For most of Amazon's history, the rules were simple: stuff your title with keywords, drive sales volume, and the algorithm would reward you with rankings. That era is over.
In a span of roughly 24 months, Amazon deployed a peer-reviewed AI knowledge system, launched a conversational shopping assistant used by over 300 million people, and then — just this month — folded that assistant directly into the main search bar under a new name. The underlying technology powering all of this has been building quietly in Amazon's labs for years, filed as patents and published as academic research that most sellers never read.
This article connects all of it in plain language — from the original A9 ranking logic that still operates underneath everything, through COSMO and Rufus, to the Alexa for Shopping launch on May 13, 2026 — and explains specifically what it means for how you market on Amazon today.
Start Here: What A9 Actually Does (And Why It Still Matters)
The A9 algorithm is the name Amazon has used since the early 2000s for its product search engine, originally built by a subsidiary called A9.com. Despite all the changes layered on top of it, the foundational logic is still running. Understanding it is the prerequisite for understanding everything else.
When a customer types a search, A9 operates in two phases. First, it runs a matching pass — pulling every product in the catalog that contains words related to that query. Then it runs a ranking pass, sorting those results by which products are most likely to generate a completed purchase.
That second phase is the one that matters for marketing. Amazon does not rank products by how many keywords they contain. It ranks them by how reliably they convert searchers into buyers. The algorithm's core goal has always been to maximize revenue per search result — not to surface the most informative or perfectly keyword-optimized listing.
A9 evaluates every listing on two axes: relevance (can Amazon match your product to the right query?) and performance (when Amazon shows your product, do people buy it?). You need both. High relevance with low conversion is invisible. High conversion with poor relevance never gets the chance to convert.
The performance signals A9 weighs include conversion rate, click-through rate, sales velocity, pricing competitiveness, review quality, inventory consistency, and seller account health. Of these, conversion rate carries the most weight — a product that turns 8% of visitors into buyers will outrank a keyword-stuffed listing that converts at 2%, all else being equal.
Two specific A9 signals deserve attention because they're widely underappreciated by sellers. First, pricing stability: products that change their price more than 20% more than once per month see ranking volatility increase significantly, because the algorithm interprets price instability as a signal of an unreliable offer. Second, stockout history: a seven-day stockout can suppress rankings by 30–50%, and full recovery takes three to four weeks of consistent performance after restocking. More importantly, repeated stockouts leave a lasting algorithmic footprint — the system flags the account as an unreliable supply source, triggering suppression that persists even after shelves are refilled.
None of this has gone away. A9 is still the foundation. What changed is what gets layered on top of it.
The "A10" Myth — And What Actually Happened
Around 2020, sellers started noticing that external traffic — visitors arriving from Google, social media, or email — seemed to give products a disproportionate ranking boost. The community coined the term "A10" to describe what felt like a new algorithm prioritizing organic behavior over paid-ad-driven sales. The name stuck, and hundreds of blog posts have been written about it since.
Here is the thing: Amazon has never confirmed the existence of an A10 algorithm. It is an informal community label for real behavioral changes in the ranking system. The effects sellers were observing were genuine, but the cause was not a cleanly named successor algorithm. It was a series of incremental updates to how A9 weights its signals — and, more significantly, the early deployment of AI systems that were formally published only later.
What is real is the shift in weighting. External traffic that converts now accounts for an estimated 15–20% of ranking contribution. The system began rewarding products that demonstrate demand beyond Amazon's own ecosystem, because such demand signals authentic market validation rather than optimized ad spend. This is why driving quality external traffic — not just any traffic — became genuinely important. Traffic that bounces tells the algorithm the opposite story.
"A10 is the community's name for real changes. COSMO is what actually explains them."
The more accurate explanation for the ranking behavior sellers were noticing after 2022 is a system called COSMO — and that's where the story gets significantly more interesting.
COSMO: The AI System Amazon Built and Published as a Research Paper
In June 2024, Amazon researchers presented a paper at SIGMOD — one of the most rigorous academic conferences in database and information retrieval research — describing a system called COSMO: the Common Sense Knowledge Generation and Serving System.
This is not a leak or speculation. It is peer-reviewed research published by Amazon's own scientists, describing a production system that has been deployed across Amazon's search applications.
COSMO is, at its core, a knowledge graph. Amazon trained large language models on hundreds of millions of customer interactions — searches, purchases, co-purchases, browsing patterns — and extracted from them a map of commonsense relationships between products, customer needs, and real-world contexts. The published system contains 6.3 million nodes and 29 million knowledge edges spanning 18 major product categories.
In benchmark testing using Amazon's own Shopping Queries Dataset from KDD Cup 2022, COSMO-enhanced search models achieved a 60% increase in macro F1 score over the best baseline when the knowledge graph data was the only variable. Even when all models were fine-tuned for the task, COSMO-backed models maintained a 28% edge in performance. These are significant margins in information retrieval research.
What does COSMO actually do in practice? It asks a fundamentally different question than A9's keyword matcher.
A9 asks: Does this listing contain the words the customer typed?
COSMO asks: Does this product solve the problem the customer described?
When a customer types "shoes for pregnant women," traditional keyword matching looks for listings that contain those exact words. COSMO knows — from the commonsense knowledge it extracted from millions of real co-purchases — that pregnant women frequently buy slip-resistant shoes. It surfaces those products even if the listing never uses the phrase "shoes for pregnant women."
COSMO maps products using 15 relationship types, which include things like: what function the product serves, who the intended audience is, what occasion it's used for, what problem it solves, what it's frequently purchased alongside, and what context makes it relevant. A listing that communicates only what a product is — through keywords — is now competing against listings that communicate who it's for, why they need it, and when they'd use it.
What this means for marketing
Your product listing is no longer primarily a keyword document. It is a context document. The listings that perform well under COSMO are the ones that read like they were written for a specific person solving a specific problem in a specific situation. That means your bullet points, your title structure, your A+ content, and even your backend product attributes all need to be built around use cases and audiences — not keyword density.
Backend product attributes in Seller Central (subject matter, target audience, intended use) now feed directly into how COSMO categorizes and surfaces your product. Most sellers fill these out minimally or leave them blank. That is now a competitive mistake.
The Patent Record: What Amazon Was Building While Nobody Was Watching
COSMO is the most significant published research, but it does not exist in isolation. Amazon Technologies has filed and received a substantial body of patents over the past decade that, read together, describe a search system being systematically rebuilt around machine learning, behavioral signals, and intent understanding.
A few worth knowing:
These describe a system that trains machine learning models on user search queries and factorizes raw data — including query data, relevance, temporal data, transaction data, impressions, demand, and supply — to rank results. Multiple goal models can run simultaneously, one optimizing for revenue generation, another for cross-product marketing. The ranking is not a fixed formula. It is a continuously retraining model.
Describes a system that assigns higher ranking priority to products based on historical user action data, then trains a machine learning model on the differential between how competing products performed. In plain terms: the system learns which products win head-to-head in real customer behavior, not just which have the most relevant keywords.
Describes a system that creates an index to retrieve and augment generation of a response to a natural language request using a generative machine learning model. When a natural language request is received, a search representation is generated, relevant portions of documents are retrieved, and a generative model produces the response. This is the technical architecture behind conversational product discovery — what Rufus and Alexa for Shopping run on.
Originally filed by A9.com — the original search engine subsidiary — and reassigned to Amazon Technologies in October 2024. The timing of the reassignment is notable: it coincides with the full deployment of COSMO and the expansion of Rufus, suggesting a consolidation of the technical infrastructure.
Taken together, these patents describe a system that has been steadily moving away from rule-based keyword ranking toward a continuously learning, behaviorally-driven model that now incorporates natural language understanding at its core. This did not happen overnight, and it was not a surprise to anyone reading the filings.
Rufus: When the Shopping Assistant Became the Search Engine
Rufus launched in beta in February 2024 and expanded to all US customers in July 2024, ahead of Prime Day. By the end of 2025, over 300 million customers had used it, and it had contributed an estimated $12 billion in incremental annualized sales according to Amazon's Q4 2025 earnings materials.
The significance of Rufus was not primarily that it could answer questions. It was that it changed where product discovery happened. Instead of a customer typing keywords and scanning a grid of results, they were having a conversation — describing what they needed, asking for comparisons, specifying constraints — and Rufus was making product decisions on their behalf.
Rufus draws on product listings, customer reviews, Q&A sections, and the contextual intelligence mapped by COSMO. When a customer asks "what's a good gift for a runner training for their first marathon," Rufus does not run a keyword query for "marathon gift." It reasons about the context, draws on COSMO's understanding of what runners at that training stage typically need, cross-references review sentiment to assess quality signals, and surfaces products that fit the inferred need.
Rufus recommendations are 83% weighted toward what serves Amazon's ecosystem, and they draw heavily on review data to generate suggestions. This means review acquisition strategy is now directly tied to AI visibility — not just social proof for human browsers. A product with thin review coverage is effectively invisible to the conversational layer of Amazon's search.
Amazon also quietly began integrating sponsored placements into Rufus conversations. Rufus can generate its own ad copy based on product descriptions, reviews, brand posts, and keywords from existing campaigns — even when those campaigns were not specifically targeting the product being discussed. This represents a fundamental shift: your advertising creative is no longer fully under your control once it enters the AI recommendation layer.
Alexa for Shopping: The Change That Happened This Month
On May 13, 2026, Amazon retired the Rufus brand and launched Alexa for Shopping — a unified AI shopping agent operating across Amazon.com, the Amazon Shopping app, Alexa.com, the Alexa app, and Echo devices. The change is not a rebrand. It is a structural change to how customers encounter products.
Previously, Rufus appeared as a chatbot in the corner of the screen — an optional tool that customers could choose to engage with. Alexa for Shopping is now embedded directly in the main search bar. When a customer types a question into the Amazon Shopping app, Alexa for Shopping automatically provides a conversational answer. The traditional keyword search results page still exists for direct product queries, but the entry point to that page now runs through an AI layer first.
Rufus and Alexa+ now share one persistent shopper profile, one product-research history, and one purchase context. The system can view up to one year of a product's price history and surface that data in comparisons. It can create shopping guides, find deals, automate routine purchases, and — critically — it can shop for products at other retailers, not just Amazon.
That last point has significant competitive implications. Alexa for Shopping is an agent that can compare your product against a competitor listed on a different platform, present the results to a customer, and facilitate the purchase — potentially not on Amazon at all. This makes cross-platform listing quality and pricing strategy more consequential than it has ever been.
The timeline that led here
Amazon presents COSMO at ACM SIGMOD — the first public disclosure of the commonsense AI knowledge system powering intent-based search.
Rufus expands to all US customers, with over 300 million users reached by year's end and $12B in incremental annualized sales by Q4 2025.
OpenAI launches Instant Checkout inside ChatGPT. Google builds Buy for Me into Gemini. The race for AI-mediated shopping intensifies.
Amazon unveils Alexa+, its generative AI-powered assistant, beginning the consolidation of Rufus and Alexa into a single agent. Amazon also launches the Alexa for Shopping FAST channel on Prime Video and Fire TV.
At CES 2026, Amazon announces Alexa+ integrations with BMW, Samsung televisions, Bosch appliances, and Oura health devices — embedding the shopping assistant across third-party hardware for the first time.
Rufus is retired. Alexa for Shopping launches as a unified agent in the Amazon search bar, the app, the website, and Echo devices. Conversational answers now appear automatically when customers type questions.
What This Means for Your Amazon Marketing — Specifically
The combined effect of COSMO, Rufus, and Alexa for Shopping is that Amazon now has two parallel discovery systems running simultaneously: the traditional keyword search that still drives the majority of clicks, and an AI conversational layer that intercepts a growing share of high-intent queries. Your marketing strategy needs to address both.
1. Your listing is now a conversation script
Alexa for Shopping and Rufus read your listing the way a researcher reads a brief — looking for evidence that your product answers a specific customer's problem. Listings built on keyword density give the AI little to work with. Listings built around use cases, audiences, and outcomes give it the raw material to surface your product in response to conversational queries you never explicitly optimized for.
Practically: rewrite your bullet points to lead with the customer situation, not the product feature. "For runners training in low-light conditions" is more useful to the AI than "LED illuminated." Both describe the same thing — one is written for a keyword crawler, the other is written for an intent engine.
2. Reviews are now your AI advertising budget
Rufus and Alexa for Shopping draw heavily on review data to generate recommendations and synthesize comparisons. A product with 12 reviews and a 4.6-star average competes very differently in the AI layer than in traditional search, where it might still rank reasonably based on conversion rate. In the conversational layer, thin review coverage means the AI has almost nothing to draw from when a customer asks for a recommendation in your category.
This reframes review acquisition not as trust-building for human browsers — though it remains that too — but as a direct input to how often and how favorably your product appears in AI-generated responses. Review recency matters as much as volume, because the AI weights current customer sentiment over historical averages.
3. Your PPC data is temporarily unreliable
With the launch of Alexa for Shopping, voice, chat, and click conversions now collapse into a single session. Your existing PPC dashboards report on paid placements in traditional search results — they do not yet have visibility into conversions that originated from an Alexa for Shopping conversation before the customer clicked to a product page. Amazon has indicated placement-level breakdowns are coming, but they are not live yet.
The practical effect is that your attributed sales figures currently undercount the contribution of the AI layer, while your ACoS may look artificially high because it is being measured against a denominator that does not include AI-assisted conversions. Before making significant bid adjustments based on recent performance data, account for this measurement gap.
4. Price history is now visible to shoppers — and to the AI
Alexa for Shopping surfaces up to a year of price history when making comparisons or recommendations. This has an immediate implication for promotional strategy: deals that follow a period of artificially inflated pricing will be identified and flagged. The AI optimizes for the customer, not the seller. Honest, substantive promotions will win AI-mediated comparisons. Price manipulation strategies that were invisible to keyword search are now explicitly surfaced in conversational responses.
5. Brand presence outside Amazon now feeds your Amazon ranking
The connection between external traffic and organic ranking — the "halo effect" that sellers have been observing since roughly 2020 — is now more consequential, not less. When someone finds your brand through a Google search, a YouTube review, or a social media post, clicks through to Amazon, and completes a purchase, the algorithm interprets that as market validation that keyword-only traffic cannot replicate. COSMO's knowledge graph likely incorporates these co-purchase and cross-platform signals into its understanding of your product's relevance.
This is why the Amazon Brand Referral Bonus — which credits sellers roughly 10% on sales driven from external traffic via Amazon Attribution links — is not just a cost-saving program. It is a documented mechanism for sending conversion-quality external signals to the algorithm while recovering a portion of your acquisition costs.
| Signal | How A9 Weighs It | How COSMO / Alexa Changes It |
|---|---|---|
| Keywords in title | Direct ranking factor | Still needed for matching; less dominant for ranking |
| Conversion rate | Primary ranking signal | Unchanged — still the most important signal |
| Customer reviews | Trust signal for buyers | Now feeds AI recommendation engine directly |
| Use case language in bullets | Minimal effect beyond keyword indexing | Core input for COSMO intent mapping |
| Backend product attributes | Indexing support | Direct COSMO categorization signal |
| External converting traffic | ~15–20% ranking contribution | Algorithmic validation signal; feeds halo effect |
| Price history | Pricing competitiveness factor | Now surfaced directly in AI comparisons |
| Q&A section | Crawlable, minor relevance boost | Source material for Rufus / Alexa responses |
| Inventory consistency | Stockout causes ranking suppression | Unchanged — repeated stockouts leave lasting footprint |
What to Do Right Now
The changes described above are not coming — they are already live. Alexa for Shopping launched on May 13, 2026. COSMO has been deployed across Amazon's search applications since Q3 2024. The following actions address the current state of the platform, not a hypothetical future.
- Audit your bullet points for intent language. For each bullet, ask: does this tell the AI who needs this product and why? "Perfect for new parents managing nighttime feeds" is intent language. "500ml capacity with leak-proof lid" is feature language. You need both, but most listings have only the latter.
- Complete every backend product attribute field in Seller Central. Subject matter, target audience, intended use, and material/fabric fields all feed COSMO's categorization. Blank fields are missed ranking opportunities in the AI layer.
- Build your Q&A section deliberately. Answer your own questions using natural language that mirrors how customers would describe the problem your product solves. Alexa for Shopping and Rufus read Q&A as source material for conversational responses.
- Restructure your deal strategy around honest pricing. With a year of price history now visible in AI-generated comparisons, manufactured discount cycles are exposed. Build your promotional calendar around genuine price events with clear value for customers.
- Do not make significant PPC bid changes based on very recent data. The measurement gap created by Alexa for Shopping is real. Wait for Amazon Ads to release placement-level breakdown data before drawing conclusions about campaign efficiency.
- Set a baseline for branded-term ACoS this week. As Alexa for Shopping gains share of the discovery funnel, branded search behavior will shift. Recording your current baseline now gives you a meaningful comparison once the AI layer's effects become measurable.
- Prioritize review recency, not just volume. Three recent, detailed reviews carry more weight in the AI recommendation layer than fifty old ones. Build a consistent review acquisition system that produces steady new reviews rather than periodic volume spikes.
- Set up Amazon Attribution links for all external channels. Every converting click from outside Amazon now sends a dual signal: it improves organic ranking through the halo effect and qualifies for Brand Referral Bonus credits. These are two separate returns on the same traffic.
The Bigger Picture
What Amazon has built is not just an updated search algorithm. It is a full-stack AI layer that sits between your product and your customer — interpreting intent, synthesizing information, making recommendations, and in some cases completing purchases without the customer ever actively searching for your specific listing.
This is the same competitive pressure that prompted Amazon to build it in the first place. OpenAI's Instant Checkout, Google's Buy for Me, Alibaba's Qwen integration — every major platform is building an AI agent that mediates the shopping decision. Amazon's answer is to make its agent the most trusted, most personalized, and most deeply integrated with purchase behavior data of any competitor.
The sellers who will perform best in this environment are not the ones who master the next keyword trick. They are the ones who build listings, review profiles, and external brand presence that give the AI the richest possible material to recommend them — accurately, consistently, and to the right customer at the right moment.
The fundamentals of good marketing have not changed. The surface they operate on has.
Leave a comment