What shoppers ask Amazon's AI about your products

Somewhere in your Amazon Ads reporting there's a file most sellers have never opened. It contains the actual sentences shoppers typed into Amazon's AI assistant right before your ad appeared. Not keywords. Not search terms. The questions.

"is the handle real wood or laminate"

"will this fit a 15-inch shelf"

"why is this one more expensive"

Those three are ones we wrote as examples, because we don't publish anyone's real shopper data. But that's the shape of it. For the first time, Amazon is showing advertisers the questions people ask before they buy, in the words they actually used.

You are almost certainly already collecting this. Of 600 US advertiser profiles we pull this report for, 334 are serving ads against Alexa for Shopping prompts. None of them opted in. Amazon auto-enrolled eligible campaigns and bills the clicks against budgets you already set.

Last updated: August 23, 2026

Why this is different from search terms

A search term tells you what category someone was browsing. A prompt tells you what they needed to know before they'd commit.

The most common ways a prompt begins, in order of volume: does, why, is, can, will, how. Does leads by a wide margin. Read that list again, because it is the whole story. These are verification words. This is not a shopper discovering that wireless earbuds exist. It is a shopper deciding whether yours clear a specific bar, and telling you exactly which bar.

They also write in sentences. Prompts run about six or seven words, against the two or three people type into the search box. Across the 21 brands we measured the median was 6.36 words, and the whole cohort sat between 5.84 and 7.47. Almost nothing else in this data is that consistent.

Not all of them are questions, though. Depending on the catalogue, anywhere from 20% to 75% arrive with no question mark at all, often as a descriptive phrase with an attribute attached. A fair number open with a brand name, sometimes a competitor's. That is intent keyword targeting rarely catches.

The twelve things shoppers want to know

Nearly everything we see falls into twelve buckets. The examples below are written by us, not lifted from anyone's report, but the categories are real:

• Compatibility and fit: will this work with a 15-inch shelf?
• Installation and setup: how hard is this to mount on drywall?
• Durability over time: does the finish chip after a year?
• Capacity and dimensions: how many cups does it hold?
• Materials and composition: is the handle real wood or laminate?
• Care and cleaning: is it dishwasher safe?
• Safety and compliance: is this BPA free?
• Variation and customization: does it come in a darker colour?
• Included contents: does the box include the mounting hardware?
• Comparison against alternatives: how is this different from the cheaper version?
• Value and brand justification: why is this one more expensive?
• Use case and scenario: is this good for a small apartment?

The mix shifts hard by category. A phone-accessory catalogue is dominated by compatibility. A homewares one skews to care and cleaning. Your own split is the useful thing, and it is sitting in your account right now.

Here is the part that should make you want to look: every one of those questions is answerable in your listing. Not with a bid. With a sentence in your bullets or a value in a structured attribute. If shoppers keep asking whether it's dishwasher safe, and your listing never says, you are paying for the impression and losing the sale to the ambiguity.

The trap that will waste your first afternoon

Amazon generates many phrasings of the same underlying question, and each phrasing is a separate row with its own metrics. Sort your prompts by impressions and the top of the list is noise, because one real question is scattered across a dozen near-identical entries.

The obvious fix does not work. Lowercase everything, strip the punctuation, collapse on content words, and you will merge about 5.6% of them. The fragmentation is semantic, not typographic. Two more examples we wrote: will this survive a dishwasher cycle and is it safe to put in the machine. Same question. Almost no shared words.

Worse, the cleanup looks like it worked. The count drops, you tick the box, and you are still staring at one question spread across a dozen buckets. Real clustering needs embeddings, or an assistant that can read the list and group it for you.

What it's actually worth

Small, and worth knowing before someone sells you otherwise.

Across 21 brands, the median took 0.32% of its Sponsored Products spend on prompts, with the middle half between 0.20% and 0.51%. That is a share of spend you are already making, not new spend. The clicks are already inside your Sponsored Products campaign totals, reported at a finer grain, so adding prompt cost on top double-counts it and treating it as incremental will quietly corrupt your ACOS and TACoS.

The traffic also converts worse than you would expect. Comparing each brand against its own Sponsored Products baseline over the same dates:

Conversion came in below parity for 19 of the 21 brands. Click-through was a coin flip, lower for only 12 of 21, so anyone quoting you a dramatic click-through multiple is reading a single account. Cost per click is near parity.

So this is not a high-intent goldmine, whatever you have been told. It is a small, oddly-priced slice of your existing spend that happens to generate the single most useful qualitative dataset Amazon has ever handed advertisers. The money is not the point. The questions are.

What to do this week

Pull your own prompts and read them. The volume is manageable, hundreds of distinct prompts for a typical profile rather than millions. Group them by intent rather than by impressions, since impression rank is distorted by the duplicate-phrasing problem above.

Then do the one pass that pays for itself: find the questions your listing does not answer. A question that keeps coming up about a spec that appears nowhere in your bullets or structured attributes is a gap you can close this afternoon. Put the answer in a structured attribute where you can, not just the description, because assistants read attributes more reliably than prose.

What you should not do is rebuild your media plan around this. At a third of a percent of spend it will not move your budget. It should move your copy.

See your own

We sync this report automatically for US accounts as soon as you connect Amazon Ads, and you read it in plain language through Claude, ChatGPT, or any MCP client. Ask "what are shoppers asking that my ads serve against, and which of those questions does my listing not answer?" and you get the answer instead of the spreadsheet.

Here's how the Prompts report works, or connect Amazon Ads free.

Method, and the fine print

Everything here is aggregated to the brand level or coarser. No verbatim shopper prompt text appears anywhere in this study, no individual brand is identifiable, and no brand's spend is published. Every example prompt was written by us and labelled as such, because the alternative is publishing one customer's shopper questions to their competitors.

Coverage numbers cover every US profile we pulled the report for between May 17 and August 21, 2026. The two streams are requested separately so the denominators differ slightly: 600 profiles for Sponsored Products and 601 for Sponsored Brands, giving 306,501 Sponsored Products prompt rows and 10,415 Sponsored Brands rows. Performance and language figures come from a cohort of 21 brands with at least 50 prompt clicks each, each compared against its own Sponsored Products baseline over identical dates. Brands below that threshold swing on a handful of clicks. Profiles with no brand assigned are excluded.

Two things we will not claim. There is no placement dimension in this report, so nobody can separate a detail-page assistant impression from a standalone chat one. And viewable impressions are zero on every row because the field is empty, not because the placement performed badly, so any viewability rate computed from it is meaningless.

One honest note on stability. We first ran this on four accounts, where click-through looked 2.7x lower than the baseline. Widening to eight accounts and then re-cutting by brand took that to 0.98x, which is no effect at all. Conversion is the finding that survived every widening and got stronger each time. Per-brand rates on this surface are noisy, and a single account will tell you a clean story that does not replicate.

We are publishing the category list without percentage shares, because our classifier leaves a large uncategorised remainder and any share we quoted would move every time we adjusted a rule.

Frequently asked questions

Is prompt spend additional to my ad spend? No. It is already inside your Sponsored Products totals, the same billed click at a finer grain. Never add it on top, and never treat it as incremental in ACOS or TACoS.

Which marketplaces have prompt data? The US only. Every profile with data here was a US profile. We probed Canada, the UK, Germany, and Australia and got nothing back.

Can I opt out of Alexa for Shopping placements? Not as a separate control. Eligible Sponsored Products and Sponsored Brands campaigns were auto-enrolled and bill against budgets you already set.

How far back does the data go? Our window opens May 17, 2026, and Amazon keeps roughly 90 days, so any "since launch" trend claim, ours included, is bounded by that.

Is Alexa for Shopping the same as Rufus? Yes. Rufus was the name before the assistant was folded into Alexa branding. The ad reporting covers the same conversational surface.

Spotted a problem with the methodology? Email info@marketplaceadpros.com.