How to Read a 2026 Statistic About Companion Apps
Statistics about AI companion apps circulate widely and hold up poorly. There are none on this page, deliberately: repeating a figure whose method I cannot see would make this site another link in the chain that lent the figure its air of authority. What is useful instead is a sequence of checks you can run on any number you meet, in the order that eliminates the most material fastest. Most claims in this area fail the first two, which takes about a minute.
Is there a named source you can reach
Follow the number backwards. This is the single highest-yield check and most people skip it.
Named organisation, dated publication, reachable document: proceed to the next check.
Attributed to “a study” or “research shows”: stop. There is nothing to evaluate.
Attributed to another article: keep following. In this category the trail frequently ends at a company announcement, which is a claim rather than a source.
Attributed to a vendor report you cannot open: the figure is the advertisement for the report. Whether the method behind it is sound is not something you or I can assess, and that is the point of publishing it that way.
Who was counted, and how were they found
A number about people is meaningless without knowing which people.
Was there a defined population? Adults in one country, app users, subscribers to a panel, visitors to a website.
How were they recruited? Through the app itself, through a survey panel, through social media, or through an intercept on a website. Each of those selects a different group, and none is the general public.
How many were there, and was the group large enough to support the breakdown being reported? Sub-group claims are frequently built on a handful of responses.
Was participation voluntary and unpaid? Both introduce different distortions, and neither is disqualifying if disclosed.
If the answer is “users recruited from the product’s own community”, the number describes enthusiasts. That is real information, described inaccurately whenever it is presented as being about the category.
Is it self-reported, and about what
Almost everything in this area is self-reported, which is unavoidable and worth pricing in.
Anything intimate is underreported. People understate use of products they find slightly embarrassing, by an amount nobody can quantify.
Anything flattering is overreported. Frequency of exercise, hours of reading, and intentions to cancel all suffer the same way.
Reported feelings are not measured outcomes. “Felt less alone” is a description of an impression at the moment of asking. It is not a finding about a state, and treating it as one is the standard slippage in wellbeing claims — the study-design version of this problem is set out in how to read a claim about psychological effects.
What is the number actually a number of
Precision about the denominator disposes of a surprising amount.
A proportion of what? Of everyone surveyed, of those who answered that question, of app users, or of users who had already reported doing the thing being asked about. These produce very different-looking figures from identical data.
Users, accounts, downloads, or sessions? These are routinely swapped, and they are not interchangeable — the reasons are set out in why nobody can tell you how many people use companion apps.
A rate or a total? And over what period.
Compared with what? A figure presented as a rise needs a stated baseline, measured the same way. Frequently the earlier number came from a different method entirely, which makes the comparison an artefact.
When was the data collected
Publication date is not collection date, and the gap is often substantial.
A figure republished under this year’s headline may rest on data gathered well before it. In a category where products change every few months, that matters more than usual.
A year in a headline is a freshness signal, not a freshness guarantee. It frequently indicates the article was written this year, not that anything in it was measured this year.
Who paid, and who benefits
Commissioned research is not worthless and it is not neutral. The question set, the population, and which findings get promoted are all shaped by whoever paid.
Check whether funding is disclosed at all. Undisclosed sponsorship is a much worse sign than disclosed sponsorship.
Ask what the figure is for. Establishing that a category is large, that a problem is urgent, or that a product is working — each has an obvious beneficiary, and noticing it is ordinary reading rather than cynicism.
Count, estimate, or projection
Three different kinds of thing, presented identically.
A count is something somebody actually tallied. Rare here.
An estimate is a model. Reasonable estimates exist; they come with assumptions and, when honestly presented, with a range.
A projection is an assumption about the future applied to an estimate. It is the weakest category and the most quoted, and how it gets built is described in how a market-size figure gets made.
Tells that a figure is decorative
Improbable precision on a quantity nobody could measure that closely.
A round, memorable number that appears everywhere in identical form.
No unit or population attached — just the number and an implication.
A percentage with no denominator anywhere in the article.
Several incompatible figures in the same piece, which indicates the writer collected numbers rather than checked them.
If a figure survives all of that
Use it, and carry its caveats with it: source, population, method, date. A well-sourced figure stated with its limitations is genuinely informative, and stating the limitations costs one clause.
And if none is available on a question you care about, the honest move is to say so rather than to reach for the nearest number. That is why this page has none.