Why AI answers about your company go stale
By Miscited · Published 27 September 2026 · 11 min read
A stale answer repeats a fact that was true once, after it stopped being true. That makes it harder to catch than an invented one, because every part of it checks out against somebody's old page.
Stale answers come from two places, and they need different fixes. Both come down to the same missing piece: knowing when each of your facts was true.
Two ways an answer gets its facts
An assistant answering a question about your company draws on one of two things, and often both.
| Training memory | Live retrieval | |
|---|---|---|
| What it is | What the model absorbed from text collected before its training cutoff | Pages the assistant fetches from the web while it answers |
| How it goes stale | The cutoff passes and your facts keep changing | It finds a page that still states the old fact |
| When a fix can show up | Only in a model trained after the fix, if then | After the corrected page is fetched again |
| A sign to look for | No sources, or sources that do not contain the claim | A citation to a page that states the old fact |
A training cutoff is the point after which a model saw no new training text. Anything that changed later, the model can only learn by retrieving a page while it answers. Retrieval solves that in principle and brings it back in practice. An assistant that searches the web can just as easily read a pricing page you forgot to retire, a directory profile nobody updated, or a news story from the week of an old announcement.
Which mode produced an answer decides which fix can work. We covered that split in how to fix a wrong AI answer about your company.
Superseded is not the same as wrong
A fact list that only records what is true today cannot catch a stale answer, because a stale answer was true. What catches it is the date. Every fact about a company has a period when it was in force: a price between two dates, an owner until a sale completed, a product name until a rebrand.
That gives you two questions to ask of any claim. Was it ever true? Was it true on the day the answer was given? A claim that fails the first is contradicted. A claim that passes the first and fails the second is stale, and it points somewhere specific: a source that was accurate when it was written and was never updated.
There is a third state worth recording: a change that has been announced and has not yet taken effect. An answer that repeats the new fact before its effective date is wrong too, in the other direction.
An illustration: who runs Aeroplan?
This is an illustrative example built from two public announcements. It is not a Miscited finding: we have not measured what any assistant says about Aeroplan. Air Canada is not a Miscited customer.
| Date | What happened |
|---|---|
| 26 Nov 2018 | An agreement for Air Canada to buy Aimia Canada, the owner and operator of Aeroplan, is announced |
| 10 Jan 2019 | Aimia completes the sale of Aimia Canada to Air Canada |
| 11 Aug 2020 | Air Canada announces a redesigned Aeroplan, launching on 8 November 2020, with Aeroplan miles to be known as Aeroplan points |
| 8 Nov 2020 | The announced launch date of the new programme |
Now take three claims an assistant could make about Aeroplan and check each against those dates.
| Claim | Made | Verdict | Why |
|---|---|---|---|
| Aeroplan is Aimia's loyalty programme | After 10 Jan 2019 | Stale | True until the sale completed that day |
| Aeroplan members collect miles | After 8 Nov 2020 | Stale | Miles were to be known as points from that date |
| Aeroplan members collect points | Sep 2020 | Early | The change was announced and not yet in force |
The second row assumes the change took effect as announced. Our sources are the announcements, and a real fact list would confirm it.
None of these claims was invented. A list of today's facts alone would flag the first two as contradictions and send you looking for a fabrication that does not exist. With dates, each one points at its likely source: pages written before 10 January 2019 in the first case and before 8 November 2020 in the second. A model whose training text ends before January 2019 cannot know about the sale unless it retrieves a newer page, and one that retrieves an old article can repeat the old owner with a citation attached. The third row shows why a fact needs the date it takes effect as well as the date it was announced.
Why a dated fact list fixes this
Everything above comes down to one habit: record each fact with the date it took effect, the date it stopped if it has, and the source that says so. For the example, the list would look like this.
| Fact | Value | From | Until | Source |
|---|---|---|---|---|
| Company behind Aeroplan | Aimia | — | 10 Jan 2019 | Aimia, 10 Jan 2019 |
| Company behind Aeroplan | Air Canada | 10 Jan 2019 | open | Aimia, 10 Jan 2019 |
| Aeroplan currency | Miles | — | 8 Nov 2020 | Air Canada, 11 Aug 2020 |
| Aeroplan currency | Points | 8 Nov 2020 | open | Air Canada, 11 Aug 2020 |
Open means no end date has been recorded, not that someone confirmed the fact this morning. A working list also records when each entry was last checked.
With a list like this, classifying a claim is mechanical. A claim that matches a row in force on the day of the answer is correct. One that matches a row whose end date has passed is stale. One that matches a row that has not started yet is early. One that matches no row at all needs checking before it can be called contradicted or unsupported. The same list tells you where to look for the source of a stale answer: pages published while the old row was in force.
The full checking method is in how to check what ChatGPT, Gemini and Perplexity say about your company, and how many runs a stale rate needs before it means anything is in how many prompts an AI visibility number actually needs. The free answer audit starts from facts read from your own site and flags answers that conflict with them. Those site facts are provisional until a person reviews them, which is the moment to add the dates.
Sources
- Aimia Completes Sale of Aimia Canada to Air Canada, Aimia, 10 January 2019.
- Air Canada Unveils Details of Transformed Aeroplan Program, Offering More Value and New Benefits, Air Canada on PR Newswire, 11 August 2020.
Questions people ask about this
Why does ChatGPT have outdated information about my company?
Either the answer came from training memory, which stops at the training cutoff, or the assistant searched the web and read a page that still states the old fact. The first changes only when a later model is trained. The second can change once the old page is corrected and fetched again. Knowing which one you are looking at decides the fix.
What is a training cutoff?
It is the point after which a model saw no new training text. Anything that changed later, such as a new price, owner or product name, the model can only learn by retrieving a page while it answers. Without retrieval, it answers from what was true before the cutoff, which is how a fluent but stale answer is produced.
Is a stale answer the same as a false one?
Not quite. A stale claim was true once and has since been superseded, while a false one was never true. The difference matters because a stale claim points to a findable source: an old page or profile published while the fact was in force. Correcting or retiring that source is where the fix starts.
How do I stop an old fact about my company being repeated?
Record the fact with the date it stopped being true, then find the pages that still state it: your own old pages first, then directories and third-party articles. Correct or retire them, re-ask the same questions and compare the results with questions you left alone, so you know the change came from your fix.