Amazon UK moves product discovery to WhatsApp
There is nothing for a seller to do about this one, and that is the point of writing it up. Amazon’s shopping assistant on WhatsApp now lets a UK shopper “reply to find deals, ask for recommendations, and discover products”, and then “complete your order on the Amazon app or website”. A conversation that used to happen inside Amazon’s own search box can now happen in a messaging app, and no part of it is visible to the seller whose product does or does not get mentioned.
No adoption figures have been published. No usage numbers, no share of orders, nothing. We are not going to estimate any, and any percentage you see attached to this story did not come from Amazon.
What the thing is
It is described as the Amazon Shopping Assistant on WhatsApp, reached in the UK on an Amazon-operated number, and it is explicitly a language model rather than a menu: “AI. It responds to what you ask in the conversation to help you discover deals and products.” The interaction pattern in the published examples is the ordinary one for an assistant of this kind. A shopper asks for something vague, the assistant asks a clarifying question, the shopper narrows, the assistant proposes products.
Two operating details are stated and both are worth holding onto. “Amazon will never ask you to provide any confidential information via WhatsApp”, which is a statement about fraud rather than about features, and a reasonable one to expect given how quickly messaging channels attract impersonation. And “Your conversation is not stored after your session ends”, which sounds like a privacy commitment and is also a design constraint: an assistant that forgets each session cannot build a picture of a shopper over time inside this channel.
The purchase itself does not close in WhatsApp. The order is completed on the Amazon app or website. So this is a discovery and consideration surface bolted onto the front of the existing funnel, not a new checkout.
From the publisher
Everything we know about running Amazon, in one course.
The commercial strategy, taught by operators who run our own brands.
What a seller does differently on Monday morning
Nothing. There is no setting, no opt-in, no placement to buy, no eligibility to check and no reporting to read. There is no seller-facing surface to this at all.
We would rather say that plainly than manufacture three bullet points. The value of the story is not a task, it is a fact about where selection is starting to be decided.
Tool comparison · FastMoss vs Kalodata
Assistants and recommendation layers expose nothing to sellers, so third-party data is the only window in. FastMoss and Kalodata do that job for TikTok Shop, reconstructing what is selling and for whom without waiting for the platform to publish it.
Comparison coming soon
FTC disclosure: our tool comparisons carry affiliate links. If you sign up through one we may earn a commission, at no extra cost to you. We rank on merit, never on commission, and the verdict is written by a human.
The part that should bother an operator
Every discovery channel a seller has ever worked in came with three things: a way to see demand, a way to influence placement, and a way to measure the result. Organic search gave impressions and rank. Sponsored Products gave an auction and a report. Even a shelf in a physical shop gave you a planogram and a sales figure.
A conversational assistant gives none of the three. You cannot see how often shoppers ask for the category you sell in. You cannot see whether your product was among the ones proposed, or which competitor was proposed instead, or how the assistant characterised yours. There is no bid, because there is no published auction. And because the conversation is not stored after the session, there is not even a record for anyone to surface later.
That is a real change in the seller’s position, and it is worth naming without dramatising it. This particular channel is one country, one messaging app, and an unknown volume of conversations. On its own it decides nothing. As a pattern it is the third or fourth instance of the same shape in about two years: a layer that reads the catalogue on the shopper’s behalf, chooses a short list, and exposes no controls to the people whose products are on it.
The rational response is not to chase the channel, because there is nothing to chase. It is to notice which inputs such a layer can actually read, since those are the only levers that exist. An assistant proposing products from a catalogue works from the structured fields: what the product is, which variant, the attributes, the review signal, the price, the availability. It does not work from a keyword tail. Sellers who have kept their titles unambiguous, their attributes complete and their variant structure clean are legible to this kind of system. Sellers whose listings were optimised for a 2019 keyword algorithm are not.
That is not a task created by this announcement. It is a task that already existed, which this announcement makes marginally more valuable.
Tool comparison · FastMoss vs Kalodata
If a channel gives you no reporting, you buy your reporting elsewhere. We compare FastMoss and Kalodata on exactly that: how much of a closed platform each one can actually see, and how reliable the reconstruction is.
Comparison coming soon
FTC disclosure: our tool comparisons carry affiliate links. If you sign up through one we may earn a commission, at no extra cost to you. We rank on merit, never on commission, and the verdict is written by a human.
What not to do about it
Stories with no available action attract bad advice, so it is worth listing the moves that look responsive and are not.
Do not go and set up a WhatsApp presence on the strength of this. Nothing about Amazon’s assistant creates a route from a shopper’s WhatsApp conversation to your own WhatsApp account, and a customer-service channel you cannot staff is worse than no channel. If you want a messaging channel, decide that on its own merits, not because Amazon launched one.
Do not buy a service that promises visibility in AI shopping assistants. There is no published placement, no auction and no eligibility criterion in this channel, which means there is nothing for such a service to influence. What can be influenced is the quality of your listing data, and you do not need a vendor to tell you that.
Do not read the absence of numbers as evidence of scale in either direction. A launch with no adoption figures is a launch with no adoption figures. It may be enormous and unmeasured, or small and quietly tested. Treating silence as a signal is how sellers end up reorganising around channels that never arrive.
And do not assume the conversation is being logged for later analysis you might one day benefit from. The stated position is that “Your conversation is not stored after your session ends”. Take that as a constraint on what any party, including Amazon, can reconstruct about a given shopper’s path through this channel.
The one thing this does change
There is a narrow, real consequence, and it is about attribution rather than tactics.
A shopper who discovers a product in a WhatsApp conversation and then completes the order on the Amazon app or website arrives at your listing already decided. From your side that looks like direct or organic traffic that converts unusually well, with no visible discovery event in front of it. It will flatter whatever channel gets credit for the session.
That is not a crisis, and the volumes today are unknown and probably small. But it belongs on the list of reasons your marketplace attribution is approximate: the same shape of problem already exists for products discovered on social video, in a search summary, or through a voice query. Each of these adds a discovery event you cannot see to a purchase event you can.
The defensible response is to stop treating last-click marketplace attribution as a measure of what created demand, and to keep at least one blunt aggregate measure, total sales against total marketing effort over a period, that does not depend on being able to see the path. That advice long predates this announcement. This is one more reason it holds.
How much weight to put on it
Not much, this week. We are marking this honestly: it is a feature launch in one market with no numbers attached, no seller controls and no measurable consequence. On its own it does not justify changing anything in a European seller’s operation, and a European seller cannot even see it, because the channel is UK-only on the evidence available.
What makes it worth ten minutes of attention is the direction rather than the size. Amazon is placing its shopping intelligence outside its own app, in a channel people already have open, and is prepared to lose the persistent session to do it. That is a company treating the assistant, not the storefront, as the product. If that continues, the medium-term question for sellers is not how to rank on Amazon’s search results page. It is how to be the product an assistant picks when the shopper never sees a results page at all.
We do not know the answer to that question, and neither does anyone claiming to sell the answer. What is observable so far is that these systems reward listings that are clear about what the product is, and are indifferent to the tactics that worked when the reader was a keyword matcher. That is a thin conclusion, and it is the one the evidence supports.
For this week: read it, note it, change nothing.
The three-line version
- If you run the channel: nothing to do. Keep your listing data clean and unambiguous, which was already true before this launched.
- If you have just taken on discovery: the gap to name is that a growing share of demand now arrives already decided, with no report behind it.
- If you sign the budget: decline anything sold to you as visibility inside AI shopping assistants. There is no placement to buy.