How we score

How the total is calculated

Multiply each sub-score by its weight and add them up. That is the whole calculation. It runs in code at build time, in weightedTotal(), and the ranking is produced by sorting on its output. There is no step where an editor assigns an overall score.

This matters because it removes the most common failure in this category. If we wanted SweetDream AI to rank higher, we would have to raise one of its visible sub-scores — on a page that shows you the sub-scores — rather than quietly adjusting a headline number. You can check every total on this site with a calculator.

Why these weights, for a UK reader

Weighting is an editorial judgement, and ours is explicitly built for someone in the UK. The two decisions worth arguing with:

Privacy and jurisdiction carries 20%, which is high for this category. The reasoning: the practical difference between complaining to the ICO about a company registered in London and pursuing a company you cannot identify in a country you cannot determine is enormous, and it only becomes apparent on the day you need it. Four of the ten apps reviewed do not name an operating company anywhere on their site. We think that should cost them.

Conversation and memory carries 25%, the most of any criterion, because it is the thing that determines whether you are still using the app in three months. Image quality is noticeable immediately; being remembered is noticeable on day thirty.

If you weight things differently — and you might reasonably care far more about image generation than about data protection — the single-criterion rankings let you ignore our weighting entirely and sort on whichever one you care about.

What this evaluation is, and is not

Desk evaluation against the published rubric, using each vendor’s live pricing and policy pages, Companies House and public registry records where an operator is named, Ofcom and ICO publications, and aggregated public user reporting. Figures carry the date they were last checked.

It is not a controlled trial. We have not run identical scripted conversations against every app under laboratory conditions, and no consumer site in this category has, whatever its marketing says. When you see a claim of "200 hours of testing" with no logs, no dates and no protocol, treat it the way you would treat any other unverifiable number.

What we can show you is the rubric, the sub-scores, the arithmetic, the date each fact was checked, and the sources for anything contestable. That is a weaker claim than a lab test and a considerably stronger one than an anecdote.

Prices and currency

Almost every app in this category bills in US dollars. We show the listed USD price and an approximate GBP figure converted at £0.79 to the dollar, as at 26 October 2026. Your card issuer's rate will differ, and many UK cards add a non-sterling transaction fee of around 3% on top.

Where an app meters images, voice or calls, we publish a real monthly cost range rather than the headline price, because for an active user on a metered app the headline price is close to fictional.

Corrections and how to challenge a score

Every review carries the date its facts were last checked. Pricing and policies in this category change often, and some of what is on this site will be out of date before we re-check it.

If a fact is wrong, tell us what it is and where the correct version is published, and we will change it and update the check date. If you think a score is wrong, the useful form of that argument is which sub-score is wrong and why — because of how the total is calculated, that is the only thing that can move a ranking.

Commercial relationships

This site earns affiliate commission, and has a commercial partnership with SweetDream AI. Partner apps are labelled everywhere they appear. Commission does not feed into any of the six criteria, and the full arrangement is set out on the disclosure page — including what you should do about it, which is to check the sub-scores rather than take our word for anything.