How do I know it's time for a full headline refresh due to creative fatigue?

It helps to understand the underlying mechanics here: replacing text inside an existing RSA is an edit to a live ad, not the creation of a new one. The ad's overall accumulated stats (total clicks, conversions, history as a single Ad ID) don't reset. But every individual new asset (a new headline or description swapped in) starts from zero — it has no history and has to build its own data from scratch, regardless of how old the ad itself is. That's exactly why "how often to refresh" comes down not to whether you can refresh at all, but to how many assets you swap at once and how often, so you're not wiping out the statistical base wholesale.

Fact-check: The premise behind this question needs a correction before anything else. Google Ads officially does not assign a "Learning" status to an RSA as a single entity. Two independent sources confirm this:

The combinations report for RSAs is real and officially documented by Google — it's a separate tab inside the asset report showing the specific headline-and-description pairings that were actually assembled into an ad, along with the impression count for each combination. You can find it via "View asset details" on a specific RSA → the Combinations tab.

Here's the important thing to understand: a lack of correlation between Ad Strength and real performance isn't an anomaly — it's a documented property of the metric itself. Google officially describes Ad Strength as a score built on best practices, meant to help you make a good first impression — in other words, the metric isn't built from your ad's actual performance data. In 2024, a Google representative (Ginny Marvin) directly confirmed that Ad Strength isn't used in Ad Rank calculations — meaning an ad with a low rating can, technically, show and perform just as well as one rated "Excellent."

When multiple RSAs run in one ad group, the system has to decide, before each individual impression, which of your own ads to enter into the auction — and only then, which combination of assets to assemble inside it. That effectively turns multiple RSAs into parallel forecasting models competing for the same auctions inside your own account, rather than against outside competitors. The industry describes the direct consequence of this as impression fragmentation: instead of one ad quickly accumulating enough data to self-optimize, impressions get spread across multiple ads, each of which ends up under-supplied with data — and the whole ad group underperforms as a result, not because the copy is weak, but because none of the variants ever built up enough statistics.

The mechanics here are straightforward and backed by independent large-sample analysis. Fully pinning assets almost always drives an RSA to a "Poor" Ad Strength status, which in turn reduces impression volume for the fully pinned RSA compared to other ads in the same group that pin less aggressively. Every pinned position cuts into the number of combinations the system can assemble and test: pinning just two headlines has been shown to cut the system's testing potential by roughly 99.5% — a sharp, nonlinear effect even from a small number of pinned positions.

Google's official position on testing RSA copy is direct: ad-level metrics like CTR and conversions don't give you the full picture, since RSAs themselves help you qualify for more auctions — so results are better judged by incremental impressions, clicks, and conversions at the ad group and campaign level, rather than by the isolated metrics of one ad or one combination. This directly explains why testing every individual combination one by one isn't the path the RSA architecture is built for — the system tests combinations inside an ad on its own logic, not the logic of a manually controlled experiment.

The fact that an ad keeps running with some assets disapproved isn't a bug — it's a direct consequence of RSA architecture: each individual asset (headline or description) goes through its own policy review independently, rather than as one indivisible unit with the ad. This is confirmed by an officially documented parallel example: when campaign-level assets (shared headlines/descriptions that can be used across multiple RSAs) fail policy review, other approved creatives and campaign-level assets are used for serving instead — the system automatically substitutes the disapproved element with one that's working. The same logic is officially documented for text disclaimers as well: if a disclaimer asset is disapproved on policy review, the ad keeps showing — a different description line, or one pinned to the first position, is used in place of the disclaimer.

Fact-check: This question's framing relies on an outdated mechanism. The "Performance label" column (Low/Good/Best) in the RSA asset report has been officially deprecated by Google, since full performance statistics are now available for each asset — and that full statistics view is available for date ranges starting June 5, 2025. The same applies to the campaign-level asset report: the "Performance label" column was removed precisely because full performance figures are now available for each asset. So "how do I connect asset performance to the label" isn't quite the right question for RSAs anymore — the label itself no longer exists in this report; it's been replaced with raw metrics.

Here's the thing to understand: a headline's effectiveness isn't a property of the text itself — it's the outcome of that text's interaction with the context of a specific ad group, and that context is genuinely different across ad groups even when the headline text is identical. Google's official account-organization guidance is built on exactly this principle: for each ad group, it's recommended to pick a narrow theme and use keywords relevant to that theme, including at least one of those keywords in the ad's headline — since a user is more likely to find an ad relevant when it explicitly mentions the term they searched for. If the same headline is used in two ad groups with different semantics, it will, by definition, match one group's queries more precisely than the other's, simply because the surrounding keywords and real search queries differ.

There's a key officially documented property worth establishing up front, since it directly shapes what's even possible to measure. Google explicitly documents that the automatic text updates that happen when a customizer fires don't reset or fragment an ad's performance data — all the stats keep accumulating on the same ad, regardless of which specific dynamic value was shown at any given moment. That's convenient in that reporting doesn't "multiply," but it also means there's no way to see directly, in the interface, how a specific countdown value or price actually affected clicks and conversions — standard reports don't break results down by which version of the dynamic text rendered at the time of a given impression.

Dynamic Keyword Insertion (DKI) technically works on the same architecture as the ad customizers and countdown covered earlier: Google explicitly describes ad customizers as working the same way as keyword insertion or countdown timers — that is, adding dynamic content directly into the ad, with all stats continuing to accumulate on the same ad regardless of which specific text was inserted for any given impression. That directly creates the main reporting risk: if a DKI ad shows against dozens of different keywords in a group, all those visually different text versions ("Buy Nike sneakers," "Buy Adidas sneakers," "Buy New Balance sneakers") get rolled up into one single ad's reporting row — you simply can't see in a standard report which inserted phrase drove more clicks or conversions.