When an AI agent repeats your listing, your product claims become its sales pitch

There is a line buried in Slobodan Manic's new test of AI shopping agents that we keep coming back to. After counting where the models' stated reasons came from, he wrote: "The model has no way to check the claim, and the shopper hears it from the agent, not from the store."

That sentence describes a shift in who is speaking when a product gets sold. It is easy to miss among the test's bigger findings about review counts and model disagreement, which we charted in today's data piece. We think it may matter more.

What the data showed

Manic gave three Claude models (Haiku 4.5, Sonnet 5 and Opus 5) lists of 20 products returned by Shopify's catalog and asked each to pick one and explain why. The models saw what an agent would see: title, store, price, rating, review count, a one-line summary and the first three features. No web access, nothing else to go on. His full write-up is on Search Engine Journal.

He then traced the language in the models' reasons back to the listings.

Phrase in the model's reason	Reasons using it	Taken from the picked product's listing
"clinically proven" (eight skincare searches)	63	61
works "11x faster" than retinol	22	22 (19 word for word)
"without clogging pores" (Opus, oily-skin sunscreen)	33 picks of that product	cited in 31 of the 33 reasons

Source: Slobodan Manic via Search Engine Journal.

In other words, the listing was not just an input to the decision. It became the explanation the shopper reads.

The voice changes, and so does the weight

When a shopper reads "clinically proven" on a brand's own product page, they read it as the brand talking. Most adults have a lifetime of practice discounting that. It sits next to a logo, a price and an add-to-cart button. The context says: this is marketing.

When the same words arrive as the agent's reason ("I picked this because it is clinically proven to..."), the context is gone. The speaker is an assistant the shopper chose, one that presents itself as working for them. The claim borrows that assistant's neutrality. It sounds like a finding rather than a pitch.

Nobody designed this as a trick. The model is doing what it was asked: choose a product and give a reason. The most quotable reasons in a product record are the brand's claims, so those are what come back.

The listing was not just an input to the decision. It became the explanation the shopper reads.

Responsibility does not move with the voice

It would be convenient for brands if the agent's voice also carried the agent's responsibility. It doesn't. The claim still originates in your listing. And the commerce plumbing being built this fall keeps the brand firmly in the seat.

Shopify's post on Shop Pay in Meta's Muse is explicit that brands remain the merchant of record and own the customer relationship and post-purchase experience. Google's direct checkout, which launched in Canada today, keeps the retailer as the seller of record too, as CBC reported.

So the agent speaks, and the brand answers for what was said. If a claim is overstated, the shopper's disappointment, the return and any complaint land with you, not with the model that repeated it.

There is a consumer-policy angle as well. Vass Bednar of the Canadian SHIELD Institute told CBC that "in an AI shopping context, choice is more constrained than on the open web." When fewer products are shown and each comes with a confident reason, the reason does more of the persuading. That will draw scrutiny over time, and the easiest claims to defend will be the ones that were true to begin with.

The same mechanism rewards plain, specific copy

This is not only a risk story. The test also showed what good listing language does in an agent's hands.

When the request was "mineral sunscreen SPF 50, under $25, for oily skin," Opus split its picks between two products. One of them, Sun Bum's lotion, had the lowest rating of the six under $25. Its listing was the only one that said "without clogging pores," and that phrase carried it to a third of Opus's picks. Badger's sunscreen had the best rating in that group and was never picked; its copy talked about "organic plant oils," which did not connect to anything the shopper asked.

The lesson is not "add more claims." It is "say the thing the customer needs, in the customer's words, and make sure it is true." A specific benefit tied to a real need is both the most persuasive line for an agent and the easiest one to stand behind.

Shopify's own guidance in the Muse post points the same way: "Agents can only recommend what they understand," it says, so clear titles, structured attributes, complete variants and reviews "matter more than ever."

What we think brands should do

We would treat agent-readable listings the way regulated categories already treat packaging: as copy with consequences. Concretely:

Write as if it will be read aloud by someone else

Read each listing and ask: if an assistant said this sentence to a customer as its reason for recommending us, would we be comfortable? Would we be able to show the evidence? If not, rewrite or remove it.

Prefer specific, checkable statements over superlatives

"Fragrance-free, tested on sensitive skin in a 4-week study of 30 people" gives an agent something concrete and gives a shopper something to verify. "Clinically proven results" gives neither. (The study wording is an example with illustrative figures, not real data.)

Put the customer's need in the first lines

Agents in the test saw only the summary and the first three features. If "for oily skin," "nut-free," "fits a 13-inch laptop" or "dishwasher safe" is true, put it where it will be read.

Keep the proof one click away

Some shoppers will check. Fractl's survey work suggests many do, which we covered in the quiet click. The page a citation lands on should back the claim quickly.

Watch what the agents actually say

You cannot see inside an agent's decision, but you can see its output. Run your category's common requests through a few assistants and read the reasons given for your products and your competitors'. A tool can do this at scale (XENA Intelligence does it for consumer brands, and SearchPanel reports on individual products), but even a monthly manual check will tell you whether your claims are being repeated accurately.

The bigger point

For most of e-commerce history, product copy had one reader: a person scanning a page. Now it has two, and the second one repeats what it reads to the first, in its own voice, with no way to tell a strong claim from a weak one.

That raises the value of honest, specific copy and lowers the value of puffery. Brands that already write labels carefully have a head start. We made a related argument about Shopify's knowledge base being the new label copy; listings are the same story at a larger scale. And if you run ChatGPT ads, the automatic rewrite setting is the paid-media version of the same question: who gets to put words in your mouth, and are those words true?

The agent will say what you wrote. Write something you would be glad to hear it say.