Research

How Training Cutoffs Change Brand Recall

What happens to a brand when a model refreshes its knowledge cutoff. Why cutoff jumps redistribute recall between competitors, how to detect it, and when to re-baseline after a release.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

A model's knowledge cutoff is the most underrated variable in brand visibility. When it moves, unprompted brand recall changes for every company in a category at once, and nothing you did caused it. Teams that do not track cutoffs routinely misattribute these shifts to their own content work.

Cutoffs Move in Jumps, Not Gradually

Retrieval is continuous: an assistant with web access sees today's page today. Parametric recall is not. It updates only when a new model ships, and the jump can be large. Gemini 3.6 Flash moved its cutoff from January 2025 to March 2026, ingesting fourteen months of web events in a single release.

PropertyRetrievalParametric recall
Update frequencyContinuousOnly on model release
What it rewardsCrawlable, well-structured, current pagesBreadth and durability of coverage before the cutoff
Time to influenceDaysMonths to years, then fixed
ReversibleYes, update the pageNo, not until the next model

Why a Cutoff Jump Redistributes Recall

Recall is relative. A model's sense of which brands matter in a category reflects the coverage density it saw during training. When the window extends, three things happen at once. Brands that earned substantial coverage in the newly-included period gain recall. Brands whose strongest coverage sits well before the window gain nothing and lose relative position as competitors enter. And brands that materially changed, rebranded, pivoted, were acquired, get re-described according to the newer information.

The third case is the one that produces the most confusing measurement, because a brand can see its mention rate hold steady while what the model says about it changes completely.

Detecting It

Four signals that a shift is cutoff-driven rather than content-driven. It appears within days of a model release rather than gradually. It affects unprompted recall questions more than retrieval-heavy ones. It moves competitors in the same category simultaneously and in opposite directions. And it does not reproduce on the previous model version, where that is still available through the API.

That last check is the decisive one and it is why keeping the prior model accessible in your measurement stack matters. Once a provider deprecates it, the counterfactual is gone. See the model deprecation tracker.

Practical Guidance

Re-baseline within two weeks of any major model release rather than carrying prior numbers forward, and record which model version each measurement came from. Weight this most heavily for default-tier models: Claude Sonnet 5 becoming the default on Free and Pro, or a Flash-tier model backing AI Overviews, resets what most users see, whereas a premium-tier release affects a much smaller population. See the July 2026 release roundup.

Strategically, the asymmetry between the two channels is the useful insight. Retrieval is fast and reversible, so it is where near-term work pays. Parametric recall is slow, durable, and fixed at training time, so coverage earned now is an investment in models that have not been trained yet.

Methodology

Framework is Presenc AI's, developed from continuous multi-platform measurement across model transitions. Cutoff dates are as published by model vendors. Vendors do not always disclose cutoffs precisely and sometimes revise them, so treat published dates as approximate.

How Presenc AI Helps

Presenc AI version-stamps every measurement and re-runs baselines automatically on major model releases, so visibility changes get attributed to the model or to your content correctly rather than by guesswork.

Frequently Asked Questions

It sets what the model knows without retrieval. When a cutoff moves forward, brands with strong coverage in the newly-included period gain unprompted recall, brands whose best coverage predates the window lose relative position as competitors enter, and brands that changed get re-described. All of this happens at once on release day.
Four signals: it appears within days of a release rather than gradually, it affects unprompted recall more than retrieval-heavy queries, it moves competitors simultaneously and in opposite directions, and it does not reproduce on the previous model version through the API. The last check is decisive.
Within two weeks of any major model release, and always record which model version produced each measurement. Prioritise default-tier models, since a change to the model backing free consumer tiers or AI Overviews resets what most users see, while a premium-tier release affects far fewer people.
Only before the cutoff, and only slowly. Parametric recall is fixed at training time and cannot be changed afterward, unlike retrieval which updates within days of a page change. Coverage earned now is effectively an investment in models that have not been trained yet.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.