How we review

Character controls (20%) asks which identity, personality, relationship and speaking-style settings are documented. Conversation options (20%) asks which text, model and voice modes are described, and whether their access units are clear. Visual controls (15%) asks whether a creator and its supported scene or identity inputs are established. Access limits (15%) asks what the plan permits and which operations consume an allowance.

Billing clarity (10%) asks whether currency, term total, renewal and separate credits are described. Privacy information (10%) asks whether visibility, retention and deletion information can be distinguished; it is not a security audit. Portability (5%) asks whether character notes or outputs can be exported or retained, without assuming that they can. Evidence quality (5%) asks whether the claim comes from an identifiable first-party source and whether sources agree.

The weights state which criteria would matter in a complete assessment. They do not turn missing results into a number. A future score of eight would require a documented, repeatable result with only a minor limitation; a five would describe partial task completion with a material limitation under the same test. We have no such recorded sessions in this edition, so no weighted total or star rating is published.

The shortlist order reflects the usefulness of available public information, not product performance. A documentation conflict can block a confident access claim without proving a feature fails. The service directory uses the same statuses across subjects, and each deep review preserves the distinction between advertised behavior and observed billing information.

“Advertised” means a first-party product or help source describes the capability. “Observed public price” records the currency and billing term visible in the source context. “Unclear” marks a conflict between public descriptions. “Not verified” marks a fact we could not establish. We do not use an illustration, filename or third-party review as proof of a provider’s output quality.

Sources checked on 26 September 2026 include Candy.ai subscription information, LoveScape’s help and official companion guides, GoLove’s terms and indexed first-party descriptions, and CrushOnAI’s safe-edition information. JOI AI’s catalogue, creator page, About, Safety, Terms, Privacy and Refund pages were also checked; monetary prices and account behavior were not independently tested. Our about page describes the publication, and the contact route accepts a source-based correction request.

Recheck subject identity, current plan totals, free limits, creative operation costs and conflicting documentation. Update the visible checked date, comparison values and modification dates together. A stronger performance claim requires a dated session record identifying the plan, task and raw results. We do not retrospectively label public-source comparisons as hands-on testing.