The WAU effect

December 23, 2025 General, Original Content, Retention

As I argue in Understanding ChatGPT’s unit economics, OpenAI’s choice to report Weekly Active Users (WAU) instead of the more common Monthly Active Users (MAU) as an indicator of product growth is curious. To my knowledge, the only other scaled consumer technology product that reports WAU is the iOS App Store — and even then, it only does so irregularly. The most recent publication of the WAU metric for the App Store was in a whitepaper from June 2025 (co-written for Apple by a former MDM podcast guest), which revealed that the App Store sees 813MM “weekly active visitors.”

Coincidentally, that number very nearly matched the 800MM WAU that OpenAI announced at its Developer Day in October. The Information reported two weeks ago that, as of mid-December, ChatGPT is now used by “nearly 900 million” people. Assuming that ChatGPT had 800MM WAU on October 1st and 899MM WAU on December 1st, ChatGPT’s WAU grew by just 12% in the last two full months.

But why use WAU instead of MAU in the first place? Two obvious reasons:

  • Strategically, OpenAI’s use of the WAU metric renders its user base scale incomparable to any other large consumer technology product (except the App Store).
  • As I argue in Understanding ChatGPT’s unit economics, ChatGPT’s user retention is likely quite low, meaning many users cycle in and out of the product over any given month, inflating MAU relative to WAU. Given that OpenAI has frequently published revenue metrics (the company stated last month that ChatGPT has reached $20BN ARR), dividing revenue by the larger MAU would reveal relatively weaker unit economics that are directly comparable to other consumer technology products.

Another, less obvious but certainly plausible reason ChatGPT might report WAU instead of MAU is that WAU masks saturation and is therefore more likely to be monotonically increasing over time. This is a nuanced point that requires an understanding of how product retention impacts cohort compounding; for background, see:

In Product retention is not binary, I propose that retention should not be viewed as a permanent, binary outcome (active / churned) but rather as a malleable status that the product operator can influence through re-engagement tactics. This distinction matters because it reframes churn not as an endpoint but as a temporary state that can be reversed at some expense. From that piece (emphasis mine / new):

Traditionally, user churn — or absence from an app — has been seen as a permanent state: if a user is churned, which is usually defined as some number of days without registering as a DAU, they are considered gone from the app for good. But as the app economy has matured and shifted into its second act, successful app developers must consider a relationship with users that might span years, if not decades … It’s perfectly reasonable to expect a user of a video streaming service to choose to churn — or stop paying a subscription — between, for example, seasons of their favorite show, or to vacillate between services as new content becomes available on each. In such a case, a user might always be a subscriber to a streaming service over the course of a year, but not one specific streaming service continuously in that time. In this case, product retention and churn are fluid concepts: not binary and certainly not permanent.

The case I describe here with streaming services should be familiar to most people, but it can be cast another way: what if a user churns because they believe the streaming service simply has nothing left to offer, despite its broad catalog of programming? And what if that user only returns because they see a large-scale promotional campaign for a new series? This would mean that the streaming service’s pool of content has no bearing on retention and that the user can only be re-activated with incrementally new content that the streaming service must pay to create. In such a model, engagement is renewed episodically and doesn’t compound, a dynamic that flatters shorter reporting windows like WAU.

This requires a content treadmill, meaning subscriber growth is a direct function of content spending. For a streaming service, this may not be a bad outcome: a user is by definition a paying subscriber, and the unit economics may resolve profitably. But what about a freemium product whose user base is dominated by non-paying users? The proposition of a content treadmill to be used for re-activation may be less sustainable.

In the context of a chatbot, a major content release — akin to a high-profile new series being published on a streaming website — is a model update that may cost hundreds of millions of dollars in compute. And since most freemium chatbot users are non-paying, the monetization opportunity with a “content release” is diluted: it must win back churned users and convince them that the experience has improved to the extent that they should subscribe.

Returning to WAU vs. MAU. Because every WAU is also a MAU, then by definition MAU must be equal to or greater than WAU. But the WAU metric is more flexible than MAU, given its shorter duration. MAU is often reported either as an absolute count for a calendar month or as a rolling average over some lookback window, whereas WAU is almost always reported exclusively as the latter (I’ve never seen a weekly-published WAU metric).

So the WAU metric is more susceptible to reporting-window engineering than MAU. A company might time content releases (model updates) to drive re-engagement such that the rolling WAU growth rate is optimized. Because by definition MAU must be at least as large as WAU — but in a low-retention setting, it can be much larger — the MAU metric’s growth rate will benefit less from such updates, given the smaller increase relative to the baseline.

And given a large enough MAU/WAU ratio, reporting the MAU metric could reveal something else: that the product has essentially reached saturation. A large, implied MAU/WAU ratio might indicate that a WAU metric nearing 1BN corresponds to 2BN or more MAU. 1BN users leave enough headroom to extrapolate impressive growth; 2BN leaves decidedly less so. Assuming the maximum carrying capacity of a consumer app is 4BN — which reasonable people could disagree about — the question of saturation and potential growth becomes relevant at 2BN MAU.

This isn’t to say that WAU is by definition an illegitimate measure of a product’s user base. Just that the decision to use it instead of the more traditional MAU should be considered, as it may implicitly reveal fundamental, underlying dynamics about user base growth that aren’t obvious in a WAU trendline.

Comments: