Keyword research on Amazon has always meant phrases: what shoppers type, how often, and which of those phrases a listing can win. A growing share of shoppers now do something else. They ask. "Are these durable?" "What do I need for cold weather golf?" "Best dinosaur toys for a 5-year-old." Amazon's shopping assistant answers, and it answers from the listing.
That changes what the research has to produce. The phrases still matter, because search still runs on them. But a listing also has to answer the questions its buyers ask, in words the assistant can find, or it is left out of the answer. This is what Amazon has said about how the assistant works, what its ads now do, and how to research the questions alongside the keywords.
Amazon introduced Rufus as an assistant "trained on Amazon's extensive product catalog, customer reviews, community Q&As, and information from across the web." On a product page, it answers "specific questions about individual products" using "listing details, customer reviews, and community Q&As." Those three are the whole of what a seller can influence directly: what the listing says, what customers say in reviews, and what the questions and answers on the page say.
This year Amazon folded Rufus into Alexa for Shopping, which it describes as combining Rufus's product knowledge with Alexa's personal knowledge of the customer. Amazon says Rufus "helped over 300 million customers in 2025." Alexa for Shopping is available in the US on the Amazon Shopping app, the website and Echo Show devices, and it can compare products, make personalised recommendations and track prices.
On 25 March 2026 Amazon made Sponsored Products prompts and Sponsored Brands prompts generally available in the US. Amazon describes them as an "AI-powered enhancement to your existing campaigns that automatically engages shoppers with relevant product information," shown "in shopping results and product detail pages on Amazon," and able to "open a dialog in Rufus or respond to the customer directly on the page."
Three details matter for a seller. Campaigns are "automatically enrolled in prompts," so they are probably running already. The prompts draw on "Amazon's first-party signals from your detail pages, Brand Store, campaign data, and more," so a thin listing produces thin prompts. And they are charged within the campaign's cost per click, and can be reviewed and paused in the Ads Console. Reading the prompts Amazon has generated is one of the few direct views a seller gets of what Amazon's systems think the listing says.
Because campaigns are enrolled in prompts automatically, most advertisers have prompts running that nobody at the brand has read. They are worth reading, as a check on the ads and as a view of how Amazon's systems understand the listing.
Find them. The prompts generated for a campaign can be reviewed in the Ads Console, where they can also be paused. Start with the campaigns that spend most and the products that sell most.
Check them for accuracy. Read each prompt as a shopper would. Does it describe the product correctly? Does it suggest a use, a benefit or a comparison the brand would not claim? A prompt that overstates is the brand's advertisement saying something untrue.
Look for what is missing. If the prompts never mention the feature that sells the product, the listing probably does not state it clearly in words. The fix is on the listing, not in the advertising.
Pause what is wrong, fix what caused it. Pause a prompt that misdescribes the product, then correct the listing details it was drawn from. Prompts are generated from the detail page, the Brand Store and campaign data, so a corrected listing is the lasting fix.
Note what Amazon thinks the product is for. The uses and audiences the prompts name are a reading of the listing by Amazon's own systems. Where they differ from who actually buys, the listing is telling Amazon a different story from the one the brand intends.
Read them again after a change. After the listing is rewritten, check whether the prompts change with it. It is one of the few direct ways to see whether the rewrite was understood.
Keep the phrase research. Search terms still decide where a listing appears in ordinary results, and
Brand Analytics and
the search terms report are still where that demand is measured. What changes is that phrases are no longer the whole of it.
Add the questions. Collect them from the places buyers already ask: the product's own Q&A, its reviews, the competitors' reviews and Q&A, and the messages customer service answers every week. Group them by what they are really asking about: size, fit, compatibility, durability, what is in the box, who it suits, how it compares with the obvious alternative. For an insulated water bottle, the groups might be how long it keeps a drink cold, whether it fits a car cup holder, whether it can go in the dishwasher, whether it leaks in a bag, and what the lid is made of. Each is a fact the listing can state, and each is a question somebody will ask.
Then map each group to the part of the listing that should answer it. The question "will this fit a 2019 model" needs the compatibility stated plainly in the bullets, not implied by a photograph. "Is it safe for a toddler" needs the age range and the material in words. An answer that exists only in an image, or only in a customer review from two years ago, is an answer the listing is leaving to chance.
Facts in words, where the assistant can read them. Measurements with units, materials, what is included, compatibility, the age or use it is for, care instructions. The title has limits, so most of this belongs in the bullets, the description and the item attributes;
the 75 character title limit covers what the title can still carry. The backend fields matter less for questions than for phrases, and
backend search terms sets out what they do.
Answer the recurring questions on the page. If buyers keep asking the same thing in the Q&A, the listing is missing a sentence. Put it in the bullets, and the question stops needing an answer from a stranger.
Treat the reviews as part of the listing. The assistant reads them, and a product whose reviews keep mentioning a flaw will have that flaw in its answers. That is a product or packaging problem to fix, not a keyword problem to write around.
Take a cordless vacuum. Keyword research for a search box finds cordless vacuum, stick vacuum and lightweight vacuum. Research for an assistant starts from the questions people ask about it, which the listing's own reviews, its customer questions and the long search terms in the search query report all show: is it good for pet hair, how long does the battery last, will it clean stairs, how loud is it.
Each question gets a stated answer. A bullet says the run time on the standard setting and on the strongest, in minutes. Another says it handles pet hair on carpet and hard floors, and why. The attributes carry the weight and the noise level as numbers, so they can be read without guessing.
Nothing is left for the assistant to infer. A listing that says long lasting battery gives it nothing to quote; one that says 40 minutes on the standard setting gives it the answer. The same facts serve the shopper reading the page, so the rewrite costs nothing on the search side.
Then check the answers. Ask the assistant the same questions about your product and read what it says. Where it is vague or wrong, the listing is usually silent on that point.
The same work pays twice. The answers that let an assistant quote the listing also cut the questions buyers send before ordering, and the returns that come from a product being something other than what somebody assumed it was.
A large share of the questions shoppers ask an assistant are comparisons: which is better for this use, what is the difference between these two, is the larger size worth it. A listing that never states how the product differs from the alternatives leaves the assistant to work it out from somewhere else.
Compare with the type, not a named rival. Amazon's listing rules restrict naming other brands, and comparisons with a named competitor invite disputes. Compare with the kind of product a shopper weighs it against: a corded vacuum, a plastic bottle, a standard size.
State the differences in words. Lighter by a stated weight, runs for a stated number of minutes longer, holds a stated amount more. A difference given as a figure is one an assistant can repeat; a difference shown only in an image is not certain to be read.
Compare within the range. Where the brand sells several versions, each listing should say how it differs from the others and who each suits. Shoppers ask which of a brand's products to buy as often as which brand to buy.
Answer the trade-off honestly. If the product is heavier because it is sturdier, say so. A listing that admits the trade-off is more useful to the shopper and gives the assistant an accurate answer, and it heads off the return from somebody who expected otherwise.
Use A+ content for the chart, the listing for the words. A comparison chart in A+ content helps a shopper reading the page. The key differences should also appear in the bullets or description, in plain text, so they are stated wherever the answer is drawn from.
What Nobody Outside Amazon Knows
Amazon has not published how the assistant chooses which products to mention or in what order. Anyone selling a formula for ranking in its answers is guessing. What Amazon has said is where the answers come from, and that is enough to act on: a listing that states the facts its buyers ask about gives the assistant something to use, and one that does not leaves the answer to reviews and competitors.
The questions buyers ask change with the season, with competitors and with the product itself. A question map built once and filed away is out of date within a few months.
Add new questions monthly. Read the month's new reviews, product questions and customer messages for questions the map does not have. Each new one gets a group, and a check that the listing answers it.
Watch for seasonal questions. Questions about gifting, delivery before a holiday, cold or warm weather use, and back to school needs appear at predictable times of year. Answering them on the listing a few weeks ahead puts the answer there when the questions start.
Update when the product changes. A new material, a changed size or a revised accessory makes some answers wrong overnight. The listing and the map change the same day as the product.
Measure whether the answers work. Three signs show the map is doing its job. Fewer repeated questions in the product's Q&A and in customer messages. Fewer returns for reasons such as not as described or did not fit. And conversion that holds or rises as the listing answers more.
Check the assistant's answers. Every few months, ask the assistant the main questions about the product and read the replies. Where an answer is wrong or vague, find the gap on the listing.
Kept current, the question map becomes the brief for every listing rewrite, every A+ module and every customer service answer, which is far more than keyword research usually delivers.
Keyword research done this way produces two maps for a product: the phrases, with their demand, and the questions, grouped and matched to the place in the listing that answers them. The prompts running in a brand's campaigns belong beside them, read as part of its
Amazon Advertising, since they show what Amazon has made of the page. A question with no answer on the page, and a prompt that says something the brand would not, are the two findings worth acting on first.