Every AI Shopping Tool Is About to Say It "Verifies" Prices. Here Is What That Should Have to Mean

Ask an AI assistant what a pair of headphones costs and it will tell you. It will give you a number, a retailer, and often a confident sentence about whether now is a good time to buy. What it will almost never tell you is where that number came from, when anyone last checked it, or whether the product it priced is actually the product you asked about.

That gap is about to become a marketing battleground. As AI assistants start answering shopping questions at scale, every product in the category will claim it verifies prices. The word will appear on landing pages within months. Almost none of them will say what they mean by it — and a shopper will have no way to tell an observation from an advertisement.

This is not a gap nobody has worked on. Schema.org has described how to express an offer for years, and the newer agentic-commerce protocols — Google's UCP, OpenAI and Stripe's ACP — define how an assistant and a merchant transact. What none of them define is the layer underneath: the evidence somebody should hold before presenting a price as verified in the first place.

So we have published the criteria WEM holds itself to, in full, including the parts that constrain us. They are at /standard/verified-offer, and they are deliberately narrow: they say nothing about how many products a catalogue contains, how good a recommendation is, or how a purchase should be executed. They cover one question only — whether a stated commerce fact is supported by evidence, and what that evidence has to look like.

The distinction that does most of the work

One criterion carries more weight than the rest, and it is the one the industry finds least comfortable. A price is either advertised or observed, and the difference decides whether it can verify anything.

An advertised price is supplied for distribution — submitted by a merchant into an affiliate, comparison or promotional feed, inside a commercial relationship of one kind or another. The merchant chooses what the feed says and when it updates. That is a perfectly legitimate way to power discovery, and it is how most of the price data on the internet is produced.

An observed price is read from the surface the retailer serves to shoppers: their API, or their page as a customer actually saw it, whether or not any commercial relationship exists. Nobody submitted it. Nobody chose the moment.

A feed is evidence of what a merchant said. An observation is evidence of what a shopper could actually see.

Both surfaces belong to the retailer, so this is not a claim that one is independent and the other is not. It is a claim about what they bind. The page is the one a customer transacts on; a feed is a copy prepared for redistribution, and a copy can drift from the original without anyone noticing. Catching that drift is the entire job, and it cannot be done from the copy — which stays true no matter how large the feed gets.

This is not a criticism of comparison feeds. It is a statement about what they can and cannot be used for. Discovery and verification are different jobs, and the data that does one well is structurally disqualified from the other.

The rules that cost us something

A standard that only forbids things your competitors do is marketing. These are the criteria that constrain WEM, and we think publishing them is the point rather than the price.

Silence where the evidence is missing. Where an offer cannot meet the criteria, we withhold it rather than estimate, infer, or fill a gap to complete a comparison. The pressure in this business is always to show something. A gap teaches you that the absence means something; a plausible guess teaches you nothing and costs you the ability to trust anything else on the page.

Availability stated only where it was checked. If a source did not report whether an item is in stock, we record that as unknown and never render it as available. Defaulting unknown to "in stock" is the easiest way to make a comparison look better than reality, so the rule forbids the default rather than trusting anyone's good intentions.

Timestamps, and disclosure of age. Every observation carries the moment it was made, and anything we publish says how old it is instead of implying "now". A price with no time attached cannot be shown wrong — which means it cannot be shown right either. "Checked four hours ago" is a weaker claim than "live", and an honest one.

Lawful to publish. Some retailers permit price data to be read but not retained, or retained but not republished. Where the terms do not allow it, we refuse the data at the point of collection rather than filtering it later. A verification corpus you cannot show anyone is not evidence, and a clean provenance story does not create a retention right.

On being funded by commission

WEM earns affiliate commission when a shopper buys through a link. We think stating that plainly is more useful than the alternative claim now appearing in this category, which is that a service takes no commission at all.

"No kickbacks" is not neutrality — it is an undisclosed business model. Someone is paying: advertisers, data buyers, or investors subsidising the service on the way to becoming one of those. Neutrality is not the absence of revenue. It is the separation of revenue from answers, stated openly enough that you can check it.

So the criterion is written as a property of the system rather than a promise about our intentions: ranking and verification are computed without reference to commercial terms. The cheapest offer wins because it is cheapest. Nothing in that code path knows what anything earns.

Why publish it rather than just do it

Two reasons, and the second one is the honest one.

The first is that an undefined word is worthless. If "verified" can mean anything, it protects nobody, and the shopper is back to trusting whoever sounds most confident. Writing the criteria down makes the claim checkable by someone who has no reason to trust us.

The second is that we would rather be measured against a published standard than an implied one. We have made mistakes in this system and found them by looking — a field that recorded "in stock" when nothing had checked, a page that displayed a price history it had created by being rendered. Both were fixed because a written rule made them visible as failures rather than as defaults. A standard you publish is a standard people can catch you failing.

Where this is going

Version 1.0 is dated and versioned so that a citation can name the edition it relied on. If it changes, the URL stays and the version moves — a standard that quietly rewrites itself is not one.

It is free to read, cite and apply. If you build commerce data and want to say you meet it, please do — and say which criteria you meet, because in our experience the ones a dataset fails are the informative ones.