Why AI Assistants Get Prices Wrong: Six Failure Modes, and How to Spot Each One

When an AI assistant quotes a price that turns out to be wrong, the usual explanation is that it hallucinated. Sometimes it did. Far more often the model reported exactly what it was given, and what it was given was wrong in a specific, diagnosable way.

This matters because the two problems have different fixes. A hallucination is a model problem. The other five are data problems, and data problems are the ones a shopper can actually catch.

1. Identity drift: the right price for the wrong product

The single most common failure, and the least likely to be noticed. The assistant searches by product name, gets back something whose title looks similar, and prices that instead. A protective case for a device rather than the device. The 100ml bottle rather than the 200ml. Last year's model, which is genuinely cheaper and genuinely not what you asked for.

How to spot it: the price is dramatically better than everywhere else. A saving that looks too good is far more often a mismatched product than a real bargain.

The underlying cause is that titles are not identifiers. "Wireless Earbuds, Noise Cancelling, Black" describes hundreds of products. A barcode describes one. Any system that matches on words alone will make this mistake at some rate, and the rate rises with the size of the catalogue.

2. Staleness presented in the present tense

The price was correct when it was collected. That may have been an hour ago, or a fortnight. The answer says "costs £249" either way, because nothing in the data structure carried the collection time forward.

How to spot it: ask the assistant when the price was last checked. If it cannot tell you, the age is unknown to it too — which means it is presenting an undated number as a current fact.

Retail prices move constantly, and they move most on exactly the products people ask about most: electronics during promotions, anything seasonal, anything discounted. The categories where an assistant is most useful are the categories where staleness bites hardest.

3. Advertised prices presented as observed ones

Most price data on the internet comes from product feeds — files merchants submit into affiliate, comparison and advertising networks. Feeds are a legitimate and useful way to power discovery. They are also a copy, prepared for redistribution, that can drift from the live page without anyone noticing.

An observed price is read from the surface the retailer serves shoppers: their own API, or their page as a customer actually saw it. Nobody submitted it; nobody chose the moment.

A feed tells you what a merchant said. An observation tells you what a shopper could see. Presenting the first as the second is the quiet failure behind a great many wrong prices.

How to spot it: this one is genuinely hard from the outside, which is the problem. Ask where the price came from. A service that cannot distinguish its own sources cannot tell you.

4. Availability that was never checked

A surprising amount of commerce data defaults unknown availability to "in stock", because the field is a boolean and something has to go in it. The result is an assistant confidently sending someone to buy an item nobody confirmed exists.

How to spot it: availability that is never uncertain. Real data has gaps. A source that reports every item as in stock is not reporting availability, it is reporting a default.

5. The total that is not the total

The item price is £42 and the item price at the other retailer is £45, so the first is cheaper — unless the first charges £7 delivery and the second does not. Cross-border listings add import handling and, sometimes, a customs charge on the doorstep.

How to spot it: a comparison that shows only item prices, with no mention of delivery, is comparing the wrong number. Where delivery is genuinely unknown, the honest presentation is to say so rather than to quietly imply zero.

6. Actual fabrication

It does happen. Asked for a price and lacking a tool that can answer, a language model will sometimes produce a plausible number from its training data — a figure that was roughly right at some point in the past, delivered with the same fluency as a checked fact.

How to spot it: round numbers, no retailer named, no date, and no link. A model that is reading a real source can usually tell you where it looked.

The question that catches most of them

If you remember one thing, make it this. When an AI assistant quotes you a price, reply with: "When was that price last checked, and at which retailer?"

A good answer names a retailer, a time and ideally a specific listing. A vague answer tells you the number is unsupported — not necessarily wrong, but not evidence either. That single question separates most of the six failure modes above from a genuinely checked fact.

And for anything expensive, open the retailer's own page before you buy. That is not a comment on AI assistants specifically; it was equally true of comparison sites for twenty years. The final price is the one at the checkout.