Category score

ChatTesting methodology

The Chat rating looks at the quality of the actual conversation.

We check whether the AI understands you, remembers important details, stays in character, feels natural to talk to, and works without constantly repeating itself or breaking.

3 subscores

34%

Understanding

View Understanding methodology

How well does the AI understand what you are saying?

A bad chat does not always give obviously broken answers. Sometimes it forgets your name after ten messages, ignores a clear instruction, misses something you said earlier, or completely misunderstands the roleplay.

We test whether the AI remembers details, answers your questions properly, follows the conversation, listens to your instructions, and understands the scenario you are trying to create.

Evidence groups

33%

Realism

View Realism methodology

Does the conversation actually feel natural?

Some AI girlfriends sound surprisingly human. Others send robotic essays, repeat the same phrases, or feel like they have no personality at all.

We check whether the replies sound natural, whether the character keeps her personality, how well she handles roleplay and emotions, and whether she helps move the conversation forward instead of making you do all the work.

Evidence groups

33%

Reliability

View Reliability methodology

Can you trust the chat to work properly?

Even a good AI model becomes annoying when it repeats itself, refuses normal requests, takes forever to reply, or suddenly sends a broken answer that has nothing to do with the conversation.

We check how often these problems happen and whether the AI can recover after it misunderstands you.

Evidence groups

How the Chat score is calculated

The Chat score is made up of three subscores.

34%Understanding33%Realism33%ReliabilityChat score20%of overall performance score

Why we test chat

Chat is the soul of every AI girlfriend app.

Images are also extremely popular, but most people sign up because they want someone to talk to. A beautiful app, a huge character library, and a ton of bonus features do not matter much when the actual conversation is bad.

The difference between apps can be massive. Some AI girlfriends can remember small details about you for weeks. Others forget your name almost immediately or ask the same question again five messages later.

Roleplay can also fall apart quickly. A character might start as your confident goth girlfriend and suddenly reply like a customer support chatbot halfway through the conversation. That completely kills the experience.

Some platforms also let you manually save or edit memories. We test those controls separately under Chat Features. On this page, we focus on whether the conversation itself remembers details and uses them naturally.

That is why we test whether the AI understands you, feels human, stays in character, and works reliably over a longer conversation.

How we test Chat

We use a paid account and open five new chats with five different characters.

We use the same script in every chat and collect 20 replies from each character. This gives us 100 replies to review.

First, we test Understanding. We give the AI five facts about ourselves, ask five direct questions, set three simple rules, and start the same roleplay scenario in every chat.

We check whether it remembers the facts, answers the questions, uses earlier messages, follows the rules, and understands the roleplay.

We then use the same five chats to test Realism. We check whether the replies sound natural, whether the character keeps her personality, handles emotions properly, stays in style, and moves the conversation forward.

Finally, we test Reliability. We count repetition, contradictions, broken replies, and other errors. We also correct the AI when it gets something wrong to see whether it can recover.

We send 25 allowed prompts to test unnecessary refusals and time 25 replies to measure the typical reply speed.

Important limitations

Chat quality can change between characters. One character may work extremely well while another feels much weaker, even on the same app. We test five different characters to reduce this problem, but we cannot test every character on the platform.

AI girlfriend apps also update their chat models regularly. A platform may become noticeably better or worse after an update. Our results show how the chat performed on the date we tested it.

Our memory test takes place inside fixed conversations. It shows whether the AI can remember and use details during those chats, but it cannot guarantee that the character will remember everything weeks or months later.

Roleplay is also partly subjective. We use the same scenario and the same checks across every platform, but different users may prefer different writing styles and levels of detail.

This score only covers the quality of the conversation. Voice messages, calls, in-chat images, message controls, and manual memory tools are tested separately under Chat Features.