EST. MMXXVI · THE INSTITUTION OF RECORD FOR AI CITATION · OPEN METHODOLOGY

The AI Citation Institute sealTHE AI CITATION INSTITUTEEx corpore, auctoritas.

The AI Citation Institute · The Answers

How often do AI engines refresh their sources?

There are two different freshnesses, and they refresh on completely different clocks. The base model has a training cutoff that is old and fixed — it does not update until the next model ships, months later. But the retrieval layer that most assistants run on top (browsing, search) pulls live pages at question time, so a browsing assistant can cite a page that went up hours ago while its underlying model is still months stale. That gap is the whole point: for getting cited, retrievability matters far more than the model's cutoff, because retrieval is what reads your page right now. Practically, that means keep content substantively fresh — real updates that change what the page says — because freshness tracks with citation. In our panel, content updated within 30 days with a substantive change earned about 3.2x more ChatGPT citations than stale content. The catch is that only real changes count: date-only bumps and bulk-editing lastmod across a whole site don't earn the lift and can get a site punished. So the honest cadence is update when something actually changes, not on a calendar.

Two clocks: training cutoff vs live retrieval

The confusion around "how current is AI" comes from treating the model and the retrieval layer as one thing. They aren't. One is frozen; the other reads the open web at the moment a question is asked.

Training cutoff compared to live retrieval
LayerHow it refreshesWhat it can cite
Training cutoffFixed until the next model ships — typically months apartOnly what existed in the training data; nothing published after the cutoff
Live retrievalAt question time, per query — pulls pages as they are nowPages published hours ago, even against a months-old base model

We don't publish specific per-engine refresh intervals as fact — the retrieval layer is opaque and varies by query.

Why retrievability beats the cutoff

If a page can be fetched and parsed at question time, a browsing assistant can use it regardless of when the base model was trained. That's why the practical lever isn't chasing the model's cutoff — it's making sure your page is reachable, readable, and current when the retriever comes looking.

The corollary: a page that is technically fresh but unretrievable (blocked, unrendered, unlinked) contributes nothing, no matter how recently it was updated.

The cadence that actually earns citations

Freshness correlates with citation, but only substantive freshness. Update a page when what it says actually changes, and the retrieval layer treats it as current.

What freshness signals do and don't earn
ActionEffect
Substantive update within 30 daysAbout 3.2x more ChatGPT citations than stale content in our panel
Date-only lastmod bumpDoesn't count — no lift
Bulk-bumping lastmod across the siteCan be punished as a manipulation signal

Instrument: a frozen five-engine panel, each query sampled k times.

Common questions

Does ChatGPT know about a page published today?+

If it browses, yes — the retrieval layer pulls live pages at question time, so a page from hours ago is fair game even though the base model's training cutoff is months old. Without browsing, it's limited to what it was trained on.

Should I update pages on a schedule to stay fresh?+

No. Update when the content substantively changes. Calendar-driven date bumps with no real change don't earn the freshness lift, and bulk-editing lastmod across a site can get you punished.

How much does substantive freshness actually matter?+

In our panel, content updated within 30 days with a real change earned about 3.2x more ChatGPT citations than stale content. Freshness tracks with citation, but only when the change is substantive.

keep reading

See what AI says about your brand right now — free, on a live engine.

Scan your domain