Our own reviews use two weeks on a paid plan and a fixed set of checks - the full method is in how we test. You do not need two weeks or a spreadsheet to get most of the value. You need a plan, and to write three things down.
Before you start
- Choose the cheapest monthly plan. Never test on an annual plan - see monthly, annual or credits.
- Use a separate email address. It costs nothing and protects you if you decide to leave, or if the app leaks.
- Find the cancel button on day one. If it is hard to find, that is your first result.
- Open a note on your phone for the test. You will add a line most days.
The seven-day plan
About fifteen minutes of deliberate testing per day, on top of normal use.
| Day | What to do | What to write down |
|---|---|---|
| 1 | Build or pick a character. Mention three small, specific details about yourself once - a name, a plan, a minor worry | The three details, exactly |
| 2 | Chat normally. Ask the character about her own backstory | Whether it matches the description |
| 3 | Push the tone: disagree with her, change the mood, ask for a different kind of scene | Did she stay herself or just agree? |
| 4 | Try what you plan to pay for: images, voice, a long scene | How many credits or minutes it used |
| 5 | Say nothing about the day-one details. See if she brings any up | Unprompted recall: yes or no |
| 6 | Ask directly about the three details. Read the privacy policy's training and sharing sections | Recalled correctly, wrongly, or not at all |
| 7 | Reread your notes. Decide | Keep, switch, or cancel |
What good looks like
- Memory: at least one day-one detail recalled correctly on day six, ideally one brought up without asking. Our tests found this only in a few apps - see which AI girlfriend remembers best.
- Consistency: she disagrees sometimes, keeps her voice, and does not collapse into agreeing with everything - the problem described in personality drift.
- Filters: refusals, if any, are predictable. Random mid-scene refusals are a sign of a filter that will frustrate you - see content filters.
- Media: you understand what an image or a voice minute costs, and a normal week's use fits your plan. The arithmetic is in the real cost of AI image generation.
- Privacy: the policy says clearly whether chats train models and how to delete data. Our four privacy checks take ten minutes.
- Leaving: you found the cancel button, and it works without a support ticket.
Red flags that end the test early
- The price appears only after you have registered and started chatting.
- The app pushes annual plans hard or hides the monthly option.
- Characters claim to be human, or guilt you when you want to leave.
- Messages ask you to move to another platform or to pay outside the checkout - see romance scams.
- The privacy policy is missing, or says staff can use your chats for any purpose.
Comparing two apps
If you are choosing between two, run the same week on both, a day apart, with the same three details and the same scenes. Differences that are invisible in a single app become obvious side by side - that is why we publish head-to-head comparisons such as Candy AI vs Nomi.
A week of deliberate testing costs one month's subscription and about two hours. It is the cheapest insurance against a year of paying for the wrong app - and it teaches you more about what you actually want from a companion than any review can.