Every sentence an AI assistant generates about your organisation came from somewhere — a training snapshot, a retrieved web page, a cached summary of a summary. Provenance is the discipline of tracing that sentence back to where it actually originated before deciding whether, and how, to correct it. Skip that step, and a correction is just another unverified claim competing with the one you are trying to fix.
What provenance means for an AI claim
Provenance is not the same question as accuracy. Accuracy asks whether a statement is true. Provenance asks where it came from, and through how many hands it passed before a model repeated it. The two questions look similar but demand different evidence: an accuracy check compares a claim to reality; a provenance trace reconstructs the chain of custody a piece of information travelled — from a primary record, through however many intermediaries, to the sentence a user reads in a chat window.
The distinction matters because two identical false claims can have entirely different origins. One might trace to your own outdated investor page, never updated after a leadership change. Another might trace to a forum post with no connection to any of your official channels. The correction that works for the first will do nothing for the second — which is exactly why provenance has to come before correction, not after.
Why corrections fail without it
Most attempts to "fix" what an AI says target the wrong layer. A communications team emails a model vendor's support address, or publishes a rebuttal on a page nobody — least of all a crawler — was ever going to visit. Neither intervention touches the layer that actually produced the error, so the error survives the correction attempt and resurfaces in the next query.
There is also a timing problem. A model's answer reflects whatever slice of the web was retrieved or absorbed into training at some point in the past, not a live read of your newsroom this morning. A correction that ignores this lag — that assumes the model will simply notice a new press release — is solving for a system that doesn't exist. Provenance tells you which layer, and which moment in time, you are actually correcting.
Anatomy of a traced claim
Take a representative case: an assistant states that a company closed a funding round at a figure that is, in fact, wrong. Tracing it typically surfaces a short chain rather than a single culprit. The model's answer often matches, almost word for word, an aggregator database entry. That entry, in turn, was compiled from a blog post that has since been deleted. The blog post, when you can still find a cached version, cited an early unofficial estimate that circulated before the round had actually closed.
At the end of that chain, the "wrong" number turns out to be an accurate transcription of a stale primary source, three hops upstream of the model itself. The error was baked in long before any AI assistant touched the story — which is precisely the kind of finding a correction strategy needs before it can target the right link in the chain, instead of the last one anybody happens to notice.
The four layers we distinguish
In practice, we separate any traced claim into four layers, because a correction aimed at the wrong one rarely holds:
- Primary source — the authoritative record itself: a regulatory filing, an official registry entry, a dated press release, a verified statement from the organisation.
- Indexed intermediary — the page, database or aggregator that repackaged the primary source, sometimes faithfully, sometimes with drift introduced along the way.
- Retrieval snapshot — whatever slice of the web a given model, or the search plugin behind it, actually pulled at the moment the answer was generated.
- Generation layer — the model's own paraphrase, compression or synthesis of what it retrieved, or, absent retrieval, of whatever it memorised during training.
A correction that only addresses the generation layer — asking a vendor to "just fix" an answer — leaves the other three layers untouched, so the same error resurfaces the moment retrieval runs again.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
When the trail goes cold
Not every claim traces cleanly. Some models answer from pure parametric memory, with no citation and no retrieval step to inspect — the information is compressed somewhere in the weights, absorbed from text that may no longer be online in any form you can locate. In other cases, tracing surfaces two or three contradictory upstream sources with no obvious way to tell which one the model actually relied on.
The honest response in both situations is to say so. A provenance trace that ends in "origin unverifiable" is still useful information — it tells you the claim should be treated as unverified rather than corrected against a specific source, and it shapes how much confidence to put in any fix you publish next.
We flag these cases differently in our own reporting, using a distinct status rather than folding them into either "traced and corrected" or "traced and disputed." The distinction matters to whoever reads the report later — a communications team deciding how to respond needs to know whether it is looking at a confirmed error with a known origin, or a claim nobody has been able to pin down at all, since the two call for very different next steps and very different levels of urgency.
Building a provenance-ready record
Organisations can make their own facts easier to trace — and harder to misattribute — well before an error ever appears. A few habits do most of the work: publish dated, authoritative statements as the single canonical instance of a given fact, rather than restating the same number slightly differently across a press page, an investor deck and a bio.
Keep prior versions accessible rather than silently overwriting them, since a model's training snapshot may lag behind your current site by months. Use structured markup so a crawler can tell your official statement apart from a copy of it. And when you do publish a correction, put it at a URL stable enough to still be there — unchanged — the next time anything re-indexes it.
None of this requires exotic tooling. A dated changelog, a single canonical page per fact, and a habit of not quietly editing history without a trace are enough to make most of an organisation's public record traceable — the same standard, in miniature, that we hold a model's answer to when we trace a claim back through it. The bar is consistency, not sophistication.
Provenance as a discipline, not a one-off
At SetRecord, every flagged claim goes through a provenance trace before we recommend how to correct it — the trace determines the target, not the other way around. A claim traced to your own stale content gets a different fix than one traced to a third-party aggregator, which gets a different fix again from one with no traceable origin at all.
That sequencing is the difference between a correction that holds and one that becomes, itself, just another unverified claim about your organisation. Our public methodology documents how we apply it case by case.
The same logic applies internally, not just to what we tell clients. Before any correction goes out, someone other than the analyst who traced it reviews the chain independently — a second pass that catches the cases where an apparently clean trace actually rested on an assumption nobody had stated out loud.