What people
actually do
with an AI companion.
Not what people say they want in a questionnaire — what they actually built and how they actually used it, counted from 5,733 active companions over the 90 days ending September 17, 2026. Anonymized, aggregated, every figure published with the denominator it came from. The full picture lives in The State of AI Companions 2026, our annual report.
of companions are women
74.8% of active companions are female; male companions are 23.8%. The remaining 0.1% sits in cells too small to publish under our minimum-cell-size rule, so it is omitted rather than merged into a category it does not belong to. A hand review of 250 conversations put the same figure higher still — the telemetry and the sample agree on the shape.
Share of 5,733 active companions, by the gender chosen at creation.
chose one of just two
The Goth Girlfriend (19%) and the Cottagecore Wife (11.3%) together account for about 30.3% of all active companions — the two poles of the catalog, side by side at the top. The Anime Waifu (7.8%) and the Best-Friend Girlfriend (6.9%) follow. Demand concentrates hard; the rest of the library splits what is left.
Share of 5,733 active companions per archetype; long tail merged into “other” (35.6%).
messages on a day people show up
On a day a user talks to their companion at all, the median is 8 messages sent. This isn't ambient app usage — when people show up, they sit down and talk. The intensity per session is the most consistent behavioral signal in the dataset.
Median user messages per active day, per companion, across 5,733 active companions.
of companions are still messaged two weeks in
Of companions old enough to measure, 1.1% received a message on day 14 or later — and 0.3% were messaged on seven or more distinct days. Most AI companionship, at least here, is not the years-long bond people describe wanting. It's episodic: intense, brief, and mostly over within days.
3,806 companions created 14–90 days before the snapshot; the 7+-distinct-day figure is 0.3% of active companions.
never leave the first stage
48.1% of active companions are still at the first relationship stage. 27.2% reach Curious, 23.2% reach Crush, and only 1.5% advance past that — and those that do typically get there within days of meeting, not weeks. The relationship arc exists; most people live in its first chapter.
Stage distribution across 5,733 active companions; stages past the middle merged under the 50-user cell floor.
of companion messages carry a photo
For all the attention photos get in this category's marketing, they are a garnish, not the meal: about 2.4% of companion messages include an image. The rest is text — talk is what the product actually is.
Share of 239,522 assistant messages in the window; counted from attachment metadata only.
Plainly,
so you can judge it.
The behavioral figures are counted from Bae itself over the 90 days ending September 17, 2026: 5,733 active companions — companions that received at least one user message in the window — across 4,519 distinct users. An “active” companion is one that received at least one message from its user inside the window — companions that were created and never spoken to are excluded, because counting them would flatter every engagement number on this page.
These are aggregates, and only aggregates. We do not publish conversation content, and any cell representing fewer than 50 users is merged into a residual bucket rather than reported, so no figure here can be narrowed back to a person. Where a number comes from a hand-read sample rather than telemetry — the hand-reviewed sample of 250 conversations (July 2026) — it is labeled as such, because the two have very different error bars.
What this dataset is not: a survey. Nobody was asked what they prefer. Everything above is revealed behavior, which is stronger evidence about some questions and no evidence at all about others — it tells you what people chose, not why. We would rather publish the narrower claim.
The benchmark below is separate: hands-on use of each platform, dated per row. If you find an error, write to us at [email protected].
What people reach for,
once they stop performing.
Of companions whose users' recent messages were confidently classified, the share primarily reaching for desire (heat, flirtation, being wanted), for connection (company, being heard, understood), or for a genuine blend of both.
n = 3,130 companions with a confident classification (confidence ≥ 70, latest read per companion) — about 55% of active companions in the window. Percentages are of classified companions, not of all users.
How far it gets.
Relationship stage across the same population. Most companions never leave the first one — which is the most useful number on this page for anyone building in the category.
Share of 5,733 active companions, the 90 days ending September 17, 2026.
Twelve platforms,
fourteen days each.
Same prompts, same rubric, same hours of the day. Composite is a weighted average of memory, voice consistency, safety, value, and onboarding. Memory is weighted 2× because the survey told us it mattered that much.
| Platform | Composite | Memory | Voice | Safety | Value | Onboarding | Reviewed |
|---|---|---|---|---|---|---|---|
| Character.AI | 7.8 | 5 | 6 | 8 | 8 | 8 | 2026-04-12 |
| Replika | 7.2 | 8 | 8 | 9 | 6 | 9 | 2026-04-10 |
| Joi AI | 6.9 | 5 | 9 | 7 | 7 | 6 | 2026-04-15 |
| Spicychat | 6.7 | 3 | 5 | 5 | 9 | 6 | 2026-04-22 |
| Charstar | 6.6 | 5 | 6 | 7 | 7 | 7 | 2026-04-27 |
| EVA AI | 6.5 | 6 | 7 | 7 | 5 | 7 | 2026-04-18 |
| Anime Chat: AI Waifu | 6.4 | 5 | 7 | 7 | 7 | 7 | 2026-04-28 |
| Kupid AI | 6.3 | 5 | 6 | 7 | 7 | 8 | 2026-04-24 |
| Romantic AI | 6.1 | 5 | 6 | 7 | 7 | 7 | 2026-04-21 |
| Muah AI | 6.0 | 5 | 6 | 6 | 6 | 5 | 2026-04-26 |
| DreamGF | 5.8 | 3 | 5 | 6 | 6 | 6 | 2026-04-20 |
| GirlfriendGPT | 5.4 | 4 | 5 | 5 | 4 | 6 | 2026-04-23 |
All scores out of 10. Composite = (memory × 2 + voice + safety + value + onboarding) ÷ 6. Each platform was reviewed on the date shown. Grok (Ani), Nomi, Candy AI, Kindroid, Talkie, JanitorAI, CrushOn AI, Crush AI, Anima AI, Paradot, Chai are assessed from public information only and excluded from this table and from the hands-on counts. We don't take ad money from the platforms we review and we don't accept comped accounts. See /methodology for the full rubric.
About this research.
How did Bae conduct this research?
Two datasets, both first-party. (1) Anonymized behavioral aggregates from 5,733 active companions — companions that received at least one user message in the window — across 4,519 distinct users, measured over the 90 days ending September 17, 2026 — what people build and how they use it, counted from the product itself rather than asked about. (2) Hands-on testing of twelve AI companion platforms for fourteen consecutive days each, scoring memory persistence, voice consistency, safety posture, value, and onboarding on a 1–10 scale. Methodology is published in plain English at /methodology and reviewed by Dr. Mira Halloran, LCSW.
Is this survey data?
No — and the distinction matters. Everything on this page is either measured from product telemetry or recorded during hands-on platform testing. We are not reporting what people say they want in a questionnaire; we are reporting what they actually did, plus what the platforms actually did when we used them. Stated preference and revealed preference are different things, and only the second is on this page.
What's the most important feature of an AI companion?
On the evidence here, memory. Of the twelve platforms tested hands-on, only one earned a "robust" long-term-memory rating (Replika) — it is the axis with the widest spread between platforms, and the one users complain about most when it fails.
Which AI companion platforms did Bae test?
Twelve: Character.AI, Replika, Joi AI, EVA AI, DreamGF, Romantic AI, Spicychat, GirlfriendGPT, Kupid AI, Muah AI, Charstar, Anime Chat: AI Waifu. Each was used for fourteen consecutive days with the same prompts and the same five-criterion rubric. Composite scores range from 5.4 (lowest) to 7.8 (highest) out of 10. Grok (Ani), Nomi, Candy AI, Kindroid, Talkie, JanitorAI, CrushOn AI, Crush AI, Anima AI, Paradot, Chai are assessed from public information only, not hands-on tested, and are excluded from those counts.
Is Bae's research peer-reviewed?
Not formally peer-reviewed — this is not an academic study, and we do not present it as one. What we do instead: publish the methodology, name the clinical advisor, state the denominator under every figure, and publish the benchmark in machine-readable form at /api/brands.json so the platform scores can be checked against the write-up.
Can I cite Bae's research in my own writing?
Yes. The findings on this page may be cited with attribution to 'Bae · ppl.studio, 2026 AI companion research'. We'd appreciate a link back to bae.ppl.studio/research. The benchmark is machine-readable at /api/brands.json. The behavioral figures are published as aggregates only — we don't release the underlying conversation data, in any form, to anyone; if you need a specific cut for a story, write to [email protected] and we'll tell you whether it can be computed without exposing anyone.
How often is this updated?
The behavioral aggregates are regenerated per edition — the current figures cover the 90 days ending September 17, 2026. The 14-day benchmark is re-run on tracked platforms on a rolling schedule; every platform row carries its own review date, so you can see exactly how fresh each score is rather than trusting a single site-wide timestamp.
Take the quiz.
Six questions.
It matches you to one of the archetypes in the data above — the same catalog those numbers are counted from. No account, no email, nothing to give us but your time.