From A9 to A10 and Rufus: How Amazon Search Ranking Actually Works in 2026

Most guides still open by explaining the "A9 algorithm" as if it were the live system deciding who ranks today. It is not, and has not been for a while.

Amazon ranking in 2026 runs in two stages. First, a retrieval and ranking system pulls the listings that match a query and orders them by relevance and likelihood to convert. Second, a conversational assistant, launched as Rufus and now folded into Alexa for Shopping, reads your listing as a source of information and decides whether to surface, compare, or recommend your product inside a chat-style answer. The listing that wins is the one written for both readers.

That second reader is the story of this piece. We will name it, walk the mechanism, separate what is documented from what the seller community made up, and show where a manual optimization pass stops scaling.

Quick Answer

How does Amazon ranking work in 2026?

• Amazon's search system, historically nicknamed A9, retrieves listings that match a query, then ranks them primarily on relevance and conversion signals.
• "A10" is a seller-community label for how ranking behaves now. Amazon has never released a public algorithm named A10; the underlying system is still A9's lineage, retuned.
• Rufus, Amazon's conversational shopping assistant, added a second layer on top of ranking. As of May 2026 it was unified into Alexa for Shopping and embedded in the search bar, results page, and product pages. It reads listings to interpret intent, not to match keywords.
• What you control: listing content, structured attributes, price competitiveness, fulfillment reliability, and the conversion experience. What you do not control: the weighting, the ordering, and the shopper's exact words.
• The practical shift: stop writing listings as keyword documents. Write them as the primary source a ranking system and a conversational assistant both read to decide whether you deserve the click and the recommendation.

We call this the Second Reader. Your listing is read twice before a shopper forms an opinion: once by the ranking system that decides whether you appear, once by the conversational assistant that decides whether you get recommended. Optimizing for one and ignoring the other is the most common mistake on the platform right now.

A9 vs. A10: One Is a Real Lineage, the Other Is Folklore

A9 comes from A9.com, the search subsidiary Amazon founded in 2003. Its public web search engine wound down in 2008, but the team and technology continued inside Amazon as the engine behind product search. So when older guides say "the A9 algorithm," they point at a genuine internal lineage, not an invented term.

A10 is a different situation. Amazon has never announced, documented, or confirmed a public ranking system called A10. Sellers and agencies started using the label around the late 2010s to describe observed behavior changes: external traffic, organic sales, and account health seemed to carry more weight than they used to, while raw ad-driven velocity carried less. It stuck as shorthand, not because Amazon shipped a version 10.

This matters practically. When someone sells you "A10 optimization" as a distinct, documented system with a known rulebook, they are selling the community's best guesses dressed as official mechanics. The observed shifts are real enough to plan around; the version number is folklore. Treat "A10" as a description of behavior and you make good decisions. Treat it as a leaked spec and you chase rules that were never published.

The clean way to hold all three names at once:

• A9: the documented lineage. Amazon's product search technology, named for the A9.com subsidiary. Still the ancestor of what runs today.
• A10: community shorthand for how ranking behaves now. Descriptive, not official. No public Amazon system carries this name.
• Rufus / Alexa for Shopping: the conversational layer, and the genuinely new thing. This is documented, launched, and live.

How Ranking Works in 2026: Retrieval First, Then Conversion Decides

Strip away the naming debate and the mechanism is stable. When a shopper types a query, the system does two things in sequence.

Step one: retrieval. The system pulls the set of listings relevant to the query, mostly on text and structure: whether your title, bullets, description, backend search terms, and structured attributes relate to what the shopper asked for. Search "insulated water bottle 32 oz" and a listing that never establishes it is insulated, or never states capacity in a field the system can read, may not enter the candidate set at all. You cannot rank for a query you were never retrieved for. This is the step most sellers underinvest in, because it is invisible: you never see the listings that failed to show up.

Step two: ranking. Among the retrieved candidates, relevance still matters but performance signals do the heavy lifting. The dominant one is query-specific conversion. Amazon makes money when shoppers purchase, so it promotes the listings that reliably convert the traffic they receive.

The signals that feed ranking fall into a few plain buckets:

• Relevance signals: title, bullets, description, backend search terms, and structured attributes.
• Conversion signals: click-through rate from search, and the rate at which those clicks become orders, against the specific query.
• Customer experience signals: price competitiveness, fulfillment speed and reliability, in-stock consistency, return rate, and review volume and quality.
• Seller and account health signals: order defect rate, feedback, and fulfillment track record.

Relevance gets you into the room, conversion decides where you stand in it, customer experience decides whether you stay. Advertising buys placement and can seed early conversion history, but paid velocity alone does not hold an organic rank if the listing does not convert on its own. This is the behavior sellers file under "A10." Amazon is optimizing for its own sell-through and reads your conversion rate as the clearest evidence of whether you help.

One caution on numbers. You will find posts assigning precise weights, "text is 35 percent, behavior is 30 percent." Amazon does not publish those; treat any exact percentage as an estimate. What is documented in Seller Central is the direction: fill every relevant structured field, write for the shopper rather than for keyword density, and earn conversion by being the obviously correct choice for the query.

Rufus Didn't Replace Ranking, It Added a Second Reader

Rufus launched as Amazon's generative shopping assistant, trained on the product catalog, reviews, community questions, and information from across the web. Amazon reported more than 300 million customers used it in 2025. In May 2026, Amazon unified Rufus into Alexa for Shopping, moving it out of a side chat window and into the search bar, results page, and product detail pages for signed-in U.S. customers. Same capability, far more central placement.

Traditional retrieval matches your listing's words against the shopper's words. The assistant does something different. Ask "what's a good water bottle that keeps ice overnight for hot yoga" and it is not hunting for a listing with that exact phrase; it reads listings to understand which products satisfy the intent behind the sentence. Your listing is no longer only a keyword document to be matched. It is a source document to be comprehended. Ambiguity, missing attributes, and marketing fluff that a keyword match would tolerate now cost you, because the assistant cannot recommend what it cannot confidently understand.

What the Second Reader rewards:

• Specific, factual claims. "Keeps drinks cold 24 hours, hot 12 hours" is comprehensible. "All-day temperature control" is not.
• Complete structured attributes. Material, capacity, dimensions, use case, and compatibility, filled in, give the assistant grounded facts to reason over.
• Reviews and Q&A that resolve real questions. The assistant reads community content as evidence. A product with reviews that answer "does it leak in a bag" has an advantage when someone asks exactly that.
• Consistency across the listing. If the title, bullets, and attributes disagree, you have given the second reader a reason to hesitate.

What the Second Reader does not change: it does not remove ranking, it does not eliminate the value of relevance and conversion, and it does not reward keyword stuffing. If anything, stuffing reads as noise to a system built to interpret meaning.

Amazon has been building the assistant-facing surface fast: it introduced its Ads Agent at unBoxed in November 2025 and opened the Amazon Ads MCP Server to global open beta on February 2, 2026. Assistants are becoming a primary interface between shoppers and the catalog, and listings are the ground truth those assistants read.

You Control the Inputs, Not the Weights

You control:

• Listing content: title, bullets, description, A+ content, images, and video.
• Backend search terms and every structured attribute field.
• Price and promotional posture.
• Fulfillment choice and in-stock reliability.
• The post-click experience that drives conversion, from images to review generation to answering questions.
• Account health, through operational discipline.

You do not control:

• The exact weighting of any signal. Amazon does not publish it and changes it.
• The ordering of results for a given shopper, which is personalized and contextual.
• The shopper's exact words, or what the conversational assistant infers from them.
• Whether a competitor with a stronger conversion history sits above you today.

The job is to make every input you control unambiguous and correct, then let conversion evidence accumulate. You are not programming the ranking system; you are giving it the cleanest read of what you sell and the strongest evidence that shoppers who want it buy it.

Common Misconceptions Left Over From an Older Amazon

Let us clear the frequent ones.

• "A9 is the current algorithm, and there is a separate A10." A9 is the lineage. A10 is a community label for behavior, not a released system. Neither is a rulebook you can look up.
• "Keyword density drives ranking." Relevance requires that the right terms appear in the right fields. Density beyond that does nothing, and reads as noise to the conversational assistant.
• "Ads buy permanent rank." Ads buy placement and can seed early conversion history. Organic rank is held by conversion, not by spend. When the campaign pauses, an unconverting listing slides.
• "External traffic is a magic ranking lever." External traffic that converts helps, because it is conversion. External traffic that bounces is just traffic. The signal Amazon reads is the sale, not the source.
• "Rufus replaced search, so listings matter less." The opposite. The conversational assistant reads your listing as its primary source of truth, which makes listing quality matter more, not less.

Here is how the old assumptions line up against 2026 reality.

That last row is the one worth sitting with. A tool that checks whether a title contains a keyword is easy to build. It cannot tell you whether "all-day cold" reads as a confident, recommendable claim or as vague filler. The gap between matching a string and judging a meaning is the gap the Second Reader opened up.

A Concrete Example

To make the Second Reader tangible, here is one listing moving through the 2026 pipeline.

Input. A stainless steel insulated bottle. Current listing:

• Title: "Premium Water Bottle Stainless Steel Sports Gym Travel Bottle BPA Free"
• Bullets emphasize "durable," "leak resistant," and "great for any occasion."
• Attributes: capacity left blank, insulation field left blank, material filled.
• Reviews: 400 reviews, several mention it kept ice overnight and did not leak in a bag.

Workflow. A shopper asks the conversational assistant, "I need a bottle that keeps ice overnight and won't leak in my gym bag for hot yoga."

The retrieval system checks the candidate set for the query terms. The listing is missing "insulated" as a stated fact and has no capacity or insulation attributes, so it is a weaker retrieval candidate for temperature and size intents. The conversational assistant, reading the listing to answer the shopper, finds "durable" and "leak resistant" but no clear statement on how long it holds temperature and no structured insulation data. The strongest evidence, that it keeps ice overnight and does not leak in a bag, is buried in the reviews rather than stated in the listing.

Expected output. A competing listing that states "keeps ice up to 24 hours," fills the insulation and capacity attributes, and says "leak-proof lid, tested in a packed bag" gets recommended for this intent. Our listing may not surface at all, because it neither retrieved strongly nor gave the second reader a confident, groundable claim.

Business decision. The fix is not more keywords. It is to make the listing say what the reviews already prove: state the temperature retention as a specific claim, fill the insulation and capacity attributes, and align the title, bullets, and structured data so both readers get one consistent, comprehensible story. Then let conversion accumulate. This is a judgment call about meaning and specificity, made listing by listing, which is precisely the work that does not compress into a keyword rule.

Where a Manual Optimization Pass Runs Out: Judgment Doesn't Scale by Hand

A single listing, you can fix in an afternoon. Read it the way a shopper would, notice the vague claims, cross-reference the reviews, fill the empty attributes, align the story. That is the right work.

Now do it across thousands of SKUs, on a schedule, as reviews accumulate new evidence, competitors sharpen their claims, and the assistant keeps raising the bar on what counts as comprehensible. The task did not get harder per listing; it got impossibly repetitive at volume. And the part that resists spreadsheets is the judgment: whether a claim is specific enough, whether the title and attributes agree, whether the reviews contain a proven fact the listing fails to state. A find-and-replace macro has no way to know what "comprehensible to a conversational reader" means.

That is the ceiling of a manual pass. Not that people cannot do the judgment, but that they cannot do it consistently, on every listing, every week, without the standard drifting as fatigue sets in.

Where Qore Fits: Codifying the Listing-Quality Judgment

Qore is the operator's judgment written down and run on a schedule. For the work in this piece, that means the semantic review a numeric rule cannot make: whether a listing states its core attributes as specific claims, whether the title and structured data agree, whether the reviews contain evidence the listing fails to surface, and whether a claim will read as confident or vague to a conversational assistant. It runs listing by listing at the volume a manual pass cannot hold, and returns the same output from the same inputs, with logic you can inspect rather than a score you take on faith. The point is one consistent standard across every listing, the standard you would apply yourself if you had time to read all of them every week.

The honest limits. Qore does not replace your bid engine; ranking still depends on conversion, and conversion still depends on price, fulfillment, and demand that a listing review does not touch. Codifying a standard requires the standard to exist first: if your team has not decided what a good listing claim looks like, Qore has nothing to enforce. It scales your judgment; it does not invent it. As of July 2026 Qore is on a public waitlist, and its actions default to manual approval, so a person signs off before anything ships.

If you are not there yet, the more useful next step is upstream: write down what a recommendable listing looks like for your category, using the Second Reader lens from this piece, before you automate anything.

Conclusion

The naming argument, A9 versus A10, was always a distraction from the change that mattered. Amazon retrieves listings that match a query, ranks them mostly on how well they convert, and rewards operators who make every input they control unambiguous and correct. What changed is that a second reader arrived. The assistant that launched as Rufus, now inside Alexa for Shopping and sitting in the main shopping path, reads your listing to comprehend it, not to match it. Write for that reader and the old one at once and you are optimizing for the Amazon that exists in 2026, not the one stale guides describe.

None of this is finished. Assistants will keep getting more central, and the bar for a comprehensible, recommendable listing will keep rising. The operators who win treat their listings as the source of truth two systems read, and hold that standard across the whole catalog rather than the one SKU they had time for this week.