📋 Overview
You ask an AI shopping assistant the exact question your best customer would ask, and it recommends three products that are not yours. That outcome is diagnosable, but the fix depends entirely on whether the assistant never had your listing in front of it or had it and could not use it. This article gives you a two-test method for telling those apart, a short list of causes to check once you know which one you are facing, and the misreads that send sellers rewriting copy that was never the problem.
🔬 Split the Problem Before You Change Anything
An assistant has to do two separate things to recommend you: find your product among the candidates it considers, and then confirm from your content that you meet the shopper’s constraints. A failure at either step looks identical from the outside — you simply are not named. Running both tests below tells you which step broke.
1. Run the open query the way a shopper would phrase it
Write 8 to 12 questions as full sentences with real constraints in them, not keyword strings. Pull the phrasing from sources you already have: buyer questions on your detail pages, return reasons, and the query text in Search Query Performance under Brand Analytics if you are brand registered.
- Run each question in the assistants your buyers actually use, including Amazon’s own on-site shopping assistant (Alexa for Shopping) and general-purpose chat tools that browse the web.
- Use a fresh session for each run, and run every question at least twice. Answers vary between sessions and over time, so a single run is a sample, not a result.
- Record four things per run: were you named, which competitors were named, which specific attributes the answer cited as reasons, and whether the answer hedged (“if you need X, check the specs”).
There is no impressions report for AI answers. This manual sampling is the measurement, which is why keeping the question list fixed matters more than making it long.
2. Run the same question against your listing text alone
Copy your title, bullets, description, and the key attribute values into a chat tool and give it a closed instruction:
“Using only the product text below, answer this shopper question: [question]. If the text does not state the answer, reply exactly ‘not stated.’ If it does, quote the sentence you used. Do not use outside knowledge.”
This removes retrieval from the equation. If the tool answers correctly and quotes a real sentence, your content can support the recommendation and the problem is upstream — you were never in the candidate set. If it replies “not stated” or quotes something that does not actually answer the question, you have a content problem and you now know exactly which fact is missing.
3. Treat the assistant’s explanation as a hypothesis, not a report
Asking “why did you choose those three?” produces a fluent explanation that the model generates after the fact. It may name real attributes, but it is not a log of what drove the answer. Use it the way you would use a hallway comment from a shopper: a list of attributes worth checking against your own listing, not evidence of what happened.
🧭 Five Causes, and How Each One Shows Up
Match what your two tests produced to the patterns below. Most listings have one dominant cause, not five.
You were never in the candidate set
Signature: the grounded test passes cleanly, but you are absent from every open run, and the products named are the ones that already rank for the keyword version of your question. Rewriting bullets will not move this. Check whether you get impressions at all for the underlying query in Search Query Performance (use the ASIN View to see your own product’s queries), and confirm your product is assigned to the right product type with its core attributes filled in. A product filed under the wrong type can be invisible to a constraint-based question even when the copy is excellent.
The deciding fact is nowhere a machine can read it
Signature: the grounded test returns “not stated” for the one constraint in the question. This is the most common and most fixable cause. The fact usually exists — it is in a secondary image, an infographic, or implied by the product name — but it is not written as plain text and not entered as a structured attribute.
Put it in both places: the attribute field that exists for it, and one bullet that states it as a complete sentence with the number, the unit, and the scope together. Bullet length is capped per bullet and varies by category, commonly 255 characters for standard sellers and up to 500 for brand-registered sellers, so spend that space on the fact rather than on adjectives.
Your listing states it, then contradicts itself
Signature: the grounded test answers, but hedges, or the open answers mention you with a caveat. A title that says one capacity, a bullet that says another in different units, and a blank attribute field give an assistant three sources that disagree. Reconcile the title, bullets, description, and attributes to one value per fact before you write anything new.
The competitor is better corroborated than you
Signature: both tests pass, you are occasionally named, and the competitors that beat you have substantially more reviews, a higher rating, or review text that repeats the exact use case in the question. Assistants weight what buyers and the wider web say about a product, not only what the seller says. You cannot write your way out of this one.
The compliant levers are slow and ordinary: ask every eligible buyer for a review through Amazon’s own request mechanism, which is available from the Order Details page for orders inside the 5-to-30-day post-delivery eligibility window, and set expectations in the listing so the reviews you earn describe the use case accurately. Incentivized reviews, review gating, and review exchanges are prohibited and carry account-level risk; there is no version of them that is worth considering here.
The recommendation is simply correct
Signature: the constraint in the question is one your product genuinely does not meet. A diagnosis that ends here is still useful. You now choose between dropping that query from your target set, or closing the gap with a variant, a bundle, or a spec change at the next production run. What you must not do is add the claim to the listing anyway — inaccurate claims produce returns and negative reviews, and those feed the same evidence pool the assistant reads next time.
💡 Pro Tip: Freeze your question list and re-run the identical wording after each change. Because answers vary run to run, the only comparison that means anything is the same question asked the same way before and after — a new, better-phrased question will produce a different answer for reasons that have nothing to do with your fix.
⚠️ Misreads That Send You Fixing the Wrong Thing
Rewriting the listing after one bad answer
Sellers do this because a single answer feels like a verdict. It is closer to one shopper’s impression on one day. Before you touch copy, run the question at least twice in fresh sessions and run the grounded test. If the grounded test passes, your copy was not the failure point and the rewrite will cost you a known-good listing.
Pasting the shopper’s question verbatim into your listing
When an assistant names a competitor, the tempting move is to insert the full conversational query into bullets or backend search terms. Backend search terms are a limited field — 249 bytes in the US marketplace, with anything past the limit ignored — and stuffing them with repeated phrasing wastes the space and conflicts with Amazon’s rules for that field. Write the answer to the question as a plain factual statement in the bullet, and keep the backend field for terms the front end does not already cover.
Fixing the bullet and leaving the attribute blank
Front-end copy and structured attributes are read differently and both get used. A fact that lives only in a bullet can be missed by a filter-style question; a fact that lives only in an attribute may not be quoted back in a conversational answer. When you add a deciding fact, add it in both places and keep the values identical.
❓ FAQs
Does Amazon report how often an AI assistant shows my product?
There is no impressions-style report for AI answers. The closest signals you control are query-level impression and click data in Search Query Performance and session and conversion trends in Business Reports. Treat your manual question list as the direct measurement and those reports as supporting context.
Can I do this diagnosis without Brand Registry?
Yes. Both tests work for any seller. Brand Registry mainly affects which supporting data you can pull — Search Query Performance is a Brand Analytics report available to brand-registered sellers — and which listing elements you can edit freely.
How often should I re-run the question set?
After any listing change you want to evaluate, after a meaningful shift in your review profile, and on a steady cadence beyond that. Assistants update their models and their retrieval behavior without notice, so a result from several months ago is not a baseline you should trust.
Should I use a tool that checks AI answers at scale for me?
Possibly, but check two things first. Confirm on the vendor’s own documentation how it collects the answers, since automated collection that violates a platform’s terms creates risk you inherit. And confirm it records the full answer text rather than just a yes/no mention, because the attributes the answer cites are the part that tells you what to fix. Capabilities in this category change quickly, so verify current behavior rather than relying on a comparison you read somewhere.