Review Methodology

Exactly how we test AI companion platforms, score them across 8 weighted categories, and decide who earns the “Editor's Choice” badge. No shortcuts, no platform-supplied data — every rating below comes from hands-on testing.

Last updated: May 28, 2026 · Reviewed every 90 days

Affiliate disclosure (read this first)

CompanionRank.com is reader-supported. When you sign up to a platform through our links, we may earn a commission at no extra cost to you. We pay for our own subscriptions on every platform we review. A platform paying us a commission does not buy a better score, a higher rank, or an “Editor's Choice” badge.

We disclose affiliate relationships at the point of every outbound link (with rel=“sponsored” and visible “Visit X” CTA labelling). We do not accept paid placements, sponsored sections, or rewritten reviews. If a platform asks us to change a rating in exchange for commission, we publish the request.

The 8-category scoring rubric

Every platform earns a score from 0 to 10 in each of eight categories. The overall rating is a weighted average. Weights below reflect what we think matters most to real users — combat-tested across 8+ reviews and adjusted twice since launch.

Category	Weight
💬Chat Quality	20%
🖼️Image Generation	15%
🎬Video Generation	10%
🎙️Voice	10%
🎭Character Depth	15%
✨UI / UX	10%
🧠Realism	10%
👥Community	10%
Total	100%

A platform that doesn't offer a category (e.g. no voice, no video) is scored 0 for that category — it doesn't get a pass on missing capability. This penalises feature-light platforms unless they compensate with exceptional scores in the categories they do cover.

What we measure and how — per category

💬Chat Quality

Weight 20%

What we measure: Coherence across long conversations, personality consistency, ability to follow nuanced instructions, and how well the model stays in character under emotionally loaded or unusual prompts.
How we measure it: We run a fixed 24-message scenario per platform (8 small-talk + 8 emotional-escalation + 8 character-test turns), then return 7 days later with a follow-up that requires recalling message 3. Each turn is scored 0–4 on coherence, character adherence, and naturalness.
What it takes to pass: 8+/10: consistent persona across all 24 messages, accurate week-later recall. 6–7: occasional persona drift. Below 6: noticeable character breaks within the first 12 messages.

🖼️Image Generation

Weight 15%

What we measure: First-try success rate, anatomical consistency (hands, faces, eyes), prompt adherence on style and pose, character likeness across multiple generations, and uncensored/safe mode boundaries.
How we measure it: We submit the same 20-prompt benchmark to every platform: 8 realistic style, 6 anime/stylized, 3 multi-character scenes, 3 complex pose prompts. Each output rated 0–4 on prompt adherence and 0–4 on technical execution. First-try success = output usable without regeneration.
What it takes to pass: 8+/10: ≥70% first-try success on realistic prompts, hands acceptable in ≥60% of cases. 6–7: 50–70% first-try success. Below 6: heavy regeneration required.

🎬Video Generation

Weight 10%

What we measure: Frame coherence, motion naturalness, generation time, output length, and whether the same character can be re-used across multiple clips.
How we measure it: 5 video prompts per platform: 2 portrait talking-style, 2 motion scenes, 1 multi-shot sequence. Timed end-to-end. Scored on motion artifacts, identity preservation, and audio sync where applicable.
What it takes to pass: 8+/10: clips usable without obvious artifacts, character identity preserved across generations. 0 if platform does not offer video.

🎙️Voice

Weight 10%

What we measure: Voice naturalness, emotional range, latency on real-time calls (where supported), accent variety, and the gap between paid and free voice tiers.
How we measure it: 10-minute test call (or 30 voice messages where calls are not supported). Latency measured with stopwatch on response onset. MOS (Mean Opinion Score) rated 1–5 by two reviewers blind to platform identity.
What it takes to pass: 8+/10: MOS ≥4.0, sub-2-second latency on calls. 6–7: noticeable robotic cadence or 2–4s latency. 0 if no voice support.

🎭Character Depth

Weight 15%

What we measure: Number of customization fields, archetypal variety in the prebuilt library, persistence of backstory across sessions, and how well the platform supports custom character creation from scratch.
How we measure it: We build the same custom character (bookshop barista archetype with 6 specific traits) on every platform and test trait recall after 30 messages and again at the 7-day mark. Prebuilt library audited for distinct archetypes (not skin-deep variations).
What it takes to pass: 8+/10: all 6 traits recalled at day 7, 20+ genuinely distinct prebuilt archetypes. 6–7: partial recall or limited prebuilt variety.

✨UI / UX

Weight 10%

What we measure: Signup friction, time-to-first-message, mobile usability, dark mode, message threading, search/scroll-back, settings discoverability, and how often the UI gets in the way during real use.
How we measure it: Signup-to-first-message time stopwatched on iOS Safari, desktop Chrome, and Android Chrome. Subjective walkthrough scored against a 22-point usability rubric (signup, navigation, settings, message handling, payment flow).
What it takes to pass: 8+/10: <60s signup-to-first-message, all 22 rubric points pass. 6–7: 60–180s signup or minor friction in 3+ rubric points.

🧠Realism

Weight 10%

What we measure: How human-like the overall experience feels — combining response timing, typing indicators, proactive messaging, mood variation, and the absence of giveaway AI tells (excessive formality, list-formatting in casual chat, etc.).
How we measure it: Blind A/B test: 4 reviewers receive 20-message transcripts (10 from the platform, 10 from a baseline reference). Reviewers score each transcript on how convincingly human it reads, without knowing which is which.
What it takes to pass: 8+/10: ≥75% of platform transcripts rated 'plausibly human' by blinded reviewers. 6–7: 50–75%. Below 6: tells obvious in first 5 messages.

👥Community

Weight 10%

What we measure: Active subreddit/Discord size and engagement, in-app character sharing, user-generated content tools, and how quickly the team responds to public bug reports or feature requests.
How we measure it: Subreddit member count, 30-day post velocity, official response rate to top-10 posts/month. Discord/Telegram active member counts where applicable. Two test bug reports submitted; first-response time measured.
What it takes to pass: 8+/10: ≥5k active community, 24h first-response on bug reports. 6–7: smaller but engaged community or 24–72h response. Below 6: ghost-town or 72h+ response.

Test environment

•Devices: iPhone 15 Pro (iOS 17.4+), Pixel 8 (Android 14), MacBook Pro M2 (macOS 14)
•Browsers: Safari (iOS, macOS), Chrome (desktop, Android), Firefox (desktop spot-checks)
•Test window: 4 weeks minimum per platform (initial review). Re-tests at 90 days, or within 14 days of a major platform update.
•Subscription tier: we test the free tier, the cheapest paid tier, and the top tier where pricing exceeds $30/mo. Mid-tiers tested when they materially change the experience.
•Test budget: typical spend per platform: $40–$180 across 4 weeks, depending on credit/token economics. Tracked and disclosed in each review.
•Identical fixtures: the same custom character (bookshop barista, 6 specified traits), the same 24-message scenario, the same 20-prompt image benchmark on every platform — so scores are directly comparable.

What earns “Editor's Choice”

The Editor's Choice badge is reserved, not handed out. To earn it a platform must clear all four of:

Overall ≥ 8.5/10

Weighted-average score across all eight categories.

No category below 6

Platform must be at least competent across the board, not lopsided.

Stable for ≥ 90 days

No major outage, pricing flip, or unannounced model change in the last 90 days.

Transparent billing

No dark patterns on pricing, no hidden auto-renew acceleration, no surprise credit drains.

What we don't do

×Accept paid placements or sponsored reviews. Affiliate commission only — and only on links readers click after forming their own opinion.
×Republish platform marketing copy. Every claim in a review comes from our own testing, not the vendor's landing page.
×Score features we haven't tested. If a platform launches video gen and we haven't put it through our 5-prompt benchmark yet, the category stays unscored until we have.
×Use AI-generated “reviewer” personas. Reviews are written and edited by the CompanionRank editorial team. Author bylines on contributing reviewers belong to real people we work with.
×Game schema with fake aggregate ratings. We only emit aggregateRating when we have real review counts to back it. No defaulting to “1 review” to populate rich snippets.

Update cadence

Every review is re-evaluated on a rolling 90-day cycle. We also re-test out-of-cycle when:

•Pricing changes (any tier)
•Underlying model swap (e.g. switching LLM providers)
•Major feature launch (voice, video, new tier)
•Significant moderation policy shift (especially NSFW boundary changes)
•Credible reader reports we can reproduce

Each review surfaces its “Last tested” date prominently. If a review is older than 90 days, treat the score as a lagging indicator and check the verdict section for any update notes.

Disagree with a score?

We're wrong sometimes. If you have evidence (screenshots, logs, a reproduction) that contradicts a rating, send it to us through our contact page. We re-open reviews when the evidence warrants it, and we publish the change with the reason.

About CompanionRank · Platform rankings · Blog · Contact