Jobs to be done · Diagnose · Ecom · Agency

How to improve your Amazon listings for Rufus.

Rufus answers shopper questions on your listing — and you cannot tell whether it answers them from your content or guesses. Your title and backend keywords were written for the old keyword index. Rufus does not use that game's rules.

Quick answer

Rufus picks products by traversing COSMO, Amazon's published knowledge graph of 15 shopper-intent relations. Add your Amazon SKUs to eCommerce Insights and the channel-aware engine scores each listing against all 15 relations, then drafts title, bullet, and A+ Content rewrites for the missing ones. See the Amazon solution and the Rufus Score docs.

The slow way: read the papers, map the graph by hand

The manual method is genuinely available, because Amazon published the research. Read the COSMO paper (SIGMOD 2024): roughly 29 million edges across 6.3 million nodes, 18 product categories, 15 relation types — function, audience, capability, purpose, location, complement, subjective need, and the rest. Then audit each listing against each relation: does the title state the function? Do the bullets state who it is for and where it is used? Does A+ Content cover the lifestyle relations?

For one hero ASIN, this is a productive afternoon. For a 300-listing catalog it is weeks, and measurement is worse: Rufus exposes no API, so checking citation means running 10–20 buyer prompts against Rufus in a logged-out session every week and tallying by hand. A stable trend takes four to six weeks of consistent manual runs — exactly the kind of recurring work that quietly stops happening.


The eCommerce Insights way

  1. Connect the Amazon channel. Add your Amazon SKUs. The channel-aware engine routes them to Rufus/COSMO scoring automatically — Shopify SKUs go to the six-engine scan, Amazon SKUs to the graph. Details on eCommerce Insights for Amazon.
  2. Read the per-relation score. Each listing is scored against the 15 COSMO relations, with the missing ones named per SKU. The leverage is asymmetric: the first six (function, audience, capability, purpose, type, category) cover roughly 70% of typical Rufus query intents in eCommerce Insights's read of the published evaluation data.
  3. Fix the title and first two bullets first. "Premium Stainless Steel Travel Mug" surfaces nothing the graph can traverse. "16oz Insulated Travel Mug for Coffee on Long Commutes — Leakproof, Fits Standard Car Cupholders, Keeps Hot 8 Hours" surfaces four relations. Rufus also weights the first two bullets more heavily than the last three, per Amazon Science's evaluation — function in bullet one, quantified capability in bullet two.
  4. Apply the rewrite diffs. eCommerce Insights drafts the rewrites — title, bullets, backend attributes, A+ Content suggestions — as reviewable diffs in the listing-rewrite workflow. A human approves every change before anything ships to Seller Central.
  5. Re-score and watch the window. Rerun the Rufus score after Amazon indexes the edits; expect citation movement in Brand Analytics search-term data over six to twelve weeks, not days.

Reviews deserve their own line: Amazon's 2024 shopping-agent research shows Rufus synthesizes review language for subjective queries ("a cozy blanket," "a thoughtful gift"). Review-collection campaigns that elicit experience language outperform another bullet rewrite for those queries. The mechanics are covered in the Amazon Rufus optimization guide.

What "good" looks like

Core relations (function, audience, capability, purpose, type, category) stated explicitly6 / 6
Relations surfaced in title + first two bullets4+
A+ Content covering lifestyle/persona relationspresent
Review base with subjective experience languagegrowing

A listing satisfying all 15 relations is rare and not the goal. The goal is no traversal dead-ends on the queries your category actually gets. If your D2C site shares the catalog, run the same products through the AEO Grader — the COSMO logic carries over to ChatGPT and Perplexity with different vehicles.

Ask AI about this job

Have your favorite AI engine apply this walkthrough to your listings.

Frequently asked questions

What is Amazon Rufus and how does it pick products?
Rufus is Amazon's customer-facing shopping assistant: an LLM with retrieval over Amazon's catalog, reviews, and Q&A. It picks products by querying COSMO, Amazon's published knowledge graph that maps shopper intent to products through 15 relation types mined from purchase behavior — roughly 29 million edges across 6.3 million nodes per the SIGMOD 2024 paper. A listing that never states a relation cannot be reached by the traversal.
Does eCommerce Insights actually score Amazon listings?
Yes. The channel-aware engine routes Amazon SKUs to Rufus/COSMO scoring across the 15 intent relations, with listing-specific rewrite recommendations for title, bullets, backend attributes, and A+ Content. Shopify SKUs route to the six-engine scan instead — every channel gets vendor-specific output. See the Rufus Score docs.
What is the single highest-leverage change for Rufus?
For listings that already state function and audience clearly: subjective review language. Amazon's 2024 research on shopping agents documents that Rufus weights review text heavily for subjective queries like "a cozy blanket." Reviews that describe subjective experience feed that matching; generic five-star "great product" reviews do not.
Do COSMO principles carry over to my Shopify store?
Largely yes. ChatGPT, Perplexity, and Google AI Overviews all reward listings that state function, audience, capability, and use case explicitly. The vehicles differ — Shopify uses body copy, Product JSON-LD, and metafields instead of backend attributes and A+ modules — but the targets are the same. See rewrite product titles for AI search for the D2C-side adaptation.
How long until listing changes affect Rufus answers?
Amazon indexes listing edits within days, but Rufus citation movement typically shows over a six-to-twelve-week window in Brand Analytics search-term data, per Amazon Science's published evaluation cadence. Treat single-week swings as noise.

Score your listings against the graph Rufus reads.

15 relations per listing, rewrites as diffs, human-approved. 14-day trial.