Somewhere in the last two years, a new reader started going through your Amazon listing. It doesn't scan for keywords. It reads your bullets, your A+ content, your reviews and your Q&A the way a diligent human would, then it summarises you to shoppers in its own words, compares you to competitors you didn't choose, and decides whether you get mentioned at all.
That reader is Amazon's AI shopping layer, and in 2026 it stopped being an experiment. Rufus, Amazon's generative AI assistant, reached hundreds of millions of customers across the US, UK and Europe, and Amazon's own earnings commentary put the incremental annualised sales it drives in the billions of dollars. Then in May, Amazon retired the standalone Rufus experience in the US and folded its successor, Alexa for Shopping, directly into the main search bar, generating AI overviews above results and running product comparisons inside the results page itself. The direction of travel could not be clearer: AI is not a feature beside search. It is becoming the search.
Here's how we're adjusting listings across the accounts we run at rankhouse, and what we'd leave alone.
What actually changed about discovery
Traditional Amazon search is lexical. A shopper types "collagen powder", the engine matches indexed terms, and ranking is settled by relevance and sales performance. The AI layer is semantic. A shopper asks "what's a good collagen for skin that doesn't taste fishy and works in coffee", and the model reasons about intent, reads listing content and reviews for evidence, and assembles an answer. Products with clear, complete, verifiable content get cited. Products whose listings are a keyword salad give the model nothing to work with, and it quietly moves on.
| Classic A9/A10 search | AI assistant (Rufus, now Alexa for Shopping) | |
|---|---|---|
| How shoppers ask | Typed keywords | Natural-language questions and comparisons |
| What gets read | Title, bullets, backend terms | The whole detail page, plus reviews and Q&A |
| Results surfaced | A full grid of options | A handful of named recommendations |
| What wins | Keyword relevance and conversion history | Complete, specific, verifiable listing content |
| What changed in 2026 | Still running underneath | Renamed Alexa for Shopping in May 2026; same logic, wider reach |
The optimisation playbook, in priority order
1. Complete your structured attributes
The least glamorous job on Amazon is now one of the highest-leverage. Every backend attribute field, material, use case, certifications, dietary flags, dimensions, is a fact the AI can retrieve confidently. Across the industry, the consistent finding is that products with complete structured data outperform keyword-stuffed listings on AI surfaces. Audit every ASIN for empty fields. It's tedious, it's free, and it compounds.
2. Write bullets that answer questions
The AI is assembling answers to natural-language questions. Bullets that state a feature and land the benefit in plain English give it quotable material. "500mg marine collagen peptides per serving, unflavoured, dissolves fully in hot or cold drinks" answers three real questions in one line. "PREMIUM QUALITY BEST COLLAGEN SUPPLEMENT UK" answers none, and the era when that line did anything useful is closing.
3. Treat reviews and Q&A as content you influence
The AI reads reviews as evidence and cites them in its answers. That raises the value of review recency, specificity and volume, and it makes the questions section a surface worth managing: seed it with the genuine questions customers ask support, and answer them properly. The compliant way to build review velocity hasn't changed, and we've written it up separately.
4. Make A+ content informational, not decorative
A+ modules written as brand posters give the model nothing. A+ written as structured information, comparison tables, how-to-use, what's-in-it, does double duty: it converts humans and feeds the machine. This aligns with something we already knew from ingredient-led search behaviour: shoppers, human and artificial, are researching the substance, not the slogan.
5. Watch the Prompts data
Amazon has started exposing AI-placement data to advertisers, including a Prompts report for Sponsored Products, and auto and broad match campaigns are the ones eligible for AI-driven placements. Two practical implications: keep auto and broad campaigns alive even in mature accounts, and start baselining your AI-surface impressions now, while these placements are young. When Amazon starts charging properly for AI real estate, and it will, the brands with baseline data will know what it's worth.
Want this level of thinking inside your account?
rankhouse is a boutique specialist agency: experienced Amazon professionals, no juniors, led personally by founder Dan Whalley. Fixed fee from £1,500 + VAT a month. Every engagement starts with a free profit audit built from your own data.
Get the free audit at rankhouse.co.uk →What not to do
- Don't torch your keyword foundations. Traditional search still drives the majority of discovery. Semantic optimisation is additive, not a replacement. The listings that win do both.
- Don't chase the assistant's phrasing. Trying to reverse-engineer specific AI answers is a treadmill. Write the clearest true description of your product and its use cases; that's the durable strategy under any model.
- Don't panic-rewrite everything at once. Change listings the way we change everything: through controlled experiments, one variable at a time, so you know what worked.
The strategic read
Every previous era of Amazon optimisation had a trick to it: the keyword hacks, the giveaway launches, the review shortcuts. The AI era is unusual because its optimal strategy is simply honesty at high resolution. Complete data, clear writing, real evidence, genuine answers. The machine is grading your listing the way your most careful customer would, at scale, on every session.
That's good news for brands with real products and disciplined operators, and bad news for everyone coasting on 2019's tricks. We know which side we build for. If you'd like your catalogue read the way the AI reads it, gaps and all, that's part of what the free audit at rankhouse covers.
Questions we get asked about this
How much of Amazon shopping actually goes through the AI layer now?
Enough to matter and growing quickly, though precision varies by source. Amazon's own reporting put Rufus in front of hundreds of millions of customers, with usage up sharply year on year and its earnings commentary attributing billions of dollars in incremental annualised sales to it. Industry analyses through early 2026 consistently estimated a meaningful and climbing share of mobile queries being mediated by AI surfaces. The May transition to Alexa for Shopping in the US, embedded directly in the main search bar with AI overviews above results, removed the last reason to treat this as an experiment: it's a default layer over the primary shopping experience, and the UK rollout direction is not ambiguous.
Will optimising for AI search hurt my traditional keyword ranking?
No, because done properly they're the same work pointed at the same truth. Traditional indexing still needs your relevant terms present in the title, bullets and backend fields, and nothing about semantic optimisation removes them. What changes is the packaging: terms embedded in clean, factual, benefit-stated sentences serve both readers, while keyword strings without grammar serve only the old one, and increasingly annoy even that. The listings losing ground in the AI era aren't the well-written ones; they're the stuffed ones, whose tricks the semantic layer simply reads through. Write the clearest true description of the product and its uses, with the vocabulary shoppers actually search, and both systems reward it.
What are structured attributes and why do they suddenly matter so much?
They're the backend data fields describing your product in machine-readable form: material, ingredients, dietary certifications, dimensions, use cases, care instructions, dozens per category. Humans rarely see them directly, which is why they've been neglected for years. The AI layer retrieves them as verified facts when assembling answers, and a product whose attributes are complete can be confidently cited for queries its bullets never anticipated, while a product with empty fields simply can't be recommended for the thing it never declared. The audit is unglamorous: export the category template, fill every applicable field on every ASIN, and treat new-field announcements as standing work. It's free, it compounds, and it's currently one of the widest gaps between well-run and neglected accounts.
Should I be doing anything with the Prompts report?
Yes: baseline it now, while AI placements are young. The Prompts report for Sponsored Products shows where your ads surfaced against conversational queries, and auto and broad match campaigns are the ones eligible for those placements, which is a concrete reason to keep discovery campaigns alive even in mature accounts. Two practical uses today: reading the actual questions shoppers ask, which is listing-content research of a quality keyword tools can't provide, and establishing what AI-surface volume you receive while the real estate is cheap, so that when Amazon prices it properly, and its advertising history says it will, you'll know from your own data what it's worth paying.
Is there anything I should deliberately not change?
Three things. Don't dismantle keyword foundations that still drive the majority of discovery; semantic work is additive. Don't chase the assistant's current phrasing or try to reverse-engineer specific answers, because the models change and the treadmill never pays; the durable strategy is clarity and completeness, which every model version rewards. And don't rewrite the whole catalogue in one heroic weekend, because unversioned mass changes make results unattributable. Change listings the way everything should change: through controlled experiments, one variable at a time, winners banked and rolled across the family. The AI era punishes tricks, but it punishes thrash almost as much.
Amazon's AI shopping layer stopped being an experiment in 2026: Rufus reached hundreds of millions of shoppers, drove billions in incremental sales by Amazon's own earnings commentary, and in May the US moved to Alexa for Shopping, embedded directly in the main search bar with AI overviews above results. The layer reads listings semantically, like a careful human, and it punishes bad content with absence rather than low rank. The playbook in priority order: complete every structured attribute field, write bullets that answer real questions in plain English, treat reviews and Q&A as evidence the AI cites, rebuild A+ as information rather than decoration, and baseline your Prompts report data while AI placements are still cheap, keeping auto and broad campaigns alive because they're the ones eligible. Don't torch keyword foundations, don't chase the assistant's phrasing, and don't mass-rewrite without experiments. The optimal AI strategy is simply honesty at high resolution, which suits disciplined operators and nobody else.
The next step is twenty minutes.
If any of this reads like your account, the fastest way to find out is the free audit: per-product profitability, where the ad spend is leaking, and what we would fix first. No pitch deck, no obligation.
Book a chat with Dan →Sources
- Alexa for Shopping: optimisation strategies for 2026 — Tinuiti. https://tinuiti.com/blog/amazon/alexa-for-shopping/
- Alexa for Shopping (Amazon Rufus): the complete guide for brands and sellers — Perpetua. https://perpetua.io/blog-alexa-for-shopping-amazon-rufus-the-complete-guide-for-brands-and-sellers/
- Amazon Rufus guide 2026 — Amalytix. https://www.amalytix.com/en/knowledge/ai/amazon-rufus-guide-2026/
About the author
Daniel Whalley is the founder of rankhouse, a boutique specialist agency for Amazon-focused growth in FMCG, health, wellness and beauty brands. He has spent 10 years inside Amazon accounts, generating £100M+ for the brands he works with, and manages £500k+ a month in ad spend across the UK, Europe and the US. He writes from inside the accounts he runs, not from the sidelines. Connect on LinkedIn → · amazon@rankhouse.co.uk