Amazon Rufus optimization: the COSMO playbook

What Rufus is

Rufus is Amazon's AI shopping assistant, embedded in the Amazon app and site. Shoppers ask it comparison questions, use-case questions ("will this fan work in a 400-square-foot room?"), and compatibility questions, and it answers conversationally — drawing on listing content, A+ Content, reviews, Q&A, and Amazon's product knowledge graph. For a brand selling on Amazon, Rufus is a second discovery surface running on different machinery than the A9 ranking algorithm, the same way ChatGPT runs on different machinery than Google. The cross-channel context is in the product AI visibility pillar; this guide is the Amazon-specific layer.

COSMO: the graph behind the assistant

Amazon has documented the commonsense layer in published research: COSMO, a large-scale e-commerce common sense knowledge generation and serving system. COSMO mines commonsense relations between products and intents — that a camping mat is used for winter camping, suits side sleepers, pairs with a sleeping bag — so the shopping experience can answer intent-shaped questions instead of just matching keywords. The practical reading for a seller: Rufus can only connect your product to "winter camping" if something in the listing, reviews, or Q&A actually says so in concrete language. Implicit positioning that a human infers from a lifestyle photo is invisible to the graph.

A9 matched keywords. Rufus traverses relations. The listing has to say the things the graph needs to connect.

Rufus vs A9: what changed

A9 rewarded keyword coverage, conversion history, and backend keyword hygiene. Those still matter for classic search placement. Rufus adds a retrieval layer that weights different signals: structured bullets that quote concrete specs, A+ Content depth, Q&A coverage, review language that spans multiple use cases, and the relations between SKUs — substitutes, complements, audiences. A keyword-stuffed title that performed for A9 reads as noise to a conversational assistant. The shift mirrors what happened in web search — the same shift covered in AI search optimization — applied to a closed marketplace.

The 15 relations worth scoring

eCommerce Insights scores Amazon SKUs through a separate channel-aware model built on the relation set Amazon publicly documents. Each relation maps to a question Rufus might ask of a product node:

#	Relation	What it captures
1	A+ content depth	Brand-story modules, comparison charts, image-with-text blocks
2	Structured bullets	Bullets quote concrete specs in a consistent format
3	Brand-store consistency	SKU listed in the brand store with consistent imagery and copy
4	Q&A coverage	Eight or more shopper questions answered on the listing
5	Review keyword variance	Reviews cover multiple use cases, not one repeated frame
6	Substitute relation	Credible substitutes surfaced (frequently-bought-with patterns)
7	Complementary relation	Complements surfaced — bundles, related accessories
8	Use-case relation	Listing maps to one or more named use cases
9	Audience relation	Listing maps to named audience segments
10	Attribute relation	Structured attributes (size, color, material) filled
11	Category relation	Listing lives in the right category node, not a generic fallback
12	Brand-tier relation	Brand positioned at a consistent tier (premium / mid / value)
13	Price-tier relation	Price sits inside the expected band for the category
14	Seasonality relation	Seasonal signals where appropriate (gifts, summer use)
15	Compatibility relation	Compatible products, platforms, and standards named

The relation set is what Amazon publicly documents; the production model may weight other signals, and the set deserves quarterly revalidation. Composite weights in eCommerce Insights' Rufus Score favor the relations that most predict Rufus citation in internal prompt testing — methodology in the Rufus Score docs.

The four kinds of Seller Central work

1. A+ Content edits. Block-level rewrites: comparison charts that name the relations explicitly, image-with-text modules that state use cases in words, brand-story modules consistent with the brand store.
2. Bullet rewrites. Five bullets, structured pattern — attribute, spec, use-case qualifier. "660 CFM whole-room circulation for spaces up to 400 sq ft" beats "Powerful airflow."
3. Brand-store updates. Module-level copy consistency, especially where several SKUs share a brand-store hub — tier and imagery mismatches degrade the brand-tier relation.
4. Q&A submissions. Proposed questions and answers that fill coverage gaps — compatibility, sizing, and use-case questions first, since those are the shapes Rufus gets asked.

Review prompts are the fifth, slower lever: asking "what were you using this for, and how did it hold up?" produces the varied use-case language that relation 5 measures. Note Seller Central work is applied manually — eCommerce Insights is read-only against Seller Central and outputs recommendations, per the Amazon solution page.

Measuring Rufus visibility

The measurement discipline is the same as the six web engines: per-SKU (per-ASIN), per-intent, on a weekly cadence. eCommerce Insights routes Amazon URLs to the Rufus Score automatically — each of the 15 relations scored 0–100 with the observation and a one-sentence recommendation, rolled into a composite with deltas per run. Multi-channel brands get the web-engine citation score and the Rufus Score side by side, which is the point of channel-aware SKU tracking: one ledger, channel-correct scoring.

Limits and unknowns

• Amazon does not publish Rufus's retrieval internals; the relation set is inferred from published research and observed behavior, revalidated quarterly.
• Sponsored placements are a separate system — everything here is organic-citation-focused.
• Rufus behavior shifts with Amazon releases; treat specifics as mid-2026 observations.
• Walmart's Sparky assistant is the nearest analog on another marketplace — secondary for most D2C brands, covered on the Sparky page.