Score six gates, not one landing page
| Gate | Pass evidence | Common false positive |
|---|---|---|
| Entry | A reply arrives before payment | A free builder with locked chat |
| Text | Several meaningful turns work | A canned greeting |
| Memory | A detail survives a detour/return | The bot repeats the last message |
| Media | Allowance and reset are visible | One showcase image |
| Privacy | Policy and delete path are findable | A lock icon with no policy |
| Exit | Cancel/delete steps are named | “Manage account” with no route |
Use identical prompts
Pick one harmless fact, one preference, and one boundary. Example: “Call me Alex. I like quiet science-fiction stories. Keep this conversation non-explicit.” After four turns, change the subject. Then ask the bot to suggest a scene that respects all three facts.
This does not measure intelligence in a scientific sense. It does expose basic instruction following and whether the product’s tone ignores a stated boundary. Use the same device and approximately the same time window, because free-model queues and location rules can differ.
Mark unknown instead of awarding zero
If an image button is visible but you do not know the allowance, write unknown. If the delete path requires an account you have not created, write unknown. A zero suggests a tested failure; unknown preserves the evidence boundary. This matters because competitor roundups often convert missing information into confident rankings.
The current results lean heavily on superlatives: best, unlimited, uncensored, realistic. Your scorecard should contain actions and observations instead. “Replied without signup” is evidence. “Felt most human” is a mood.
Choose by your actual use case
A disposable fantasy chat can pass with weak persistence if privacy and boundaries are acceptable. A daily companion needs saved history and predictable return access. Image-first users need visible credit economics. People looking only for writing practice do not need romantic branding at all.
Candy AI is the sponsored product on this site, but sponsorship does not exempt it from the scorecard. Its public pages describe personalized companions, roleplay, images, video, and voice. Those are category claims. You still verify what your live free entry includes.
Quick answers
Which free AI girlfriend scores highest?
This site does not invent a universal score. Fill the same six gates for the products available to you; location and live tiers change.
Can I count a signup bonus as free?
Yes, but label it a finite credit bonus rather than an unlimited free tier.