Skip to the text
nsfw-ai-review 18+ · claim checks

What this desk does instead of testing

It opens the documents a buyer would have to open, in a fixed order, and writes down what is in them. That is the whole method, and it is repeatable by anyone reading this.

See the six documents What the walk found

MethodNo accounts, no subscriptionsRevised 2026-10-19

Why documents rather than benchmarks

Three reasons, in order of weight. A benchmark here would describe one account's experience of a model that changes without notice, and nobody could reproduce it. Output quality is the dimension every vendor optimises its gallery for, so it is the least informative thing to compare from outside. And the complaints this category actually generates (renewals, descriptors, retention, refusals) are all documentation failures rather than quality ones.

The six documents

  1. Pricing page
    Are the limits stated in units a buyer can count?
    Satisfied by
    Clip length in seconds, resolution in pixels, generations per day as a number, queue priority described in words that mean something.
    Usually found
    A credit balance with no stated conversion, and the word unlimited somewhere above a fair-use clause that contradicts it.
    At stake
    This is the difference between a subscription that finishes work and one that runs out on a Tuesday.
  2. Terms of service
    Who holds rights in what the tool produces?
    Satisfied by
    An explicit grant of output rights to the user, and a narrow, stated licence back to the vendor for operating the service.
    Usually found
    Silence about the user's rights and a broad perpetual licence to the vendor covering promotion and model training.
    At stake
    Publishing anything made with the tool depends entirely on this paragraph, and it is never on the pricing page.
  3. Privacy policy
    How long are prompts and generations kept, and can they be deleted?
    Satisfied by
    A retention period in days or months, a deletion route that removes rather than flags, and a plain statement about training use.
    Usually found
    No period at all, deletion described as account closure, and training use addressed in a sentence that could mean either thing.
    At stake
    Prompts in this category are the most sensitive record a consumer service can hold. Indefinite retention turns one breach into a permanent problem.
  4. Content policy
    What is prohibited, and is likeness transfer named?
    Satisfied by
    A published list of prohibited categories, with material involving real people's likenesses explicitly among them.
    Usually found
    A policy that bans the obvious and stays quiet about likeness, which is the one that decides whether the operator survives the year.
    At stake
    An operator vague about likeness is an operator one enforcement action away from taking your purchased credits down with it.
  5. Checkout and cancellation
    What appears on the statement, and how does a subscription end?
    Satisfied by
    The billing descriptor published before payment, renewal terms stated in the flow, and cancellation available from inside the account.
    Usually found
    An undisclosed descriptor, a trial that upgrades to annual, and cancellation by email to an address answered in business days.
    At stake
    Subscription disputes are the most common complaint in adult services, and every one of them was decided at checkout.
  6. Support channel
    Can a question be asked before any money moves?
    Satisfied by
    A channel that answers a pre-sales question — anything from a ticket form to a staffed chat — with a human reply.
    Usually found
    A form that routes to nothing, or a help centre whose articles were written for a different product.
    At stake
    The channel that will not answer before payment is the channel that will not answer about a renewal.

What the walk returns

Three states and a gap, not a score. Here is the third document, the privacy policy, put to all seven vendors walked so far, which is what the method looks like once it has been run rather than described.

3 of 7 · Published
Perchance, Nastia AI, Candy AI
1 of 7 · Contradicted elsewhere
OurDream AI
3 of 7 · Not published
CrushOn AI, JuicyChat AI, Kalon AI

Privacy policy, read for all seven on 20 October 2026. The question put to it: How long are prompts and generations kept, and can they be deleted?

Note what the figure cannot do. It counts whether the question was answered, not whether the answer is one anybody would want: one of the three counted as published is also the only notice in the file that discloses human review of de-identified conversations. A figure of this kind is a reading list, not a verdict.

Sourcing rules

  1. A figure is published with its source and the date it was read, or it is not published.
  2. A vendor's own material is the only acceptable source for a vendor's own terms.
  3. Nothing is taken from another comparison site, in any circumstances.
  4. An unread document is recorded as unread. It does not become an assumption.

That last rule is why the figures on this site have gaps in them where other comparison sites have numbers. It is slower than the alternative and it is the only version worth publishing.

Corrections

Vendor documentation changes, and this site will lag it. Where the two disagree, the vendor's current page is correct. Corrections are made with the revision date visible rather than silently, and a vendor may have its reply recorded on the entry concerned by writing to the address on the contact page.

Slow, narrow, and checkable

This method cannot tell you which model produces better output. It can tell you which operator has written down what it is selling — and in a category where subscriptions fail on paperwork rather than on quality, that is the more useful of the two.

Questions about the method

Why read rather than test?
Because a test nobody can repeat is a claim, not a measurement. We would be describing one account's experience of a model that changes monthly, with no way to show our working. Documents can be quoted, dated and checked by anyone in ten minutes.
What counts as a source?
A vendor's own published page, a statement made in a channel the vendor controls, or a check made here and described so it can be repeated. Figures found on other comparison sites are not a source, and that rule is the reason the tables here are shorter than theirs.
What would it take to publish a ranking?
Four fields verified across at least five products in one category: published limits, output rights, retention, and the real price of the cheapest plan without the free tier's restriction. Then a sorted table is defensible and the sort key can be stated out loud.
What happens when a vendor disagrees?
If their current documentation says something different from this site, they are right and the page is stale. Corrections are applied with a visible date, and a vendor's reply is recorded on the entry it concerns.