AI Companion Engagement Metrics: What to Measure Yourself

Updated 2026-10-05. AI girlfriend guides

We do not have industry-wide engagement data, and we have not published engagement data of our own. This page previously carried specific figures — average session length, daily engagement share, feature adoption rates — plus a "well-being sweet spot" of 15 to 45 minutes per day, attributed to a study in JMIR. We could not locate that study. The figures have been removed rather than replaced with new guesses. What remains is the framework: what each metric means, how it is computed, where it misleads, and what we can and cannot tell you about our own product.

Key takeaways

Session length, session frequency, message volume, feature adoption, and daily-to-monthly engagement are standard product metrics with precise definitions. None of them is technically difficult. All of them are easy to quote badly.
The honest version of most engagement claims is a definition plus a question, not a number.
Engagement volume concentrates in a minority of users in most consumer products. We have no AI-companion-specific figures and will not imply otherwise.
We publish no engagement figures of our own, and the product does not show you a usage dashboard.
There is no established "healthy amount" of AI companion use. Any claim that a particular number of minutes per day produces better well-being is unsupported.

What each metric actually means

Session length is the wall-clock time between a user's first and last message in a single sitting. Session frequency is how many such sittings occur per week or month. Together they are the most quoted engagement pair in this category, and they are the two most context-dependent.

AI companion sessions are structurally unusual in a way that makes the average nearly meaningless. There is no natural session boundary and no natural stopping cue — the user decides when to stop, and the stopping point is arbitrary. A user who sends three messages over six hours and a user who sends six messages in six minutes produce identical message counts and wildly different session lengths. Any figure for "average session duration" in this category should be treated as a property of the platform's interface design as much as of its users.

Messages per session and messages per day are volume counts, and they inherit the same problem: a message is not a unit of meaning. Users who write short paragraphs and users who write one-word replies are counted identically.

Message length is usually reported as a user-to-AI word ratio. AI responses are structurally longer than user messages, because the system prompt asks for emotional validation, callbacks to shared history, and elaborated context. That is a prompt design decision, not a finding about users. We are not aware of a citable dataset for this ratio in AI companion products.

Feature adoption is the share of users who have used a given feature at least once. It is conditional on the feature existing, so cross-platform comparisons are close to meaningless: a platform that only offers text messaging cannot have low voice adoption for a good reason. "Adoption" also hides the denominator problem — a feature most users cannot discover will show low adoption for reasons that have nothing to do with demand.

Daily-to-monthly engagement (DAU/MAU) is the share of monthly active users active on a given day. It is the single most useful retention-shaped metric in consumer products, and it is the one most often quoted in this category without a definition of what counts as active.

The three ways engagement claims mislead

Confounding by design. "Users who adopt two or more features retain better" is a nearly tautological observation. People who are already engaged try more features. The direction of causation runs from engagement to adoption at least as often as the reverse, which makes the "drive feature adoption in onboarding" strategy unproven by the very correlation usually cited to justify it.

Survivorship. Metrics computed on the current user base exclude everyone who left. Retention curves and engagement distributions computed over active users only will always look healthier than the ones computed over a signup cohort.

Silent definition changes. Whether a user who has not returned for a week has "churned" is a choice, not a fact. Whether multiple accounts belong to one person is unknowable from inside any single platform. Any cross-platform engagement comparison is comparing different definitions of the same word.

Volume concentration

In most consumer products, message and session volume is heavily concentrated: a minority of accounts generate a large share of total activity, and the average user is not a meaningful unit of description. We have no AI-companion-specific distribution, so we are not going to imply one, and we would treat any specific "heavy users account for N percent of messages" claim you encounter as unsourced unless it comes with a named dataset.

This has a practical consequence for reading any average. A category average describes the average of a skewed distribution, which means it describes almost no individual. If you care about a typical user rather than an average one, the median is the figure you want, and it is rarely published.

Engagement and well-being

There is no established threshold for how much AI companion use is healthy, and the specific claims that circulate here — a 15-to-45-minute daily sweet spot, a U-shaped curve where heavy users do worse, a "healthy maximum" — trace back to a study we could not locate. They should not be repeated.

The real research on the wellbeing question is thin, and it does not support a usage threshold.

A preregistered 21-day randomized trial assigned 183 participants either to chat with Replika or to play word games as an active control. It found no significant difference in social health, loneliness, or effect on relationships with family and friends, in either direction (arXiv 2509.19515).

The longitudinal evidence points the other way, and it is the more uncomfortable strand. In a two-wave study of 1,182 Character.AI users, with 439 responding at roughly 12 months, sustained engagement predicted lower psychological well-being, and the pathway the authors identified was reduced in-person interaction with other people (Zhang et al., arXiv 2609.07243).

Note what neither study gives you: a threshold, a dose-response curve, or a safe daily limit. A trial of three weeks cannot detect a cumulative effect, and an observational study that identifies displacement as a pathway cannot tell you how much use is too much. Anyone who tells you the number is making it up.

What you can measure about your own use

You do not need a company's dataset to know how you are using a product. These are observable without instrumentation:

Volume and time. Your own message counts and timestamps are in your history. The app surfaces your recent conversations; you do not need a dashboard to know whether you open it daily or twice a month.
Feature coverage. Which features have you actually used, and which did you set up and never touch.
Timing. When you reach for it, and what precedes that.
Displacement. The only measure that matters and the only one no dataset will give you: whether sessions have been replacing time with people, or sitting alongside it.
Stop reasons. What actually made you cancel, unsubscribe, or go quiet — a reason only you have.

If you want the analytical version of this, the useful artifact is a private record you keep yourself. It does not need to be rigorous to be informative.

What our product can and cannot tell you

Cannot: We publish no engagement figures, we have no comparable cross-platform dataset, and we do not surface per-account usage analytics in the product. If you are looking here for a benchmark, there is not one.

Can: We can tell you what the product does — what it stores, where your conversations live, and how they are deleted. That is a product question rather than a metrics question, and it is the one we can answer precisely.

Frequently Asked Questions

How long do people spend with AI companions daily?

We do not have a figure for this category and are not going to supply one. On your own usage, your history contains the timestamps; the honest average is heavily skewed by a minority of heavy users, so the median is the more useful number.

What engagement metrics are worth tracking?

For a product: session length and frequency, messages per session, feature adoption among users who can see the feature, and DAU/MAU. For yourself: volume, timing, and displacement. The last one is the only one any of these datasets measures badly.

Do users adopt more features over time?

More engaged users try more features, and the causation runs both ways. This is a well-known confounding problem in product analytics rather than a demonstrated growth lever.

Is there a healthy amount of AI companion use?

No threshold has been established. The 21-day randomized trial found no significant effect on social health or loneliness versus a control, and the 12-month observational study found sustained engagement predicted lower well-being through reduced in-person contact. Neither identifies a safe limit.

Does engagement correlate with wellbeing?

The only credible evidence available runs from engagement toward lower wellbeing, via displacement of in-person contact, and it is observational. There is no evidence for the U-shaped curve this page previously described.

This article is for informational purposes only and does not constitute medical or professional advice. Luvlock AI provides tools for companionship but encourages maintaining a balanced approach to relationships and mental health.

Ready to experience AI companionship for yourself? Get started with Luvlock AI →