How we score
Four pillars, written anchors, numbered sources, and a plain statement of what has and has not been tested.
What a score is
Every tool gets a paper score out of 10. It is built from four pillars, each marked out of 10 against written anchors and then weighted. It measures what a company has committed to in writing and what is on the public record.
What a paper score is not
It is not a verdict on how good the product is to use. We have not used these products, and a document cannot tell you whether a companion remembers your name after a fortnight. Where a review is a desk review, it says so at the top. Where a pillar could not be assessed, the score is marked provisional and the pillar is shown as n/a.
The anchors
Privacy, 30% of the score
What the policy lets the company do with your chats: train on them, keep them, share them, read them.
| 9 to 10 | The company says it cannot read your content: it is encrypted on your device or not stored. |
|---|---|
| 7 to 8 | No retention or strong encryption, clear deletion, few third parties. |
| 5 to 6 | Encrypted in storage, limited or de-identified training, deletion available, some ad tracking. |
| 3 to 4 | Broad rights to train on chats, long retention, deletion by email, or a policy silent on chats. |
| 1 to 2 | Sensitive data sold or shared, no deletion, or an unresolved regulator finding. |
Price clarity, 25% of the score
Whether you can see the real cost before you sign up, and how refunds and renewals are written.
| 9 to 10 | Full price list on the open web, flat pricing, fair refunds. |
|---|---|
| 7 to 8 | Full public price list; flat or nearly flat; standard no-refund terms. |
| 5 to 6 | Public prices but core features metered, or an intro price that steps up; or a simple plan that is not printed on the company's own pages. |
| 3 to 4 | No public price list, stacked add-ons, or overlapping plans. |
| 1 to 2 | Hidden prices plus consumable credits plus documented inconsistency. |
Track record, 25% of the score
Fines, breaches, lawsuits and sudden changes on the public record, and whether the operator is named.
| 9 to 10 | A long record with nothing against it and a fully named operator. |
|---|---|
| 7 to 8 | No regulator action or breach found. 8 if the operator is named with an address, 7 if not. |
| 5 to 6 | Documented controversies, or one significant incident. |
| 3 to 4 | A regulator's fine, or a pattern of abrupt changes for paying users. |
| 1 to 2 | A breach handled badly, or repeated enforcement. |
Rules and age checks, 20% of the score
Whether the limits are written down, including on minors and real people, and whether adults are verified.
| 9 to 10 | Detailed published rules, explicit bans on sexual content with minors and real people, privacy-preserving age checks, appeals. |
|---|---|
| 7 to 8 | Most of the above, with one gap. |
| 5 to 6 | Published rules with several gaps, or an age check only where the law demands it. |
| 3 to 4 | Thin or no content rules, or no age check described. |
| 1 to 2 | No published rules and no age requirement worth the name. |
Where the facts come from
- First, the company's own pages: pricing, privacy policy, terms, content rules, help centre.
- Then regulators, courts and named research groups.
- Then reporting from named outlets.
- Prices a company does not publish are marked "reported" and attributed to the outlet that reported them.
- Every claim carries a numbered source. Every page carries the date its paperwork was last checked.
How hands-on testing works
A tool is marked "Tested" only when someone has paid for a plan and followed the test script for 14 days. The script covers sign-up and checkout, a memory test with planted facts, the tool's behaviour at its own stated limits, voice and image features as advertised, cancellation, and a deletion request. It produces five marks out of 10:
- Conversation quality
- Memory over 14 days
- Features as advertised
- Reliability
- Account control
The hands-on score sits beside the paper score. We do not blend them, because they answer different questions.
What we do not list
Tools built to sexualise real people. See deepfakes and consent. Tools with weak age checks are listed and marked down for it.
Updates and corrections
Prices and policies in this category change monthly. Each page shows when its paperwork was last checked. If you find something out of date, the contact details are on the About page, and corrections are made with a dated note.