Ask most Amazon sellers how the platform works, and you get the same answer: it is a two-shelf game. There is an organic rank, which you earn through sales velocity, reviews and a clean listing. And there is paid, which you buy through Sponsored Products and the rest of the ad stack. Every brand has a strategy for both. Both are measured, optimised and argued over in weekly meetings.
There is now a third shelf, and almost nobody is managing it. When a shopper asks Amazon's AI assistant what to buy, the answer it returns is not a search results page. It is a short, curated list built under different rules. Amazon says the Rufus engine behind its shopping assistant helped more than 300 million customers in 2025 (Amazon), and the company is steadily nudging shoppers from the search box toward that assistant. If your bestsellers dominate organic and paid but never surface in the assistant, you are invisible on the shelf that is growing fastest.
That was the argument Christian Umbach, co-founder and CEO of Autopilot, made on a recent episode of MerchantSpring's Marketplace Masters, where he shared results from a study his team ran across roughly 70,000 Amazon products. The findings are specific enough to act on, so it is worth walking through what they found and what it changes.
How you study an AI assistant at scale
The hard part of studying a shopping assistant is that you cannot scrape it. Amazon detects automated collection and either blocks it or quietly distorts what it returns, which poisons the data. So Autopilot did it the slow way. The team designed around 2,000 non-branded queries spread across categories, roughly 200 to 250 each in apparel, supplements, home and kitchen, beauty and a few others, then ran them by hand while logged into a US account in June 2026.
For each query, they recorded two things side by side: what the traditional search results page showed, split into organic and sponsored, and what the assistant recommended when asked the same question in plain language, such as “what is the best running shoe?” They also kept pushing the assistant past its first pick, asking for more options up to ten. In total, the run captured about 13,500 assistant recommendations against roughly 110,000 products from search results pages, or close to 70,000 unique ASINs once duplicates were removed. This is a first-party observation of what a human shopper actually sees, not an inference from Amazon's API.
The assistant is not just a smarter search box
The central question was simple: Does the assistant recommend the same products that win organic and paid, or does it follow its own logic? The data points clearly to its own logic.
Across the study, 64% of the assistant's top picks were not in the organic top ten. The single number-one recommendation did overlap with the organic top ten about 55% of the time, so a strong organic position still helps. But the further down the assistant's list you go, the weaker that link becomes. By the eighth recommendation, overlap with the organic top ten fell to roughly 18%. Marketplace Pulse framed the same pattern from the other side: being in the organic top ten gives a product only about a 6% chance of appearing in the assistant's results (Marketplace Pulse).
The finding Umbach called the most striking: about 40% of the products the assistant recommended were nowhere on the traditional results page at all, neither organic nor sponsored. Given how heavily sales concentrate on the first screen of Amazon search, that is a real opening. A brand that has not yet cracked page-one rank can still land in the assistant's answer, which is incremental demand it could not otherwise reach.
You cannot buy your way in yet
The obvious next question from any marketer is whether ad spend opens the door. Right now, it does not. Sponsored listings made up 20.7% of results on ordinary search pages but only 14% of the assistant's picks, so paid placement is underrepresented in the assistant relative to search. More tellingly, only 2.4% of the assistant's recommendations were advertised products that were not already earning their place elsewhere on the page. Buying visibility directly inside the assistant is not currently a viable path.
Advertising still matters indirectly. Spend drives visibility, visibility drives sales, and sales velocity is one of the signals that lifts a product into both page-one rank and the assistant's consideration set over time. Worth watching: Amazon began rolling out sponsored product prompts in March 2026, and Umbach expects clearer ad formats for AI placements to emerge. The current picture is a snapshot, not a fixed rule.
What actually earns a spot
The study cannot reverse-engineer Amazon's exact prompt, and Umbach was careful about where the data stops being statistically firm. But the patterns across recommended products were consistent enough to guide action. Every product the assistant suggested had a rating of 4.0 or above. Everyone was in stock. Most carried a velocity signal, the “bought 50 times last month” style callout, and most had listings with real substance in the title and bullets rather than filler.
Read together, those are the same fundamentals that make a happy customer: good ratings, availability, proof of demand and a listing that describes the product accurately enough for the assistant to match it to a shopper's intent. Two practical moves follow. Early in a product's life, when reviews are thin, Amazon Vine is a sensible way to build review volume and quality at scale (Amazon Seller Central). And at the catalogue level, the job is to feed the assistant clean, complete attribute data so it can place your product in the right context, for example, recognising that the shopper is a college student kitting out a first dorm room, and matching accordingly.
Category matters, and apparel is the wildcard
The overlap between the assistant and organic search was not uniform. Averaged across categories, about 36% of assistant results came from the organic top ten. In supplements, it ran higher, around 53%, meaning rank protected you reasonably well. In apparel, it was the lowest, so the assistant's picks diverged most from search.
Umbach's read is that apparel carries the highest degree of personalisation. Amazon leans on custom preferences there, so if it knows a shopper favours blue shirts, that shapes what surfaces in a way traditional search does not always mirror. For apparel brands, this cuts both ways: rank buys you less, but there is more room to win by presenting the full range and the specific use cases each product serves, including micro-seasonal spikes like everything-orange demand around Halloween.
Apparel is also an early winner from Amazon's updated title policy and the introduction of item highlights. Titles are capped shorter now, but sellers get a dedicated highlights field of up to 125 characters to call out attributes like waterproofing or compliance claims. That is close to an extra marketing slot, and most listings waste it. Umbach noted that many apparel listings still run three to five bullets of generic buzzwords, “made in the US, cotton,” and little else. Lifting catalogue quality across thousands of variant listings is unglamorous work, but Autopilot sees it move conversion rates, and conversion feeds the velocity signal the assistant rewards.
The shelf extends beyond Amazon
The same dynamic is playing out in general AI assistants, and the sourcing differs by platform. A study by Peak Ace, published in Search Engine Land, found that about 45% of the product-carousel items ChatGPT recommended matched the Google Shopping top 40 exactly (Search Engine Land). So where Amazon's assistant leans on Amazon's own catalogue, ChatGPT currently leans heavily on Google's shopping ecosystem.
This creates a specific problem for brands whose strength is on Amazon. Amazon holds rich, structured product data because it forces sellers to provide it, but it walls that data off from the open web. ChatGPT, Claude and the rest do not get a low-cost feed of it. Umbach's team has started building agent-readable pages, closely tied to a brand's Shopify site, that expose product details, unique selling points and item highlights in a form bots can read, so that when Meta, Anthropic or others crawl, the information is there. Early results are showing up as citations in AI-visibility tools and as AI-assistant traffic in Google Analytics, with add-to-cart events attributed to it. It is early and unsettled, but the underlying point holds: the disciplined product data Amazon trained sellers to maintain is becoming an asset across a much wider discovery surface.
The scale behind this is why it matters. OpenAI has said roughly 20% of ChatGPT queries already carry some shopping intent. Discovery is not moving to a single new place; it is fragmenting across assistants that each source differently. Content optimisation is not dying. It is splitting into two jobs: writing for the human reading the page, and structuring that same information so an AI can find, trust and cite it.
What to do about it
The takeaway is not to abandon rank and ads. They still drive the velocity and reviews that feed everything else. The takeaway is that they are no longer the whole game. Audit whether your bestsellers actually appear when you ask Amazon's assistant a plain-language buying question in your category. Treat the fundamentals the assistant rewards as non-optional: 4.0-plus ratings, reliable stock, genuine review volume, and listings that use every field, including item highlights, to describe the product precisely. Then extend the same product data outward so the assistants beyond Amazon can read it too.
The third shelf is small today, and it plays by rules that are still forming. But it is the shelf Amazon is actively steering shoppers toward, and the brands auditing it now will understand it before their competitors know it exists.
See the shelf you're missing
The third shelf runs on the same fundamentals you can already measure: ratings, availability, review velocity and listings that convert. The brands that win it are the ones watching those signals across every account, not guessing. MerchantSpring gives sellers, agencies and enterprise brands one view of marketplace performance, so you can spot the listings dragging your velocity down before they cost you a recommendation. See it on your own accounts with a walkthrough of your data.
