Where Does an AI's Price Actually Come From? Tracing the Commerce Data Supply Chain

When an AI assistant tells you a kettle costs £39.99, that number has a history. It was produced somewhere, copied several times, reformatted, aged, and finally read aloud in the present tense. Almost none of that history survives the journey, which is why the same product can carry three different prices across three tools that are all technically working correctly.

Understanding the chain is the fastest way to understand why AI shopping answers disagree — and which link to suspect when one of them is wrong.

Hop 1: the retailer's own systems

Every price starts in a merchant's commerce platform. It is real, current and authoritative here, and it is also more complicated than a single number: there may be member pricing, regional variation, a promotion that applies at basket level rather than item level, and delivery rules that change what the shopper actually pays.

The important observation is that "the price" is already an approximation at the source. Anything downstream inherits that ambiguity without inheriting the context that explains it.

Hop 2: publication

The merchant publishes outward, usually in more than one form: the product page a shopper sees, structured data embedded in that page for machines, an API for approved partners, and product feeds submitted into affiliate, comparison and advertising networks.

These are not guaranteed to agree, and in practice they routinely do not. Feeds are generated on a schedule; pages change continuously. The gap between them is not a bug in anyone's system, it is the difference between a copy and an original.

Hop 3: aggregation

Networks and aggregators collect feeds from thousands of merchants and normalise them into one shape. This is genuinely valuable work and it is also where most silent degradation occurs.

• Timestamps get lost or replaced with the time of aggregation rather than the time of collection.
• Missing fields get defaults — availability being the notorious one.
• Currency and market context flatten, so a listing priced for one country appears alongside listings for another.
• Provenance disappears entirely. Downstream, a feed price and an API reading are both just numbers in the same column.

That last one is the load-bearing loss. Once provenance is gone it cannot be reconstructed, and every consumer of that data is now unable to distinguish an advertisement from an observation.

Hop 4: the comparison or verification layer

Here the data is matched across retailers — the step that decides whether two listings are the same product — and prepared for display. Errors introduced here are the most visible to shoppers, because they produce comparisons that are obviously wrong to a human eye and invisible to the system that made them.

Services differ enormously in what they do at this stage. Some re-read the retailer directly and use feeds only to discover which retailers to check. Some pass the feed through and present it as a comparison. Both are called price comparison; they are not the same product.

Hop 5: the assistant

Finally a model reads whatever arrived and turns it into a sentence. This is where the last piece of nuance is stripped, because natural language has no obvious way to carry a caveat that the data structure did not force it to carry.

A field that says "observed four hours ago" produces a different sentence from a field that says "39.99". The metadata is not decoration; it is the only thing standing between a reading and a claim.

What survives a well-built chain

The fix is not to shorten the chain — aggregation is useful and nobody is going to stop doing it. The fix is to carry four things through every hop, none of which is expensive.

1. A product identifier, so identity does not have to be re-guessed from titles at each stage.
2. The time of the original observation, never overwritten by the time of a later copy.
3. A provenance label, so an advertised price and an observed one remain distinguishable at the end.
4. Explicit unknowns, so a field that nobody checked arrives as unchecked rather than as a flattering default.

A system that preserves those four can say honestly how confident it is. A system that loses them has to sound confident, because it has nothing left to be uncertain with.

The question to ask a supplier

If you buy commerce data, or build on it, one question separates most vendors: "For a given price in your response, can you tell me where it came from and when it was collected — not when you processed it?"

It sounds mundane. In practice a surprising number of pipelines cannot answer it, and the ones that cannot are the ones whose data should not be used to make claims about what things cost.