- A citation refreshes on the schedule of whatever triggers the next retrieval. A mention sourced from training data has no schedule: the model reproduces a memorized pattern, not a live lookup, and nothing about correcting your site today reaches back into what it already learned.
- The only thing that actually updates a memorized fact is a new training run, and that happens on the provider's timeline, not on request. No lab offers a way to patch one specific fact in a deployed model.
- Real knowledge-cutoff dates make the gap concrete: as of September 2026, cutoffs range from Claude Opus 5 (May 2026) down to GPT-4o (October 2023), a model nearly three years stale, still live in production API integrations.
- The practical implication: correcting a source today prevents the error from being picked up by future training and fixes any retrieval-based mention immediately. It doesn't retroactively correct what an already-deployed model memorized.
Mention vs. citation vs. recommendation already states the mechanism briefly: a mention sourced from training data has no retrieval event to trigger a refresh, so a stale fact can persist indefinitely. This piece is the fuller version: why citations and mentions age on different timelines, what a "refresh" actually requires for each, and how long the gap can realistically run.
Why can a citation refresh when a mention structurally can't?
A citation is downstream of retrieval: something had to be fetched for something to be cited. That means a citation's accuracy is only ever as stale as the last time the source page was fetched. Fix the page, and the next retrieval-enabled query has a real chance of returning the corrected fact, because the model is looking the answer up again rather than recalling it.
A mention sourced from training data works differently: the model isn't looking anything up. It's reproducing a fact pattern it learned once, during a training run that happened at some point in the past, from a dataset that itself stopped changing on a specific date. There's no "next retrieval" for a memorized fact, because nothing was retrieved to produce it in the first place, which is exactly why what to do when AI gets your brand wrong treats correcting a source as the fix for citation-based errors specifically, and flags uncited, training-data-sourced errors as a structurally harder problem.
What actually updates a memorized fact?
Only one thing does: a new training run on a dataset that includes the corrected information, released as a new model version. There's no mechanism, from any major lab, for patching a single memorized fact inside an already-deployed model: no equivalent of editing a database row. A provider decides to retrain, at a cadence driven by cost and product strategy, not by any individual correction request.
Knowledge cutoff dates make this concrete rather than abstract. As of September 2026, the most recent cutoffs among major models include Claude Opus 5 (May 2026), Gemini 3.7 Flash (March 2026), and GPT-5.6 (February 2026). Older models remain live well past their own cutoff. GPT-4o, still active in API integrations as of this writing, has an October 2023 cutoff: a model answering questions today from a knowledge base that's nearly three years old, with no path for that specific deployment to catch up short of a provider removing it or a caller switching model versions.
How long can a stale fact realistically survive?
The honest answer: as long as whichever model snapshot memorized it stays in active use, which current cutoff data shows can be years rather than weeks or months. A brand's stale pricing, a discontinued feature, a category it no longer competes in, any of these, if memorized during a training run, can keep surfacing in mentions from that model version for as long as that version stays deployed somewhere a user or an API caller can reach it, regardless of how quickly the brand corrects the actual, current source of truth.
This is a different risk profile than a citation error. A citation error is bounded by the next crawl. A training-data-sourced mention error is bounded only by a provider's own model-retirement timeline, which isn't public, isn't predictable per-fact, and isn't something a brand has any lever over directly.
What does this mean practically?
Correcting every source you control (your own site, Wikidata, G2, and the rest of the entity-consistency footprint) is still the right move, and it does real work: it's what a future training run would pick up, and it's what fixes any citation-based mention of the same fact immediately, the moment a retrieval happens against the corrected page. What it doesn't do is reach backward into a model that's already deployed and already memorized the old version. That's not a gap in your correction effort. It's a structural property of how these systems update, and the honest response is bounding expectations to match it rather than promising a fix on a timeline nobody can actually guarantee.
Frequently asked questions
Why can a citation refresh while a mention structurally can't?
A citation is downstream of a live retrieval: the next time a crawler or a web-search-enabled query fetches the page, it gets whatever is currently there. A mention sourced from training data has no such trigger. The model isn't checking your site again; it's reproducing a pattern it learned once, during training, and that pattern doesn't update until the model itself is retrained, which per this site's own research on mention vs. citation, can mean a stale fact persists indefinitely with no retrieval event to ever correct it.
What actually triggers a model to update a memorized fact?
A new training run, not a correction request. Knowledge cutoff dates only move when a provider releases a new model version trained on more recent data. There's no fixed schedule, and no mechanism for updating one specific fact in isolation. As of September 2026, cutoffs range from Claude Opus 5 (May 2026) and GPT-5.6 (February 2026) down to older models like GPT-4o (October 2023) that remain live in API integrations.
How long can a stale fact realistically survive, then?
As long as whichever model snapshot memorized it stays in active use (which, per current knowledge-cutoff data, can mean years, not weeks, for older model versions still running in production API integrations even after a newer model exists). There's no guaranteed timeline for a memorized fact to age out on its own.
What can you actually do about a stale, training-data-sourced fact?
Correct every source you control now, so newer model training runs pick up the right fact going forward, and so any citation-based (retrieval-sourced) mention of the same fact gets the correction immediately. What you can't do is force an already-trained model to forget what it memorized. That's a real, bounded limitation worth stating plainly rather than promising a fix this category of error doesn't support.
Sources referenced
This piece is the full mechanism behind a limitation already stated in mention vs. citation vs. recommendation. For the actual correction workflow this limitation bounds, see what to do when AI gets your brand wrong.
