Rufus is gone. In May, Amazon replaced it with Alexa for Shopping, and shoppers have stopped typing keywords. They now ask for outcomes: "best leakproof mug for commuting," "quietest humidifier for a nursery."

Which means you have no idea whether Alexa is putting your product in that answer, or your competitor's.

Lumian built a free Alexa Visibility Tool that settles it. Drop in your ASIN and it runs real Alexa queries against your listing, then shows exactly where you surface, where you're invisible, and what to fix to close the gap.

Takes under a minute, and you'll know where you stand before your competitors do.

Run your ASIN free

Amazon has started showing shoppers the conclusions it draws about them. Most people found this mildly horrifying. For those of us selling on Amazon, the useful question is what those conclusions are built from. (Spoiler: largely our products.)

Since May, Amazon customers have had an account page called About You, listing what Amazon has concluded about them from purchase history, searches, saved Lists, reviews and conversations with Alexa for Shopping, the conversational assistant launched the same month (the listening assistant and the dossier, delivered in the same breath). Shoppers can edit or delete entries, or simply ask the assistant what it knows.

Nobody much noticed until October, when shoppers started posting their profiles on social media. Most entries were mundane: keeps aquariums, drinks coffee, plays video games. Others were less so. One shopper learned Amazon had concluded she has "flat buttocks", apparently on the strength of a purchase of shaping leggings. Others were told they can't keep plants alive, treat their cats as children, or have no friends. Some were simply wrong, including one that recorded a partner under the wrong name (a mistake that usually ends relationships rather than recommendation sessions).

The genuinely instructive material, though, came from one customer who requested a full export of what the assistant had compiled about him over nine years. It held more than a hundred plain-language statements about his life, the large majority accurate, each with a topic, a confidence score, a timestamp and a short explanation of how direct the evidence was, how often it had occurred and how large the "inference leap" had been. (Amazon marking its own homework, but with decimals.)

Amazon's classic recommendation logic was correlation between products: people who bought this also bought that. The customer was implied, but the system's working material was the catalogue.

Inference works differently. The system turns transactions into statements about the person, and those statements become the inputs for what happens next. One entry in the export shows the mechanism neatly: the user had a pet called Conrad needing regular dental care, "suggesting conrad is a dog", with a confidence score of 0.571. The evidence was a reminder to brush Conrad's teeth and, a few days later, an order of pet wipes. Neither means much alone; together, they produced a durable conclusion. Months later, when the user asked for paw wipes, the assistant referred to the dog by name. (Excellent customer service or the opening scene of a thriller, depending on your disposition.)

This matters because conversational requests are short. Ask "what should I get for cleaning up after walks?" and most of the meaning is left unsaid, so the assistant fills the gap with what it believes about you. Two shoppers saying the same words can expect different answers, because they're no longer really asking the same question.

If that's how the system reasons, every purchase on Amazon now does two jobs. The first is obvious: it's a sale. Lovely. The second is quieter and lasts longer: it's evidence in a statement Amazon writes about the buyer, shaping what they see across categories, potentially for years.

And products differ enormously in how much evidence they provide. A pack of batteries says almost nothing about the person who bought it (beyond, presumably, ownership of a remote). Trail running shoes, a puppy training kit, a sourdough starter or a beginner's aquarium say a great deal: each implies an identity, an ongoing activity and a predictable set of future needs.

The nine-year export was built largely from that second kind. The user was classified as a runner because he'd bought running shoes, judged to cook with premium ingredients because of orders from an upmarket supermarket, and tagged as preferring plant-based protein because of what he'd ordered. Each time, one or two products wrote a trait into the profile.

Now the strategic bit, which comes with an "if" attached. If inferred traits persist, and if they shape later recommendations, then the brand whose product wrote the trait may be well placed to benefit afterwards. For fifteen years, Amazon strategy has been about winning the purchase in front of you: the keyword, the ranking, the click. In a profile-driven system, the more valuable moment may be the purchase that first established this customer as a runner, a new dog owner or a home baker, because every later recommendation in that territory is made with that trait in view. First impressions, it turns out, were never just a dating problem.

That shift reframes some very familiar decisions.

Entry-level products and starter kits have usually been treated as low-margin acquisition tools, justified by repeat purchase. They may increasingly matter as the way a brand gets associated with a trait in the first place. The question moves from "what does this product earn?" to "what does this product say about whoever bought it?"

Bundles look different too. The export shows the system weighting directness and frequency when deciding how confident to be. A starter bundle of several clearly related items sends a denser, more consistent signal in one transaction than any single item can. One order, one coherent story about the buyer.

Range architecture carries more weight. A brand selling the shoes, the socks, the energy gels and the recovery tools is positioned to benefit from the trait its first sale helped create. A brand selling one item leaves the rest of that need to whoever the system recommends next. (Creating the runner and then letting someone else sell them the socks is, let's say, a generous business model.)

Then there's evidence quality: an inference is only as good as the system's understanding of what a product is and who it's for.

For years, a common listing tactic has been breadth: copy written to catch every possible search term, describing a product as suitable for men, women, children, the home, the office and travel. (The perfect product for everyone, everywhere, otherwise known as the perfect product for no one in particular.) In a keyword world, that breadth widened the net. As evidence, it's weak, because it tells the system almost nothing specific about the person who bought it.

A listing that states clearly what the product is for, who typically uses it and in what situation gives the system a cleaner basis for its conclusion, which neatly inverts a habit a decade of keyword optimisation trained into the industry. (Sorry.) The test becomes simple: if the system read only this page, would it know what kind of person just bought from you?

The export also shows what happens when evidence is thin. The same profile concluded the user used a self-driving car service, which he didn't, and that he was interested in magnetism because he'd asked about it once. (One question, and curiosity becomes a personality trait.) A wrong trait helps nobody, least of all the brand that triggered it.

The limits are significant. Brands can't see profiles, can't see which traits their products generate, and can't target them directly. Ad audience segments exist, but the trait layer itself isn't exposed. (Amazon will cheerfully tell a shopper it thinks she has flat buttocks. It is noticeably less chatty with the people supplying the leggings.) What we can control is how clearly a product signals its use and user, how the first purchase is composed, and whether the range around it is coherent enough to serve the needs the system will infer.

The second consequence is about measurement. Organic rank, keyword tracking and share of voice all assume a search term produces one results page that every shopper broadly shares. Under profile-conditioned recommendations, a shopper filed as a marathon runner and one filed as an occasional gym-goer may get quite different answers to the same request. Rank stops being a position and becomes a spread of outcomes.

Many third-party trackers check rankings from neutral sessions with no history, reporting the experience of a shopper with no profile: precisely the kind the system wants fewer of. Not worthless, but a shrinking share of what customers actually see. Even Amazon's own keyword reporting is an aggregate, so a product can hold a steady impression share while winning with one group and quietly vanishing for another. (Averages: the statistical equivalent of "I'm fine.")

A more useful question might be: for which kinds of customer is this product the natural recommendation, and is that group growing? We can approximate it by watching which entry products lead to second purchases, how new-to-brand customers behave after their first order, and which items in a range get bought together over time.

So why would Amazon show its working at all, when a silent engine is far easier to run? Three plausible reasons, not mutually exclusive.

Data quality. Every shopper who deletes a wrong entry or confirms a right one is labelling Amazon's data for free, and October's public profile-picking may have produced more of it in weeks than months of quiet availability. (Unpaid data annotation, rebranded as self-discovery.)

Trust. A conversational assistant says its inferences out loud, and when it named a dog it had never been told about, the user felt unease rather than delight. An editable profile answers "how did you know that?", a question that matters when Amazon says it has sold more than 600 million Alexa devices.

Regulation. Amazon Store has been designated a very large online platform under the EU's Digital Services Act since 2023, and the General Court dismissed Amazon's challenge in November 2025. Such platforms must explain their recommender systems plainly and offer a non-profiling option.

Whatever the motive, the customer Amazon optimises for is now partly written by the customer. Delete a trait, and whatever advantage it gave the brands it favoured goes with it. The leggings did their commercial job perfectly well, but "flat buttocks" is precisely the kind of trait a customer removes (possibly with some force). Accepted inferences stick. Intrusive ones may not last the week.

Plenty remains open, from whether trait-level data will ever reach sellers to how many shoppers will edit their profiles. If enough Europeans opt out of profiling, the EU marketplace could start behaving like the keyword-led store we once knew, while the US moves further away from it.

What we do know is that every product we sell is now quietly testifying about the person who bought it. And nobody yet knows what a useful dashboard looks like when every customer is, to some degree, shopping in a different store.

Jo

P.S. If Amazon is 57.1% sure about a dog it has never met, I'm choosing not to ask what it has concluded from my own order history.

🗣️ Murf AI - AI voiceover platform with preset voices, dubbing and a script assistant.

🌐 Replit - Cloud IDE with an autonomous agent that builds, hosts and deploys apps.

🎨 Midjourney - Benchmark AI image generator for artistic, editorial imagery.

📂 Elicit - Automate literature reviews and extract data from research papers.

😍It was great, all I needed

🙂Good, not great

🙄It sucked