Clear sources. Honest conclusions.

Discovery comes first.

The collection starts with public character-platform directories, product lists and searches in seven languages. The first systematic expansion read five directories and collected 294 listing observations. Those observations were combined with an earlier manual catalogue and individual search findings.

A directory mention is evidence that a platform was found. It is not sufficient evidence that a current feature works, that an outbound link is official or that a subscription is good value.

Three evidence states.

• Official sources: a platform’s public page or documentation describes character creation. No account-level performance test is implied.
• Research listing: the platform has been discovered, but creation controls, official identity or current access still need verification.
• Archived: the record is retained for historical context and excluded from the active directory.

Per-feature comparisons require separate evidence. An official source about the creator does not automatically verify voice, imagery, group scenes or sharing. “Not checked” is deliberately used instead of “not supported.”

Check the destination, too.

Names and URLs are normalized, but conflicting links are retained for review. During this collection, one directory’s Janitor entry pointed to Lovescape. The destination was not accepted as Janitor’s official site. Similar names and partner redirects do not establish a rebrand.

Sources are shown on each platform page. Unknown domains stay unknown. An inaccessible page is not automatically treated as a closed product.

Keep each check’s scope and date.

The review history and source register record which fields were clarified, which sources were reread and which questions remain open. An entry revision, a source read, a publisher update and a hands-on test are different events. Updating one field does not replace all older read dates.

An access attempt can have a recent date while every capability stays unverified. When supporting sources disagree, we retain the statements and their effect on a decision until their scope is established.

Characters, with context.

Our character collection combines public profiles with clear use cases, characters featured in a platform’s own collection and names linked from official updates. Each card identifies its source and snapshot date. Creator credits are shown when the source names them; a platform collection is not treated as an individual author credit.

Public interactions and Chats describe activity, not unique users, conversation quality or safety. Likes, Favorites and Downloads are separate counters. Chub’s labeled page-title values are identified separately from its rounded icon labels. Anonymous is a publication label, not a verified person. Rounded profile labels and individual observation dates are retained; catalogue and profile values can differ. Counts are not compared across platforms. We do not infer what an unlabeled counter measures. Placement in this collection is an editorial choice, not a hands-on recommendation or a universal popularity ranking.

“Seen in Trending · RPG” records a dated SpicyChat catalogue observation with that filter selected. “Editorial selection” identifies profiles chosen from PolyBuzz’s RPG tag page without claiming a popularity order. Discovery links preserve that context. Their unlabelled counters are omitted. Starting scenes and evaluation questions are original Characters.best exercises; they are not copied character definitions or tested replies.

Portraits are local snapshots of publicly displayed avatars, with neutral crops for this editorial layout. Missing portraits are identified explicitly. Profiles and availability may change after the snapshot; no private accounts or conversations were accessed.

Web presence is one piece of the picture.

Our platform insights use the Majestic Million dataset, updated 3 Oct 2026. It ranks domains by the number of distinct referring subnets. Linking networks and IP addresses come from that same snapshot. Rank movement compares the current and previous index positions; the source does not identify a previous observation date in the CSV.

We match the exact hostname, normalize “www” and never substitute a parent domain for an unlisted subdomain. Repositories, app-store listings and shared generator hosts are excluded. Measurements cover the whole matched domain, including other products it may host. A domain outside this million-entry index has unknown traffic; it is not assigned zero visitors. Domain presence does not change a platform’s editorial evidence status.

The dataset is selected and reformatted by Characters.best under CC BY 3.0, with source attribution. It contains no monthly visits, unique users, visitor countries, age or gender. There is no automatic refresh during a visit.

Platform pages link to Semrush’s free Traffic Checker for external traffic estimates and Google Trends for relative search interest. Semrush describes one free daily check without an account; additional access may need registration or payment. We do not scrape or republish its reports. Google Trends scales interest from 0 to 100 within the selected query, period and location. It does not supply absolute searches or website visitors.

Platform passports and matching

The platform finder compares your requirements with recorded workflows and sourced passport fields. Known requirements appear with an explanation and a link to the relevant evidence. Missing evidence stays open; it does not count as a supported feature.

Documented matches have sourced passport fields and no outstanding checks for the requirements selected. Research candidates have gaps. Their direction comes from the editorial research catalogue or the documented workflows; it does not establish suitability or quality. Results put documented matches first, then candidates with fewer gaps, then sort by name. Traffic and commercial relationships do not affect this order.

Monthly budgets use USD offers billed each month, with both the first payment and a verified renewal within the ceiling. Annual or quarterly equivalents are not treated as monthly payments. An unknown renewal stays a research gap. Extras, taxes and regional differences remain outside the filter. A free starting point may be a limited allowance, free creation or a free interface requiring a separate model backend.

Memory fields describe saved-context controls, not measured recall. Private character visibility is separate from account-data processing. Language performance, output quality and plan entitlement are not tested. The 18+ filter hides explicit labels; it is not an assessment of unmarked platforms. Requirements are retained in the page URL, without an account or external matching API.

How we document prices.

Pricing notes identify the public source, date read, currency, billing interval, introductory amount, renewal amount and included credits. Public offers are not completed purchases. Regional offers and account-level allowances can differ. The date read does not mean the provider updated its article on that date.

Cost comparison counts payments required to cover the selected period, beginning with a new subscription. A yearly plan is paid upfront even for a one-month comparison. An introductory payment is used only once. Renewals at the end of the selected period are excluded; incomplete billing periods still require a full payment. Future charges reuse a quoted amount when a separate renewal price is unknown, with that assumption stated in the result.

The calculation uses published USD offers, without currency conversion. Extra credits, character add-ons and taxes are excluded. Missing prices stay unknown. Token, Chip and Neuron balances are not interchangeable or converted to a number of images without a documented per-action cost. We have not tested checkout or cancellation inside an account.

What a hands-on test will add.

A future product test should record the account tier, exact task, date and output. Character creation, conversation behavior, image consistency, voice, privacy controls and checkout should be evaluated separately.

We have not completed those tests in this research edition. Prices, discounts, memory quality and paid-plan performance are not assigned invented values or scores.

How the content was prepared.

This edition uses AI-assisted discovery, structured extraction and source-based writing. Platform guides combine public product descriptions with original evaluation questions. They are not testimonials from invented users or claims by a fictional testing team.

Platform discovery began on 3 October 2026. The latest catalogue observations are dated 4 October 2026; individual guide and source checks retain their own dates. The directory retains open questions and does not claim exhaustive coverage.

Explore the sourced directory or read our commercial disclosure policy.