Rufus doesn't rank your A+ content. It filters your structured attributes first, then ranks the survivors. Most CPG catalogues on Amazon EU were built for a human reader scrolling on a tab. They are 30 to 45 percent incomplete on the fields that gate the AI shelf. The fix isn't a content refresh. It's a 2-year catalogue architecture project that touches PIM, regulatory, brand, and ecommerce ops simultaneously. The brands that start in Q3 will compound. The brands that wait will be invisible to a discovery surface that already mediates 38 percent of Amazon's sessions.
A shopper opens the Amazon app on a Tuesday night.
They type: "best fragrance-free moisturiser for sensitive skin under 30 euros that's vegan."
Rufus returns five products.
Three are from indie skincare brands that did under 10 million euros in revenue last year. One is an Amazon-acquired private label. One is a Korean brand most European CPGs haven't heard of.
The global beauty brand selling the same product, with the same ingredients, the same price point, and a multi-million-euro annual A+ content investment behind it, doesn't appear.
Not because of the price. Not because of the reviews. Because the backend attribute field marked fragrance-free was empty.
That's not a content problem. That's a catalogue problem. And it's structural.
Rufus doesn't read your A+ content. It filters your attributes.
Most teams I talk to still believe Rufus is a smarter version of A9. It isn't. It runs a different mechanic, and the difference is where the budget is now misallocated.
Amazon's own engineering documentation describes Rufus as a retrieval-augmented generation (RAG) system on Amazon Bedrock, orchestrating multiple LLMs — Anthropic's Claude Sonnet, Amazon Nova, and a custom Amazon shopping model — selected per query by a router. Underneath sits COSMO, Amazon's Common Sense Knowledge Graph, mining query-purchase pairs and co-purchase pairs to build relationships like capable_of, used_for, suitable_for. COSMO has shown a 60 percent improvement in search relevance and a 0.7 percent live sales uplift on 10 percent of US traffic.
The architecture matters because of the order of operations. Agency consensus, validated against Amazon's published material, is that Rufus runs a two-stage retrieval: structured backend attributes filter the eligible product set first, often from millions to hundreds. Then unstructured content — titles, bullets, reviews, Q&A — is semantically ranked.
In plain language: if your target_audience, skin_type, formulation, dietary_preference, material_composition fields are missing or generic, Rufus eliminates your product before it reads a single bullet. Your A+ content investment is downstream of a filter your team didn't know existed.
And it gets worse. The text inside an HTML A+ module's text-only sub-blocks is readable. The lifestyle photography, the comparison-chart modules baked into PNGs, the brand-narrative imagery — none of it is documented as a Rufus retrieval source. Most CPG A+ budget goes to exactly those formats. Flywheel Digital, who has tracked Rufus since launch, observed it directly: image content appears absent from the training set, which makes written content the lever.
Atomicamz, building from the COSMO architecture, observes that products with 90 percent or higher completion of category-specific attributes show 2 to 3 times better Rufus visibility. Amazon Growth Lab estimates the catalogue contains over 750 data fields used for ranking and discovery, "yet most sellers optimize only the 10 to 20 visible fields and leave the rest incomplete."
That's the gap. And it's almost universal across CPG.
The 22% problem
If you want one number that captures how broken the current shelf-to-AI relationship is, it's this one.
Profitero and Mars United Commerce ran a controlled study of Rufus recommendations against organic search results in March 2026. The headline finding: only 22 percent of products on page 1 of Amazon search also appear in Rufus's responses. Thirty-six percent of products Rufus recommends do not appear on page one of search at all. And over half of audited brand content was wrong.
That's a different shelf. Not an extension of the old one. A new one, with different gatekeepers, populated by different brands.
Independent data from Amalytix (n=1,300+ products across 500 search terms) confirms the filter shape. Rufus functionally excludes products under 4 stars. Median rating of recommended products: 4.5. Median review count: roughly 2,991. 94 percent FBA. 92 percent Prime. 98 percent in-stock at retrieval time. Private-label bias is minimal — Amazon Basics is 0.6 percent of recommendations.
So Rufus isn't biased toward Amazon. It's biased toward complete catalogues with verified review depth. That's a meritocratic shelf, but only if you've done the structural work to be eligible.
For a shopper, the implication is invisible. They get five products, they pick one, they move on. For a brand with 1,500 ASINs on Amazon EU and a 2-million-euro media plan, the implication is that a third of your category is being discovered through a surface your investment doesn't reach.
The catalogue was built for the wrong reader
This is the part most CMOs miss. The catalogue you have today is the catalogue you built — and you built it for a reader that has been slowly displaced.
Walk through how a CPG listing actually gets created at scale. Brand team writes the hero narrative for global launch in English. Local marketing teams translate into German, French, Italian, Spanish. The digital shelf agency repackages narrative into Amazon listing format: title, five bullets, description, A+ modules. Regulatory reviews claims with a 2- to 4-week lag. The KAM team uploads via flat file. Backend attributes get filled by whoever uploads, often defaulted to "unspecified" or skipped entirely. A+ modules get built as image-rich, story-driven PNGs.
The chain breaks in three places.
Backend attributes are an afterthought. The team obsessing over the headline image is not the team filling target_audience, use_case, skin_type, formulation_type, material_composition. Those fields get whatever the PIM happened to have, which is often nothing.
PIM was built for syndication, not AI. Most CPG PIM systems were architected to push product data to retailer portals. Their schemas mirror retailer requirements from 2018, not Amazon's 2024 mandatory attribute push that added 274 required fields across 200 product types in August 2023.
There is no single owner of catalogue quality. Brand owns narrative. Regulatory owns claims. Digital owns A+ creative. Ecommerce ops owns upload. Nobody owns is this listing complete and AI-ready? That orphan accountability is the deepest root of the problem. AI-readiness is a horizontal problem in a vertical organisation.
The result: a catalogue that scrolls beautifully on a phone and is almost entirely invisible to the shelf compressing the consideration set to five products.
What's missing isn't content. It's structure.
Five specific gaps, in priority order, based on Rufus's documented retrieval mechanics.
Attribute completeness. Industry estimates from agency conferences put typical CPG attribute completion on Amazon EU at 55 to 70 percent. The missing 30 to 45 percent is concentrated in exactly the fields Rufus uses as a filter: intent attributes (use_case, target_audience, skin_concern), formulation specifics, structured ingredient and material data.
Claim structure. Marketing claims are written for human persuasion: emotional, narrative, vague. AI prefers atomised, machine-readable claims. "Reveals visibly younger-looking skin" is invisible. clinical_claim: "97% saw reduced fine lines in 28 days, n=42" is retrievable. The trend across beauty, supplements, and baby food is unmistakable, and CPG copy decks aren't being rewritten in this shape.
INCI and ingredient data. For beauty, INCI (International Nomenclature of Cosmetic Ingredients) is the single highest-leverage attribute set for AI-driven discovery. Most CPG beauty listings on Amazon EU upload INCI as an unstructured string in the description field. Rufus, Perplexity, and Google AI Mode are already matching beauty queries to ingredient-level signals. "best moisturiser without parabens for combination skin" cannot be answered from prose. Only from structured ingredient flags.
Use-case coverage. Tinuiti's 2026 analysis shows the average Rufus query is 2.4 times longer than a traditional Amazon search. The query patterns are: "best [product] for [persona]", "[product] suitable for [condition]", "[product] gift for [occasion]". CPG titles are written around brand + product + size + hero ingredient. They miss persona, condition, occasion entirely. Backend search terms field — 250 bytes, hidden, intent-rich — is empty on most CPG ASINs.
Localisation, not translation. Amazon EU treats DE, FR, IT, ES as separate marketplaces with separate Rufus indices. German Rufus surfaces products with logic-dense, technical specification language. French Rufus weights INCI compliance heavily. Spanish Rufus mixes native search with English brand terms. Translating the same English listing across four markets misses each one. Most CPG brands are doing exactly this because human localisation doesn't scale across 1,500-plus ASINs without structural investment.
Each gap is fixable on its own. Together, they require a system rebuild.
This is a 2-year project. Which is exactly why you start in Q3.
For a typical CPG brand on Amazon EU with a 1,000-ASIN portfolio, full catalogue remediation runs 14 to 20 months. The math doesn't compress.
Phase 1 alone — attribute audit, prioritisation by revenue cohort, top 200 SKUs remediated — is a 90-day exercise that requires a dedicated catalogue lead, regulatory liaison time, PIM administration, and an agency partner that has actually rebuilt for Amazon's 2024 attribute spec (most haven't). Phase 2 covers PIM schema upgrade, structured claim taxonomy, and the localisation layer for DE and FR. Phase 3 is long-tail remediation and cross-channel consistency: Schema.org JSON-LD on the DTC site, Google Merchant Center push, Perplexity Merchant Program enrolment, OpenAI product feed for ChatGPT Shopping.
A 500-ASIN portfolio compresses to 9 to 12 months. A 2,000-ASIN portfolio expands to 22 to 30 months, almost always with a parallel PIM migration that adds 6 to 12 more.
Cost shape, based on public agency rate cards and conference benchmarks: 150 to 350 thousand euros for 500 ASINs, 300 to 700 thousand euros for 1,000 ASINs, 700 thousand to 1.5 million euros for 2,000 ASINs. PIM rebuild adds another 250 thousand to 1 million euros depending on platform.
The 2-year framing is honest. And it's the reason this can't be a 2027 project. Every quarter you wait, AI systems train on your incompleteness as a signal that your product is less relevant. The gradient compounds against you. The agencies with deep Amazon EU attribute expertise are scarce, and they're booking 18-month engagements now. The brands that lock the best partners in Q3 2026 are setting the terms for what their category looks like in 2028.
This isn't an IT project. It's a competitive timing decision. The first-mover window is roughly 18 to 24 months. After that, AI-native catalogue is table stakes — and the brands that didn't build will be losing share with no clear lever to fix it quickly.
Five moves you can ship before the JBP
Most operators reading this can't fund a 2-year project tomorrow. That doesn't mean there's nothing to do. These five moves cost weeks, not quarters. They won't make you AI-native. They will defend share while the bigger project gets funded.
→ Populate the backend search terms field on top 100 ASINs with intent terms. The 250-byte hidden field most operators leave empty. Fill it with persona, condition, problem, occasion: sensitive skin, fragrance-free, vegan, gift for mum, for toddlers. Two days per market. Immediate Rufus candidate-set lift.
→ Rewrite the first bullet of the top 50 ASINs to lead with use case. Replace the brand-story opener with a direct intent statement: "Daily moisturiser for dry, sensitive skin prone to redness. Suitable for ages 25+. Fragrance-free, dermatologist-tested." Rufus reads first sentences harder than the rest of a bullet.
→ Audit and populate the 5 highest-leverage backend attributes per category. For beauty: target_audience, skin_type, skin_concern, formulation_type, vegan/cruelty_free. For food: dietary_preference, allergen, flavour_profile, meal_occasion, serving_size. Pull from PIM where possible. Don't wait for full PIM cleanup.
→ Mine your Q&A and answer 10 high-intent questions per top ASIN. Q&A is one of Rufus's most-cited sources. Most CPG brands answer defensively in legal language. Start answering with structured, specific, intent-matched language.
→ Convert top 20 A+ Content pages from image-only to HTML structured modules. If your A+ Content is one big JPG per module, Rufus can't read it. Rebuild in HTML with structured headings, body copy, and feature/benefit tables. Premium A+ if you have it.
These five together: 4 to 8 weeks of cross-functional effort. They sit in front of the JBP. They give the KAM team something concrete to put in the negotiation as evidence the brand is AI-ready. And they buy 12 months of share defence while the strategic project gets sponsored.
The question worth asking
Most brand teams are still asking how do we improve our Amazon performance?
The question that actually matters in 2026 is different. Are we structurally eligible for the surface that's choosing the next five products?
Different question. Different answer. The first produces optimisations on a shelf that's compressing. The second produces a catalogue that compounds.
Open the Amazon app. Ask Rufus for the best version of your hero product under 30 euros that fits the most-asked customer scenario in your category. If your brand isn't in the five, that's not a search problem. That's the next 24 months of your category share.
The catalogue you built for humans won't read itself for AI.
Build the right one.