at all. Only 14.3% of picks were running a sponsored ad on that search page, and 83% of those already ranked organically anyway.
The study has since been corroborated in write-ups from other outlets that reviewed the same underlying data, and Autopilotbrand’s own site describes it as the first public study of how Amazon’s AI actually picks products — useful context for treating the numbers as a real, if early, signal rather than a one-off anomaly.
What the Study Actually Measured
Researchers posed a best-of question to the assistant — “what is the best queen mattress?” — and compared its answer against the standard category search, “queen mattress.” That framing matters: it isolates what happens when a shopper asks the assistant to recommend rather than to list, and the gap between the two shows the assistant reaching into a different, often deeper, part of the catalog than the page most sellers actively optimize for.
Neither of the Two Usual Routes to Visibility Appears to Matter Yet
Rank and ads are the two established paths to visibility in Amazon search: rank earned slowly through the sales velocity that ad budgets are spent to manufacture in the first place, and ads bought outright at auction. On this evidence, neither shapes what the AI assistant recommends. Rank is the more surprising omission of the two — it’s the metric most seller teams build entire dashboards around, and the one thing the assistant most conspicuously doesn’t follow.
A Counterpoint Worth Knowing
Not every read of the data agrees on how untouched by advertising the AI shelf really is. Separate analysis from Similarweb has characterized the assistant’s carousel as a “pay-to-play shelf,” where Sponsored Products placements and a shopper’s own purchase history shape what gets shown — a meaningfully different emphasis than the Autopilotbrand study’s finding that sponsored placement barely correlates with recommendations. Both can be partially true at once: an assistant that ignores organic rank while still weighting some paid signals and personalization is a different, messier picture than either “ads decide everything” or “ads decide nothing.” Treat any single study here as an early read on a system that’s still changing quickly, not a settled ranking algorithm.
How the Assistant Actually Decides, as Far as Anyone Can Tell
Independent breakdowns of the underlying system describe a multi-stage process that doesn’t map onto traditional keyword optimization. A product first needs baseline organic eligibility — Amazon’s A9/A10 search algorithm still governs whether a listing enters the pool of candidates at all, based on keyword relevance and performance signals. From there, Amazon’s COSMO model maps a shopper’s query to structured backend attributes — item type, intended use, material, compatibility — rather than to a listing’s consumer-facing marketing copy. A listing with polished bullet copy but incomplete backend attribute fields is reportedly close to invisible to the assistant at this stage, regardless of how well it ranks in ordinary search. From there the assistant reads full listing content, A+ Content, reviews, and Q&A, and its image models evaluate product photos directly rather than relying on alt text or captions.
Why This Won’t Stay a Blank Slate
Amazon has already begun monetizing this layer even as this particular study found no correlation yet: Sponsored Products and Brand Prompts inside Alexa for Shopping moved from beta to general availability in March 2026 and are now billable under standard cost-per-click terms. Amazon spent $68.6 billion on advertising revenue in 2025, and ad load has a documented history of eventually finding every high-traffic surface on the platform — exactly the pattern regular search followed before the current level of sponsored crowding set in. The AI shelf’s economics are not written yet, but there’s no reason to assume the current gap between influence-by-rank and influence-by-AI-recommendation stays this wide indefinitely.
What to Actually Do About It Now
Because the assistant reads structured data and specificity rather than keyword density, the practical response looks different from ordinary Amazon SEO work:
- Fill in every backend attribute field completely — item type, intended use, material, dimensions, compatibility — since an empty field is a question the assistant can’t answer about a product, independent of how well the listing ranks.
- Write bullet points and A+ Content as direct, verifiable claims about who the product is for and what specific problem it solves, rather than keyword-dense marketing language.
- Build a review and Q&A base that’s specific to real use cases, since the assistant appears to cross-reference listing claims against review content before recommending a product.
- Test a product’s own category with best-of style prompts directly in the assistant, the same way the study did, to see what currently gets recommended in that space and how a listing compares.
For a fuller walkthrough of how to structure a listing specifically for this kind of conversational, best-of query rather than a typed keyword search, optimize your listing for Rufus-style AI shopping questions covers the mechanics in more depth.
The Bottom Line
For now, the AI shelf is a rare surface on Amazon where rank incumbency is a weaker moat, and a product that has never cracked the search page’s top results can still appear where a category leader doesn’t. Search looked exactly like that once too, before ad load found it. The honest caveat is that this is one large study from a single US account, captured early in the life of a system Amazon is actively still building out and starting to monetize. Sellers who spend the next few months learning how this layer actually selects products will be the ones who notice, and can react to, the day its economics change.
