How we rank AI companion platforms
Most rankings in this category are an affiliate payout table with adjectives on top. Ours is arithmetic you can check: six criteria, fixed weights, sub-scores published on every review page, and an overall figure that is calculated rather than chosen.
This page exists so you can disagree with us precisely. If you think voice deserves more than 15% of the weight, you can recalculate the leaderboard yourself from the numbers we publish.
The six criteria
Each platform is scored from 0 to 10 on all six. The weights below never change between platforms, and they never change because of a commercial relationship.
Conversation
25%Coherence over a long session, consistency of persona, whether the companion drives the conversation or only mirrors it, and how much it remembers days later. The heaviest weight, because it is the product.
Visuals
20%Character consistency first, aesthetic quality second, generation speed third. A beautiful image of the wrong person scores badly. Video output, where it exists, is judged here too.
Voice
15%Naturalness, emotional range, latency, and whether live calls exist at all. A platform with no voice feature is scored on what it does offer, not given a zero — but it cannot score highly.
Customisation
15%Depth of the character creator, granularity of appearance and personality controls, how well the platform holds those choices over time, and the size and quality of the ready-made catalogue.
Value
15%Realistic monthly cost including typical token spend — not the advertised annual rate. Renewal pricing, free-tier usefulness and the honesty of the checkout all feed into this.
Trust & privacy
10%Operator transparency, jurisdiction, data handling, billing discretion, payment options and how hard it is to cancel. Reports of unauthorised charges hit this score hard.
The testing protocol
Every platform is tested from a paid account bought at retail. We do not accept press accounts, comped upgrades or review units, because a review account is not the product a reader will buy.
The same brief is used everywhere: one identical character description, one identical opening scenario, and three identical image prompts. Standardising the inputs is what makes the outputs comparable — otherwise you are comparing prompts, not platforms.
- A long single session to test coherence and whether the persona holds.
- A return session three to seven days later, cold, to test long-term memory.
- Three fixed image prompts generated in batches, checked for character consistency.
- Voice messages and, where available, a live call.
- Full checkout and cancellation, both timed and documented.
How the overall score is calculated
The overall score is the weighted mean of the six sub-scores: conversation 25%, visuals 20%, voice 15%, customisation 15%, value 15%, trust and privacy 10%.
It is computed at build time from the sub-scores, so the number in the ring and the numbers in the breakdown are mathematically the same claim. We display one decimal place but sort on the full value, which is why two platforms showing 8.3 can still have a definite order.
Independence and how we are paid
We earn a commission when a reader subscribes through one of our links. The commission rate differs between partners — some pay us more than double what others pay — and it has no input into the ranking.
The safeguard is structural rather than a promise: sub-scores are set from the test protocol before any commercial data is looked at, the overall score is computed from those sub-scores in code, and the ranking is a sort. There is no step in that pipeline where a payout rate could be applied even if someone wanted to.
We also rank platforms we cannot earn from, and we recommend them by name in alternatives where they are the better fit. If the two things ever conflict, the ranking wins — a comparison site that shades its scores is worth nothing to anyone, including its partners.
What gets a platform excluded
Not everything that exists deserves a listing. A platform is excluded, or removed after listing, for any of the following.
- A pattern of charges users did not authorise, or renewals at a price that was not disclosed at checkout.
- A cancellation flow that does not complete, or that requires emailing support to stop a subscription.
- Any content policy permitting material involving minors, real people who have not consented, or non-consensual scenarios presented as real.
- An operator that cannot be identified at all — no company, no jurisdiction, no way to reach anyone.
Corrections and updates
Prices in this category change constantly, so pricing is re-verified monthly and scores are reviewed every quarter or after a significant product change. The date shown on each page is when the underlying data was last verified, not when the text was last edited.
If something on this site is wrong, tell us and we will fix it and say what changed. That includes platform operators — a correction with evidence gets applied whether or not we have a commercial relationship with you.