A9, A10, COSMO, Alexa for Shopping: which of these is real

Three of those four names refer to something that exists. One does not, and it is the one you have read the most about. Here is what Amazon has actually built, actually published, and actually renamed — and what the industry invented.

A9 was real. It is also gone.

Amazon confirmed A9.com was a genuine Amazon subsidiary founded in 2003 under Udi Manber. The name is a numeronym for “algorithm” — A, then nine letters. It built product search for Amazon and for other retailers, along with Search Inside the Book and the visual search that was later rebranded as Lens.

The 2016 SIGIR paper that remains the single best public description of how Amazon ranks products — Amazon Search: The Joy of Ranking Products — carries the author affiliation “A9.com” on its face.

Amazon confirmed In 2019, Amazon took down the A9.com site and redirected the domain to Amazon's homepage. The organisation was absorbed into Amazon Search. There has been no entity called A9 for roughly seven years.

A10 was never anything

Contested “A10” is not an Amazon system. It has never been announced by Amazon. It appears in no Amazon documentation, no patent, no publication, no engineering blog and no Seller Central page. It has no version number because there is no versioned product for it to be a version of.

It is industry coinage from roughly 2019–2020, invented to package a genuine observation: that Amazon appeared to be weighting performance and off-Amazon signals more heavily than the early-2010s understanding suggested. That observation may well be sound. The name, and the implied versioned release with a documented factor list, is fiction.

The reason this matters is not pedantry. It is that “A10” content is where the fabricated precision lives. Ranked and weighted factor lists — “conversion rate is 27% of the algorithm” — are downstream of pretending there is a documented system to enumerate. There is not.

COSMO is real, published, and badly covered

Amazon confirmed COSMO is a genuine Amazon system, documented in a peer-reviewed paper at SIGMOD 2024: COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon.

What it actually is: a pipeline that distils commonsense assertions from large language models, filters them with critic classifiers trained on around 30,000 human annotations, then scales through a fine-tuned model across 18 product categories to produce millions of knowledge assertions. It captures user intent and behavioural context — how people think about and use products — rather than product attributes. The paper states it has been deployed in Amazon search applications, particularly search navigation, with significant improvement in offline and online A/B tests.

Contested Everything beyond that is extrapolation. Amazon has not published COSMO's weight in ranking, whether it touches core ranking as opposed to navigation and query understanding, or any seller-facing optimisation surface. The substantial body of “how to optimise for COSMO” content is inference from a systems paper, and should be read that way.

What is defensible to take from it: Amazon is investing heavily in intent understanding rather than string matching. That direction is consistent with everything else it has published on semantic retrieval, and it argues for listings that describe use case, context and audience rather than stacking synonyms.

Rufus is now Alexa for Shopping

Amazon confirmed Rufus launched in February 2024 as a generative-AI conversational shopping assistant. On 13 May 2026 Amazon launched Alexa for Shopping, powered by Alexa+, which replaces it. It is voice- and touch-enabled across mobile, desktop and Echo Show, and adds product comparison, price history, price-drop alerts, automatic purchase at a target price, recurring orders and a “Buy for Me” capability that shops beyond Amazon. US customers only at launch.

Amazon confirmed On scale, Amazon's own disclosures are the usable numbers. Rufus reached over 300 million customers in 2025 and was credited with roughly $12 billion in incremental annualised sales. On the 30 April 2026 earnings call Andy Jassy reported monthly active users up 115% year on year, engagement up 400%, and that customers using it were 60% more likely to complete a purchase.

Publish that alongside the caveat it deserves: “$12 billion incremental” is an Amazon-internal counterfactual attribution, resting on assumptions about behaviour that introduce real uncertainty.

Why any of this matters for your listings

Observed Here is the finding that makes the distinction practical rather than academic. Amalytix analysed more than 1,300 AI-assistant recommendations by taking the top 500 generic search terms, converting them to conversational queries, and recording the ASINs returned.

Recommended products showed weak correlation with traditional organic search rank: a median organic rank of 41, with the 75th percentile undetectable in organic search altogether. What the recommended products did have in common: a 4.0-star minimum and 4.5-star median, a median of 2,991 reviews, six to nine images, 55% with A+ Content, 32% with Premium A+, 92% Prime-eligible, 94% FBA, and 75% unique brands — the assistant actively diversifies across brands.

The conclusion that follows is genuinely differentiated: AI-assistant visibility and organic keyword rank are partly independent surfaces. Winning one does not deliver the other. That is a strategic fact, and it is invisible if you think of “the Amazon algorithm” as a single thing with a version number.

The four systems, and how they differ →

Sources