How We Test AI Waifu Apps

• Every app is paid for out of our own pocket, on the plan a normal user would buy
• A month of daily use minimum before anything gets written
• We test the things that only show up after you have paid: memory, refusals, cancellation
• We take affiliate commission, and it does not change the order

Most pages ranking for "best AI girlfriend" were written by someone who opened three free tiers for twenty minutes. You can tell, because they all describe the same things: the signup flow, the character builder, the marketing copy on the pricing page. Those are the only parts visible without paying. Everything that actually decides whether an app is worth keeping sits behind the wall.

What we buy, and with whose money

Every app covered here was subscribed to on a normal paid plan, bought with our own card, at the price anyone else pays. No press accounts, no comped subscriptions, no vendor-supplied test logins. That matters more than it sounds: a comped account often sits on internal limits that differ from the retail plan, so the reviewer is genuinely describing a product the reader cannot buy.

We buy monthly rather than annual, deliberately, even though annual is cheaper. Annual pricing hides the renewal experience, and the renewal is one of the things we want to see.

How long, and how much

A month of daily use before anything is published. Not a session, not a weekend. The reason is that these products are indistinguishable in the first week. Every one of them is charming on day one, because day one is a fresh context window and a well-written opening persona. The differences appear somewhere between week two and week four, when the app either remembers who you are or quietly does not.

Daily also means daily, including the boring days. A companion tested only during long deliberate sessions gets judged on its best behaviour. Most real use is five minutes on a commute, and how an app handles five minutes is a different question from how it handles an evening.

The five things we actually measure

• Memory across sessions. Facts given in week one, brought up unprompted in week four, without being re-fed. This is the single largest difference between apps and the hardest to fake.
• Refusal behaviour. Where the content wall sits, and more importantly how it is written: an in-character deflection or a system error dropped into the middle of a scene.
• The true cost of a month. Subscription plus whatever the app meters on top, measured against real use rather than the pricing page.
• Character consistency. Whether the same character talks the same way in week four as in week one, or drifts toward a generic voice.
• The exit. How cancellation works, whether it is in the settings or by email, and what the charge is called on the statement.

The last one gets skipped by almost everybody, which is exactly why we do it. An app that is pleasant to use and unpleasant to leave is a worse product than its reviews suggest, and you only find that out at the moment you are already annoyed. We wrote up what that process looks like across the field separately in the cancellation and refund guide.

What we deliberately do not do

We do not score out of ten. A number implies a measurement, and most of what matters here is a fit between a product and a person: the app that suits someone who wants a written character is the wrong app for someone who wants to design one. A number flattens that into a false ranking, and it is also exactly the shape of markup that search engines treat as a fake review signal when there is no methodology behind it.

We do not publish vendor performance claims as findings. Response times, model names and user counts from a marketing page are not observations, and repeating them with our name attached would make them look verified when they are not.

We do not review apps we cannot pay for. Several well-known names are missing from this site for that reason alone. Writing about a product from its own landing page is the practice this page exists to distance itself from.

How we are paid, and what that changes

Links to apps on this site are affiliate links. If you subscribe after clicking one, we receive a share of what you pay, usually for as long as you stay subscribed. That is the entire business model, there is no advertising and nothing is sold to you directly.

The honest thing to say about that is what it does and does not influence. It does influence which apps get covered at all: we cover apps we can be paid for, because a month of testing costs money and time. It does not influence the order, and the practical evidence is that the ranking has moved against our own interest more than once — the app that pays us the most is not at the top of the ranking, and two apps we were paid for were removed entirely when we could no longer verify their billing.

When pages get updated

Prices in this category move, and what a subscription includes moves more often than the price itself: the usual change is a generation allowance quietly tightening rather than a number going up. Every review carries the date it was last checked, and a page whose date is more than a few months old should be read as a snapshot rather than as current.

Content policy moves too, and rarely because a platform changed its mind. Payment processors set rules for adult merchants, merchants comply, and users experience the compliance as an app getting stricter for no stated reason. That is why we test refusal behaviour on every re-check rather than assuming last quarter's result still holds.

Where this leaves you

The method is not complicated and it is not secret. Pay for the thing, use it long enough for the novelty to wear off, and write down the parts nobody can see from outside. Most of the value in this site is not cleverness, it is the month. If you want the output rather than the method, the ranking is where it lands, and how these apps work underneath explains why the memory test is the one that separates them.