Key Takeaways
- Amazon’s search system no longer matches keywords. It runs a knowledge graph called COSMO and a shopping AI called Rufus that require structured, complete listing data to recommend products.
- Five backend fields in Seller Central — Search Terms, Intended Use, Target Audience, Subject Matter, and Other Attributes — are the most commonly left blank, and empty fields sever connections in the knowledge graph that Rufus uses to make recommendations.
- A 25-to-1 sales difference between two products from the same brand came down to data completeness: one listing had every backend field filled out, the other left most fields empty.
- Rufus uses Retrieval-Augmented Generation to answer shopper questions directly from listing content, so bullet points written as complete, readable sentences perform significantly better than keyword-stuffed headers.
- Semantic changes to listings take 7 to 14 days to propagate through COSMO’s knowledge graph, so ranking improvements will not appear overnight after backend updates.
General Summary
Amazon’s search infrastructure has shifted away from keyword matching toward a graph-based AI architecture. COSMO, Amazon’s Common Sense Knowledge Generation system, maps relationships between products, buyer intent, and purchase behavior at scale. Rufus, the shopping AI built on top of it, answers shopper queries by retrieving information directly from listing content through a process called Retrieval-Augmented Generation. The practical consequence for sellers is significant: listings that provide incomplete structured data are not recommended. Rufus does not guess. It routes shoppers toward whichever listing provides the most complete, machine-readable information. Across more than 600 million products, the listings that give Rufus the most to work with receive disproportionate traffic. The sellers who understand this shift and audit their backend fields are capturing organic rankings that keyword optimization alone cannot recover.
Extractive Summary
Amazon’s search now operates through two systems: COSMO, a knowledge graph, and Rufus, a shopping AI that uses Retrieval-Augmented Generation to answer shopper questions from listing data. Five backend fields in Seller Central — Search Terms, Intended Use, Target Audience, Subject Matter, and Other Attributes — are the primary connection points Rufus uses to match products to queries. A 25-to-1 sales difference between two products from the same brand was explained entirely by backend data completeness. Rufus processes listing bullet points as source material for answering shopper questions, so conversational, complete sentences outperform keyword-dense formatting. Images are processed through computer vision and OCR, Q&A sections are indexed as ground truth, and semantic listing changes take 7 to 14 days to fully propagate through COSMO.
Abstractive Summary
The transition from keyword-based to knowledge-graph-based search represents the most fundamental shift in Amazon listing strategy since Sponsored Products became the dominant traffic driver in the mid-2010s. For most of Amazon’s history, the central question of listing optimization was: which words does the algorithm match? That question no longer applies. The new question is: how much structured data can this listing supply to an AI that needs to speak on the product’s behalf? Sellers who grew their businesses on keyword density and search term volume are finding that the techniques that once worked actively underperform in a Rufus-driven environment. The algorithm now penalizes ambiguity. A listing that might have ranked adequately under A9 because it contained the right words can fail entirely under COSMO because those words exist without context, without categorization, without the labeled attributes that allow Rufus to confidently answer a specific shopper question. The sellers gaining ground in this environment share a common trait: they have stopped treating backend fields as optional metadata and started treating them as the primary mechanism through which their products get discovered.
What Does Amazon’s AI Actually See When It Reads Your Listing?
Amazon’s AI reads your listing as a structured data set, not as a page of text, and it builds a network of relationships from every field you fill in. The old system, Amazon’s A9 algorithm, worked through keyword matching. A shopper typed “stainless steel garlic press” and A9 looked for listings containing those words. Relevance was largely a function of keyword placement and density. That mechanism is no longer primary.
Amazon now runs two interconnected systems. COSMO, which stands for Common Sense Knowledge Generation, transforms Amazon’s product database into a knowledge graph. It does not store product data as flat text. It maps relationships between products, buyer intent, purchase behavior, and contextual signals. COSMO connects a product to the problems it solves, the people who buy it, and the situations in which it gets used.
Rufus is the shopping AI built on top of COSMO. When a shopper searches “best ergonomic chair for lower back pain,” Rufus does not look for listings containing those exact words. It queries the knowledge graph for products associated with lumbar support, extended sitting comfort, and orthopedic positioning. Products that have supplied that information through structured backend fields appear. Products that have not are bypassed.
Rufus operates on Retrieval-Augmented Generation, or RAG. This means Rufus pulls content directly from your listing to construct answers to shopper questions. If a shopper asks Rufus “does this press work for ginger root?” and your listing contains a sentence that says “designed to crush fibrous ginger root without bending” — Rufus lifts that sentence and uses it. If your listing contains no such information, Rufus does not fabricate an answer. It skips your product.
The mental model that matters here: every backend field is a connection point in the knowledge graph. Every empty field is a severed connection. Rufus follows connections. When yours are missing, it routes traffic to a competitor whose data is complete. The difference between a well-connected listing and a sparse one is not marginal in a system this size. With over 600 million products indexed, Rufus routes every additional unit of traffic toward the listings it can confidently recommend. Confidence comes from data.
Why Did Amazon Build This Kind of System?
Amazon built COSMO and Rufus because keyword matching could not scale to handle the complexity of modern shopper intent. A shopper searching “gift for a dad who loves grilling” is not describing a product. They are describing a situation, a relationship, and a purpose. Keyword matching has no mechanism to bridge that gap. A knowledge graph does.
Amazon’s primary business interest in search is accuracy, not coverage. A bad recommendation costs Amazon a sale and erodes trust in Rufus over time. So the system is built to be conservative. It will confidently recommend fewer products rather than speculatively recommend many. That conservatism means listings with incomplete data are not edge-cased in — they are excluded.
For sellers, the implication is direct. The old approach of optimizing for maximum keyword coverage worked against a system that rewarded presence. The new approach requires optimizing for maximum data completeness against a system that rewards confidence. These require different tactics, different field priorities, and a different understanding of what your listing is actually doing.
Which Backend Fields Does Rufus Rely on Most?
Rufus relies on five backend fields in Seller Central more than any other structured data source: Search Terms, Intended Use, Target Audience, Subject Matter, and Other Attributes. In most listing audits, three or four of these are completely empty. Each one does a distinct job inside the knowledge graph.
What Should the Search Terms Field Actually Contain?
The Search Terms field accepts up to 250 bytes of hidden keywords, and the byte limit is the most important technical constraint to understand. If you exceed 250 bytes, Amazon does not truncate the excess. It ignores the entire field. Every keyword you entered becomes unindexed.
Search Terms should contain synonyms, common misspellings, alternate spellings, and regional variations that do not appear in your title, bullets, or description. Do not repeat words already present in your front-end content. Amazon’s indexing system already captures those. Use spaces between terms, not commas. Write all entries in lowercase. And count bytes, not characters — certain characters consume more than one byte.
A garlic press listing might use its 250 bytes for terms like “garlic crusher mincer peeler tool kitchen gadget press unpeeled cloves herb squeezer ginger” — covering the vocabulary a shopper might use without repeating what the title already contains.
What Does the Intended Use Field Tell Amazon’s Knowledge Graph?
The Intended Use field tells Amazon the activities, locations, events, and conditions for which your product is designed, and this information becomes the mechanism by which Rufus matches your product to situational queries. A travel pillow listing with Intended Use populated as “airplane travel, car rides, camping, neck support during long flights” becomes findable for a shopper searching “best pillow for long flights” even if those exact words do not appear in the title.
Without Intended Use, Rufus has to infer context from your front-end content. Rufus does not make confident inferences. It works from explicit, labeled data. If you have not told Amazon what your product is for in the field Amazon created for that purpose, Rufus will not fill in the gap.
A skincare product might include: “daily moisturizer, dry skin treatment, post-shower application, nighttime skincare routine, sensitive skin care.” Each phrase connects the product to a different query pattern. Each one represents a traffic stream the product would miss if the field were blank.
How Does the Target Audience Field Affect Who Sees Your Product?
The Target Audience field specifies the demographic or psychographic profile of the intended buyer, and Amazon uses this to filter recommendations when shoppers express identity-based search intent. Searches like “gifts for women over 50,” “products for new parents,” or “professional kitchen tools” carry audience signals that the knowledge graph matches to listings with corresponding Target Audience data.
Amazon expanded Target Audience Keywords in late 2025 across Pet Supplies, Home and Kitchen, and Beauty. If your category has this field available, it should be filled. The specificity of the entry matters: “adults over 50 with joint pain” will outperform “adults” as a target audience signal because it connects to more specific query patterns.
What Goes Into Subject Matter and Other Attributes?
Subject Matter is where you describe what the product is about and what problem it solves, written as if explaining the product to an AI that has never encountered it. Because that is precisely what this field is for. Rufus reads Subject Matter as context when it cannot fully infer purpose from the product name and category alone.
Other Attributes covers anything that aids categorization but did not fit into the other fields. Material type, use case details, product benefits not captured elsewhere. Both fields should be treated as indexable content, not optional notes.
Beyond these five fields, category-specific attributes extend the opportunity further. Depending on your category, you may have access to Material Composition, Care Instructions, Active Ingredients, Product Benefits, Special Features, Pattern, Color, Size, and Weight. Every one of these fields indexes. Amazon confirmed this: their AI decides which content to surface in responses, but it reads everything that has been supplied. Empty fields mean fewer connections. Fewer connections mean less traffic.
Are There Fields Only Accessible Through Flat File Uploads?
Some listing attributes are not visible in the standard Seller Central interface and can only be accessed through a flat file bulk upload. Downloading your category’s flat file template reveals columns for attributes that never appear in the UI, and those attributes index the same way as any other field. If you have only ever used the standard interface, you may have significant attribute gaps you are not aware of.
Additional bullet point slots follow the same logic. Amazon has expanded many categories to allow 10 bullets instead of 5. Each bullet is indexable, Rufus-readable content. A listing using 5 of 10 available bullet slots is forfeiting structured, machine-readable real estate at no cost.
What Does a 25-to-1 Sales Difference Actually Look Like in Practice?
A 25-to-1 sales difference between two products from the same brand, in the same category, at similar price points, was explained entirely by the difference in backend data completeness. Product A had every relevant field populated. Product B had most fields empty. Amazon’s algorithm could not recommend Product B because it lacked the data required to make a confident recommendation.
Product A supplied: a fully populated Target Audience, a complete Product Benefits list, Active Ingredients documented, Intended Use detailed, all available bullet slots used, and backend search terms optimized within the byte limit. Rufus had everything it needed to match that product to specific shopper queries and to answer follow-up questions about it.
Product B supplied almost none of this. No target audience, no product benefits, no active ingredients, incomplete search terms, fewer bullets. Rufus had no mechanism to say “this is suitable for dry skin” because the Intended Use field was empty. COSMO could not connect Product B to searches from “women over 40 looking for organic skincare” because the Target Audience field was blank. The knowledge graph had no connections to follow. So traffic did not follow either.
This is not a case where Product A had better ads or better reviews or a lower price. The listings existed in the same ecosystem with comparable front-end assets. The backend data gap produced the entire performance gap.
How Should Bullet Points Be Written for Rufus to Use Them?
Bullet points should be written as complete, readable sentences that Rufus can lift directly and present as answers to shopper questions, not as keyword-dense headers separated by commas. The technical description for this is RAG-ready content: content that a Retrieval-Augmented Generation system can extract without modification and use as a coherent response.
The distinction matters in practice. A bullet reading “HEAVY DUTY AND DURABLE: Made of 304 stainless steel, anti-rust, non-slip handle, dishwasher safe, great for garlic and ginger, best kitchen gadget” is a list of attributes assembled for human scanners. Rufus cannot quote it to a shopper. It contains no complete thoughts.
A bullet reading “Constructed from solid 304 stainless steel, this press features a reinforced hinge mechanism designed to crush unpeeled garlic cloves and fibrous ginger root without bending” is a sentence Rufus can extract and present as a direct answer to a shopper asking “does this work for ginger?” The information is the same. The structure is entirely different. And the structure is what determines whether Rufus can use it.
One seller rewrote 3,000 listings from keyword-stuffed to conversational format and recovered 72% of organic rankings within a week. The content was not new. The product did not change. The format changed: from keyword aggregations to complete sentences that Rufus could retrieve and present. That format change was the entire variable.
A practical test: read your bullet points out loud. If they sound like the output of a keyword tool — all-caps headers, comma-separated features, word after word without syntax — rewrite them as sentences a knowledgeable colleague would say to explain the product. That is the standard Rufus applies.
Which Listing Elements Beyond Backend Fields Does Rufus Index?
Rufus indexes images, Q&A sections, and review content in addition to structured listing fields, which means optimization opportunities extend well beyond Seller Central’s keyword and attribute tabs. Each of these elements feeds the knowledge graph through different mechanisms.
How Does Rufus Read Product Images?
Rufus is multimodal, meaning it processes both text and images. Text overlays on infographic images are read through optical character recognition and become indexable data points. If an infographic image displays “Supports up to 250 lbs” in clear text, that specification enters the knowledge graph as a product attribute. A listing can include claims in image format that the structured fields do not capture, and Rufus will still read them.
Lifestyle images are processed through computer vision. Amazon’s system tags lifestyle content with contextual labels: “outdoor use,” “kitchen setting,” “office desk environment.” These tags become attributes in the knowledge graph. A product shown in a camping context acquires the contextual tag of outdoor use even if outdoor use does not appear in any text field. That tag connects it to queries about camping.
Why Does the Q&A Section Carry More Weight Than Most Sellers Realize?
Rufus indexes Customer Questions and Answers as ground truth content, treating each Q&A pair as a verified statement about the product. This makes the Q&A section one of the few places where you can influence what Rufus says about your product without modifying the listing itself. Seeding five or more questions that match how your target customer actually searches creates structured, extractable answers Rufus can present verbatim.
Effective Q&A seeding means identifying the specific questions your target buyer has before purchasing and answering them in complete, factual sentences. “Does this work for sensitive skin?” answered with a full, specific response becomes a RAG-ready unit Rufus can retrieve. A vague or partial answer does not function the same way.
How Long Does It Take for Backend Changes to Affect Rankings?
Backend field updates and semantic listing changes take 7 to 14 days to fully propagate through COSMO’s knowledge graph, which is a meaningfully longer lag than sellers accustomed to A9’s keyword indexing expect. The old system could reflect keyword changes in rankings within 24 hours. COSMO does not operate on that timeline because it is rebuilding graph connections, not flipping keyword flags.
The practical consequence is that sellers who update their backend fields and check rankings the next morning will see no change. The knowledge graph has not rebuilt yet. Sellers who interpret that absence of change as evidence the update did not work and revert their listings interrupt the propagation process and return to their previous state.
Two weeks is the minimum observation window before drawing any conclusions from backend field changes. That timeline applies equally to bullet point rewrites, subject matter updates, and any other semantic content change. Treat it as a property of the system rather than a sign that something went wrong.
How Do You Verify That Your Keywords Are Actually Being Indexed?
Keyword indexation can be verified by searching your target keyword alongside your ASIN in the Amazon search bar. If your product appears in the results, the keyword is indexed. If it does not appear, something in the listing is preventing indexation, whether a byte-limit violation in Search Terms, a category mismatch, or a listing quality flag.
Several free browser extensions automate this check across multiple keywords simultaneously. The manual method works for spot-checking individual terms. For a full audit of a catalog with many SKUs, the extension approach reduces time considerably. Either way, indexation verification should follow every backend field update, not precede it.
What Is the Highest-ROI Listing Change a Seller Can Make Right Now?
Filling in every empty backend field across your active catalog is the highest-ROI listing change available to most sellers right now, because it requires no additional ad spend, no new creative assets, and no changes to the product itself. It addresses the exact mechanism Amazon’s AI uses to route traffic, and most catalogs have significant gaps.
The audit process is straightforward. Open each listing in Seller Central. Navigate to the Keywords and Description tab. Check Search Terms, Intended Use, Target Audience, Subject Matter, and Other Attributes. Document which fields are empty. Then download your category’s flat file template to surface attributes that are not visible in the UI. Every empty field is a missed connection in the knowledge graph. Every filled field is a signal Rufus can use.
After completing the backend audit, evaluate bullet point format. Identify bullets written as keyword lists and rewrite them as complete sentences. Prioritize the first two bullets on each listing, since Rufus weights earlier content more heavily when constructing answers. Then move to Q&A: seed five or more questions that reflect actual shopper searches and answer them in full, specific sentences.
This is the shift the last few years required: listings written for shoppers, then rewritten for an algorithm, now need to be written for an AI that speaks to shoppers on your behalf. The AI can only say what you have given it. Empty fields produce silence. Silence produces no recommendation. No recommendation produces no sale.
Amazon’s AI no longer matches keywords. It reads structured listing data to recommend products. Learn which backend fields Rufus requires and how to fill them to recover organic rankings.
