Max Cruz tested five “free” AI girlfriend tiers in May 2026 — Candy AI, DreamGF, Kupid AI, Joi AI, Replika. Burner email, zero credit cards, signup-to-paywall verified on each. Four of the five demos hide the exact feature you’d be paying for. Here’s what each free tier actually shows.
A review goes deep on one platform. An experiment goes wide — one question across many platforms at once. “Is the free tier honest.” “Do the annual discounts survive renewal.” “How fast do the tokens really drain.” Those are category-wide questions a single review structurally cannot answer, and they’re what this section exists to test.
Reviews are vertical — everything about one platform. Experiments are horizontal — one variable held constant across a whole field. The two answer different questions. If you want to know whether a specific platform is worth your money, read its review. If you want to know whether the entire category pulls the same trick on everyone, no single review can show you, because the pattern only becomes visible when you line several platforms up side by side.
“I tested this for 30 days” is the most abused sentence in affiliate writing. A real test carries specifics — what happened on which day, what broke, what the actual numbers were when the paywall hit. A fake test carries vibes and a referral link. The reliable tell: could the writer have produced this without doing the test? If the whole thing could’ve been written from the platform’s marketing page, it was.
Here’s the constraint nobody admits: one person cannot run a genuine 90-day test on five platforms simultaneously. So the experiments here state exactly what was verified firsthand versus what’s cross-referenced from other testers’ published findings, with attribution. The scope is part of the result. An experiment that claims more than one person could plausibly do is lying about its method, which means it’s lying about its conclusions too.
Patterns no single review catches. The billing-complaint structure that’s identical across every platform in a category. The free-tier bait model that every competitor copied from the same playbook. The token economics that look reasonable on one platform and predatory once you see them lined up. The category has tells, and they only surface when you test across it instead of within it.
Check the scope claim first. If an experiment says it did more than one person realistically could, doubt the rest of it. Mine tell you the limits up front — what was tested, what was read, where the line is. That’s not a weakness in the method. It’s the only thing that makes the conclusions worth anything.