Optimizing Amazon Listings for Rufus and Alexa for Shopping: The 2026 AI Discovery Playbook

A shopper opens the Amazon app and types "quiet air purifier for a nursery that won't dry out the air." Three years ago that query was a mess for the algorithm: too many words, no clean keyword match, half of them irrelevant to the old index. Today Amazon reads it the way a person would. It infers a baby in the room, a preference for low noise, a worry about humidity, and it assembles an answer from the products whose listings actually address those things. The brands that get surfaced aren't the ones who stuffed "air purifier" into the title nine times. They're the ones whose catalog data told the machine what the product does, who it's for, and why it fits.

That shift has a name now. On May 13, 2026, Amazon retired the standalone Rufus brand and folded its shopping intelligence into a conversational agent called Alexa for Shopping, on by default for every signed-in U.S. customer right in the search bar. Rufus, now Alexa for Shopping, sits on top of a ranking stack most listings still aren't built for. Across our client portfolio, having managed more than $450M in Amazon revenue across 100+ brands, we've spent the last few months rebuilding listings for this stack, and the pattern is consistent: keyword-only optimization is running out of road.

The three-layer stack behind Amazon AI search

Amazon discovery in 2026 runs on three layers, and they stack rather than replace each other.

• A9 is the original keyword-matching engine. It still decides whether your listing enters the candidate pool at all. If a relevant term isn't indexed, you don't get considered. Keywords didn't die.
• COSMO is the semantic layer Amazon added on top of A9. It's a large language model that reads intent and context. A9 asks whether a listing contains the words the customer typed. COSMO asks whether the product solves the problem the customer described. Search "shoes for a wedding" and COSMO infers formal dress shoes, even when a listing never uses that exact phrase.
• Rufus, now Alexa for Shopping, is the generative layer the shopper actually talks to. It writes the answer, runs the side-by-side comparison inside the results page, and decides which products to cite.

Amazon hasn't published a full COSMO spec, so treat the internals as observed behavior rather than documented fact. The direction, though, isn't ambiguous. Conversion rate has overtaken raw sales volume as the signal the system weights most heavily, because a model optimizing for "did this answer the shopper" cares more about whether clicks turn into orders than about gross units alone. If your listing pulls traffic it can't convert, the new stack notices faster than the old one did.

This isn't the ChatGPT question. It's happening inside Amazon.

We wrote earlier about AI becoming an off-Amazon sales channel, the ChatGPT and Perplexity and Google AI Mode question. This is a different problem. Alexa for Shopping lives inside Amazon and draws almost entirely on Amazon-owned data: your listing copy, your A+ content, your reviews, community Q&A, pricing, availability, and your structured attribute fields. It's built on Amazon Bedrock using Claude and Amazon's own catalog models, and for most queries it reads your catalog, not the open web. That's good news for sellers. The inputs are things you control.

The scale is why this matters now and not next year. Rufus launched in 2024 and reached more than 300 million customers in 2025, with monthly users up over 115% year over year and engagement up roughly 400%. Amazon has said the assistant drove around $12B in incremental sales in the prior year. One practitioner estimate puts 15% to 20% of mobile queries now flowing through natural-language interpretation rather than literal keyword match. Treat that figure as directional and single-sourced, but the trend line is real. When a discovery surface reaches that many buyers, optimizing for it stops being optional.

What the AI layer actually reads: your attribute fields

The most underused lever on Amazon right now is sitting in the Seller Central attribute fields most brands leave half-empty. Subject keywords, intended use, target audience, material, item form, all the fields that were optional under A9 and quietly got skipped. That structured data is exactly what the AI layer reads to decide whether your product matches a described need. Filling it in isn't housekeeping anymore. It's discovery.

In our listing audits, the completeness gap is usually the fastest win. A brand doing $8M with 200 SKUs might have 30% of its attribute fields populated. Closing that gap hands COSMO and Alexa the raw material to recommend the product for queries the title never mentions. Nothing about the copy has to change for that first lift. You're just telling the machine what it couldn't previously see.

Writing bullets and A+ the machine can cite

Practitioner testing shows the assistant cites specific bullet content when it generates answers. That makes every bullet a potential AI-citable data point, which changes how you write them. Vague benefit language gives the model nothing to quote. Specific, verifiable attributes give it everything.

Compare two versions of the same bullet.

Keyword-era bullet	AI-era bullet
POWERFUL PURIFICATION: Our premium air purifier uses advanced technology for the cleanest air in your home	58 dB noise level, quieter than a normal conversation, with HEPA H13 filtration that captures 99.97% of pet dander and allergens, designed for rooms under 800 sq ft
PERFECT FOR ANY ROOM: Great for bedrooms, living rooms, and more	Sized for nurseries, bedrooms, and offices up to 800 sq ft; auto mode adjusts fan speed to air quality so it runs near-silent overnight

The right column answers questions a shopper actually asks: how loud, how clean, how big a room. Four verifiable attributes beat one adjective every time. COSMO also looks for cause-and-effect language, words like "designed for," "because," and "prevents," that signal genuine fit rather than keyword decoration. Write the way you'd answer a customer's question, not the way you'd cram a search box.

The same logic runs through A+ content. A+ that reads like a brochure gives the model marketing copy. A+ that reads like an informed answer, with comparison modules, use-case breakdowns, and real specifics, gives it something to cite. Rebuilding that across a full catalog, ASIN by ASIN, is the unglamorous core of Amazon listing optimization, and it's where most of the AI-readiness work actually lives.

Reviews, recency, and the signals you don't write

Not every input is copy you control directly. The AI layer weights review quality and, notably, review recency. A product with 4.6 stars and a steady stream of recent reviews reads as a safer recommendation than one with 4.7 stars whose last review landed eight months ago. Community Q&A feeds it too. This is where operational rigor shows up. Brands that run genuine post-purchase review generation and keep their Q&A answered give the model current, trustworthy signal. Brands that coasted on old review counts look stale to a system that reads freshness.

Be honest about the trade-off. You can't fabricate recency, and you shouldn't try. What you can build is the operational habit, the review requests, the Q&A monitoring, the fixing of product issues that generate bad reviews in the first place, that keeps the signal healthy over time. That's slow, compounding work, and it's exactly the kind of thing that separates a brand managed at scale from one running on autopilot.

Amazon PPC in the age of conversational search

Discovery changes don't stop at organic results. If shoppers search in full questions, exact-match-only campaign structures leave that demand on the table. Broad match and auto campaigns turn into intent-discovery engines, surfacing the conversational themes your exact keywords never captured. Your backend search terms should carry question-based phrases, "best organic dog food for puppies," not just the head term "organic dog food."

There's a compounding effect worth understanding. Listings that genuinely answer shopper questions earn better relevance, and better relevance tends to mean lower CPCs and stronger sponsored placement in AI-mediated results. Amazon CPCs averaged around $1.12 in 2025 and are projected toward $1.18 to $1.25 in 2026, with the usual Q4 spike. Those are directional figures, and we'd sanity-check them against your own managed-spend numbers before acting on them. The point holds regardless: listing quality and ad efficiency have become the same problem, which is why we wire campaign structure to listing quality inside Amazon advertising and PPC.

The AI-readiness checklist we run

This is the short version of the audit we run across a portfolio before touching a single bid.

Area	What to check	Why the AI layer cares
Attribute fields	Are subject keywords, intended use, target audience, and material filled on every ASIN?	Structured attributes are the cleanest signal COSMO reads to match described needs
Bullet specificity	Do bullets lead with verifiable numbers and use-cases, not adjectives?	The assistant cites specific bullet content in its generated answers
Cause-and-effect copy	Do you use "designed for," "because," and "prevents" to state real fit?	COSMO reads logical anchors to judge genuine relevance
A+ content	Does A+ read as informative answers rather than a brochure?	Informational A+ gives the model something to cite
Review recency	Is there a steady stream of recent reviews, not just a high old count?	The AI layer weights freshness, not just the star rating
Backend keywords	Do search terms include question-based phrases?	Conversational queries flow through natural-language interpretation
Campaign structure	Are broad and auto campaigns running as intent-discovery, not exact-only?	They surface conversational demand that exact match misses

Where to start

Pick your ten highest-revenue ASINs and run one pass this week: fill every attribute field, rewrite the top three bullets to lead with verifiable specifics, and check when your last review actually landed. That's a day of work that tells the AI layer what your products are for, and it usually moves impressions before it moves anything else. If you'd rather a team do it across the whole catalog and tune the ad strategy to match, that's what full-service management is built for. When you're ready, let's talk.