Why Nobody Can Tell You How Many People Use Companion Apps
There is no reliable figure for how many people use AI companion apps, and you will not find one here. That is not evasion — it is the finding. Two things make the question unanswerable as posed: nobody agrees on which products belong in the category, and every metric that is actually available counts something other than people. Once you can see both problems, the numbers circulating become easy to read, and easy to stop repeating.
The category has no boundary
Before anything can be counted, someone has to decide what is being counted. In this case, several defensible decisions are available and they produce wildly different totals.
Are general-purpose assistants included? Enormous numbers of people talk to general chatbots for company, advice, and conversation without ever installing anything marketed as a companion. If that counts, the total is one thing. If not, another.
Are character and roleplay platforms included? Many users treat them as creative writing tools, some as companionship, and most as a mixture. The user’s purpose is not visible in any metric.
Are adult-oriented products included? They are frequently excluded from mainstream tallies and frequently large.
Are messaging-platform bots included? A companion running inside another app leaves no separate install to count at all.
What about someone who installed one once, out of curiosity? Almost every count includes them.
Whoever chooses the boundary chooses the total, and the boundary is usually set by someone with an interest in the answer being impressive.
Every available metric counts something else
Even with a fixed boundary, the underlying numbers do not measure people.
Downloads are not installs. The same person downloads on a phone and a tablet, reinstalls after a reset, and tries the app twice.
Installs are not accounts. Plenty of installs never get through signup.
Accounts are not users. People make several, abandon them, and never delete them. Accounts accumulate permanently and only ever go up.
Accounts are not active accounts, and every company defines “active” for itself — a definition that can be a day, a month, or “has ever returned”.
Active is not paying, and paying is not engaged.
Some accounts are not people at all. Automated signups exist wherever there is a free tier.
And a single person may span several products, so adding two companies’ figures together double-counts by an unknown amount. This is the error that makes category totals almost meaningless: they are usually sums of numbers that overlap.
Who is in a position to know, and why they don’t say
Operators know their own figures and publish them selectively, at moments chosen for effect — a funding announcement, a launch, a press push. A company disclosing its own metric is choosing which metric to disclose, which is a form of editing that requires no dishonesty at all.
App-intelligence vendors sell estimates, built by modelling from store signals, panels of consenting devices, and observed rankings. These are models. They can be reasonable and they are not counts, and their error bars are rarely reproduced when the figures are quoted onward.
Survey firms ask samples, which means self-report about a behaviour people have reason to understate. Anything intimate or slightly embarrassing is underreported, by an amount nobody can quantify.
Nobody has an independent, comprehensive view, and no public body collects one.
Why company-reported engagement is not comparable
Two companies can both be truthful and still produce numbers that cannot be placed side by side.
Each defines the window. Daily, monthly, or since-signup.
Each defines a session. A single message, an app open, or a minimum duration.
Each decides what counts as the product. Web and app, free and paid, one region or all.
Each chooses when to measure. After a marketing push rather than before.
None of that is fabrication. It is why comparison tables assembled from company statements are not comparisons of anything.
What a figure would need before it was worth repeating
A short test, and very little survives it.
A stated definition of what is being counted.
A stated boundary for which products are in scope.
A stated method, including how the data was collected and from whom.
A collection date, separate from the publication date — figures in this area are often much older than the article carrying them.
Disclosure of who paid for it.
A named, reachable source, rather than a citation to a summary of a summary. That laundering process is the subject of how to read a statistic about this category.
If a number arrives without those, the honest reading is that you have learned nothing about usage and something about the publisher.
What you can observe without any of this
That the products exist and are actively developed, which tells you somebody is using them.
That store charts place them relative to each other, which is a ranking and not a count, and is influenced by promotion.
That communities around individual apps are visibly populated, which tells you about the most engaged users and nothing about the total.
That new products keep appearing, which is a signal about investment rather than about users.
That is genuinely all a reader can establish, and it is enough for every decision an ordinary reader actually faces. The size of the category has no bearing on whether a particular app handles your data acceptably or whether your own use of it is doing you any good.
Why the question gets asked at all
Usually not out of curiosity. A large user figure is the premise that makes the category significant — for a journalist deciding a story is worth writing, an investor deciding a market is worth entering, or a company arguing it is now mainstream. The same function is served by the market projections examined in how a market-size figure gets made, and by the societal framing described in the loneliness-epidemic framing.
Noticing what a number is for is the fastest way to work out how much weight it was ever meant to carry.