What Amazon said
On its third-quarter earnings call on Thursday, Amazon CEO Andy Jassy said more than 400 million customers had used Alexa for Shopping in the past 12 months, up from 350 million at the end of July. He said those customers spend 50% more than customers who don’t use it. Modern Retail reported it on 29 October 2026.
Two more facts from the same report. Alexa for Shopping started life as Rufus, released in February 2024, and got its new name in May 2026. And Amazon has blocked ChatGPT and Meta’s Muse shopping agent from its site, while Jassy says he wants as many customers as possible on Amazon’s own assistant.
What the numbers actually show
The excitement reads like this: 400 million people now shop by talking to an AI, they spend half as much again, so voice commerce has arrived. I don’t think the numbers say that.
Start with the 50%. Amazon compared people who used the assistant with people who didn’t. The people who try a new shopping feature on Amazon are the people who already shop on Amazon the most. Heavy shoppers find the new button. Light shoppers don’t. So “users spend 50% more” is at least partly a description of who picked it up, not of what it did to them. Amazon didn’t publish a holdout test, and until it does, the 50% is a correlation.
Then look at what the assistant is being built to do. The new holiday features are gift lists and managing Subscribe and Save. That is reordering, renewing and remembering. The shopper already knows what they want, and the assistant removes a few taps from getting it again. That is a long way from an open question and the AI persuading someone to buy something new.
And the product only became “Alexa” by renaming. For two years it was Rufus. Whatever a customer speaks into it, the thing underneath is reading product data, order history, prices and reviews, and that is all text.
Where voice helps, and where it doesn’t
My position on voice AI hasn’t moved. It helps on input. I dictate my instructions to Claude with Wispr Flow because talking is faster than typing. “Reorder the dog food” is an instruction, and voice is a fine way to give it.
It helps far less on output and on two-way commerce. Nobody wants a speaker reading out six product options with their specifications. And more of these conversations will be agent to agent. A shopper’s assistant asking a merchant’s system which sizes are in stock will exchange data, not speech.
So the 400 million figure tells me that a very large retailer has put a reorder layer on top of the best product database in retail.
The part a smaller business should actually read
The line in the report that matters is that Amazon blocks ChatGPT and Muse from its site and wants its shoppers using its own agent.
I’ve said agentic commerce is too early to call and it is 50/50 which way the platforms go. Here is the split happening in public. Amazon is big enough to keep outside agents out and run its own.
A smaller shop, a clinic, a service business or a brand selling through its own site doesn’t. If the open agents are shut out of Amazon, where do they go to answer “which of these is better for a sensitive scalp” or “who delivers by Friday in Singapore”? They go to whatever they can read: your site, your product pages, your policies. If that information is thin or written in adjectives, the agent names someone else.
That is why interest in GEO is rising, and why my own work has moved that way. The website is now one surface among many. Winning an agent’s recommendation takes attributes that are materially different from the next shop’s, written down in a way a machine can quote.
The conversation after the recommendation is also text and data. When a customer’s request has to become an action in your systems, a stock check, a booking, a refund within a limit, it is a data exchange, which is the layer AeroChat is built on. Voice may start it, but data runs it.
What to do
Don’t build a voice strategy off this announcement. Amazon’s number describes Amazon’s shoppers reordering on Amazon.
Do make your product and service data agent-readable now. For every product or service, write the attributes a buyer would ask about: who it’s for, what it does, what it costs, delivery times, sizes, limits, exclusions. Numbers and plain facts, no marketing language. Put them on your own pages and on other people’s pages too, because an agent checks more than one source before it names you.
Then test it. Ask ChatGPT, Gemini and Google’s AI Mode the questions your customers ask, and see whether you are named and whether what they say about you is accurate. Date the result, because the answers change week to week, and fix whichever part is broken.