I asked an assistant to find me a mid-priced espresso grinder for a home setup, under a specific budget, that would not wake my daughter at 6am. It named four models, explained the burr difference, and told me which two had documented noise complaints.
Not one of those four brands had optimized for that conversation. They got surfaced because reviewers, forums, and spec sheets happened to contain the attributes the model needed. Three other brands that make perfectly good grinders were absent, not because their products are worse, but because nothing published about them answered the question in usable form.
That is product discovery now. The shelf is being assembled by something that reads specs and complaints, and most catalogs are written for a browsing human who is already on the page.
The buyer is asking constrained questions, not category questions
Nobody asks an assistant "what is a good grinder." They ask for a grinder under $300, quiet, that handles light roasts, from a brand that ships to Curaçao. Four constraints in one sentence.
To be surfaced, your product data has to be matchable against constraints. Which means the constraint has to exist as text somewhere a retrieval system can find it.
Go look at your best selling product page right now. Count how many of these are stated in plain text rather than implied by an image or a lifestyle paragraph: exact dimensions, weight, materials, power draw, noise level, compatibility, what is in the box, what is not in the box, warranty length, country of manufacture, shipping restrictions.
On most of the ecommerce sites I audit, fewer than half are stated. The rest live in a photo of a spec table or in the mind of the product manager.
Fix one: write the attributes nobody thinks to write
The attributes that win constrained queries are usually the ones brands consider too obvious or too negative to mention.
Too obvious: "runs on standard US 120V" or "requires a 2 inch clearance behind the unit." Obvious to you. Decisive to a buyer with a constraint.
Too negative: "does not fit under standard 18 inch cabinets," "not dishwasher safe," "batteries not included." Brands avoid these and it costs them. A model asked for something that fits under an 18 inch cabinet will confidently exclude every product that did not say, and will happily recommend the one honest competitor who did.
I now push clients to publish a short "what this is not for" block on every product page. It reduces returns, it reduces support tickets, and it makes the page matchable on exclusion queries. It also builds exactly the kind of trust signal that lifts conversion with humans at the same time.
Fix two: your structured data is probably incomplete
Product schema with a name and a price is table stakes and does almost nothing. The fields that actually differentiate are the ones most implementations skip: gtin, brand, sku, material, color, size, weight, additionalProperty for anything custom, aggregateRating with real reviews, offers with availability and shippingDetails, and returnPolicy.
Shipping and returns as structured data matter far more than people expect, because "can I get it here and can I send it back" is one of the most common constraints in a purchase conversation.
If you are hand rolling this, a schema markup generator will get you a valid skeleton fast, and then you fill in the fields that are specific to your catalog. The broader argument for going past the defaults is in structured data beyond the basics.
Fix three: the reviews are doing the persuading, not your copy
When an assistant explains why one product beats another, the reasoning is usually reconstructed from reviews, not from marketing copy. Marketing copy is discounted because every brand claims the same things.
This changes what review strategy is for. It is no longer just social proof on the page. It is the training material for how your product gets described to a stranger.
Which means the review prompt matters enormously. "How would you rate us, one to five" produces star ratings and no text. "What were you comparing this against, and what made you choose it" produces exactly the comparative language that gets synthesized into recommendations.
I have watched a client change one review request email and, within two quarters, shift the vocabulary that showed up in AI descriptions of their product. Same product. Different available evidence. The psychology behind why buyers respond to that framing is the same mechanism I unpack in social proof psychology.
Fix four: comparison content, published by you, honestly
Buyers ask assistants to compare. If the only comparison content in your category comes from affiliate sites monetizing whoever pays most, that is the evidence base being used.
Publish your own comparisons. The catch is that they have to be honest, including the cases where you lose, or they will not be retrieved and repeated. A comparison page that concludes you win every scenario is recognizable as marketing and gets discounted.
The version that works: "Choose us if A, B, or C. Choose the competitor if D or E." I have never seen that cost a client a sale. I have repeatedly seen it get cited, because it is the only document in the category that behaves like an advisor instead of a salesperson.
Fix five: make sure you can be crawled and rendered at all
The unglamorous one. A meaningful share of ecommerce sites render product attributes only through client side JavaScript, gate specs behind a tab that loads on click, or block the crawlers that feed AI systems in robots.txt without realizing it.
Check your robots file deliberately rather than inheriting whatever the platform shipped. Write one you actually understand rather than inheriting a default. Then confirm that a plain fetch of your product URL returns the spec text in the HTML source. If it does not, nothing else on this list matters.
The rest of the technical baseline is the same as it has always been, and I keep the current version in the technical SEO checklist.
What this does not change
It does not change that the product has to be good. It does not change price sensitivity. It does not eliminate brand preference, and in fact I think strong brands get more resilient here, because a buyer who names your brand in the prompt has already skipped the comparison step entirely. Everything I have argued about brand differentiation gets more valuable, not less, as discovery gets automated.
What it changes is who gets into the consideration set when the buyer has no preference. That used to be decided by ad budget and shelf placement. It is increasingly decided by whether your product data can survive a constrained question.
Start with your worst converting hero product
Not your best. Your best product is probably converting on brand recognition and will teach you nothing.
Take a product you believe in that underperforms. Write out the twelve constraints a buyer might apply to that category. Check how many your page answers in plain text. Fix the gaps, add the exclusions, ship complete product schema, and change your review request to ask what they compared it against.
Then wait a quarter and ask an assistant to recommend something in that category with three of those constraints. My experience across client catalogs is that this single page audit, done properly, takes about four hours and is worth more than the last six months of adding lifestyle photography.