Review methodology
This page documents the exact process behind every AI chat platform review published on polybuzz-ai.io. The framework draws on public AI evaluation guidance from bodies like the National Institute of Standards and Technology (NIST) and independent research indexed at Stanford HAI, adapted for consumer AI companion products. If a claim we make about a platform isn't reproducible using the steps below, we treat that as a bug in our own review, not in the platform.
Hands-on testing period
Every platform we review gets a minimum two-week hands-on testing period before publication. Melissa uses the platform herself, daily, across the categories below. Two weeks is the floor, not the goal. Most platforms in the AI companion category take four to six weeks to fully evaluate because memory behaviour, moderation quirks, and drift patterns only surface over long sessions.
Testing happens across at least three devices and browsers: a desktop browser (Chrome, latest stable), an iOS mobile browser (Safari, latest), and an Android browser (Chrome, latest). Where a platform ships a native app on the app store, we test that too, but our default is the web app when both exist.
Review categories
Every review scores a platform across the same seven categories. Categories carry equal weight. We don't inflate one to hide weakness in another.
1. Feature completeness
What's in the free tier and what's behind the paywall. This includes companion library size, custom character builder capability, memory length, response speed, media features (images, voice), and any specialised modes.
2. Moderation and content policy
How the platform handles adult content and where filters fire. We test with a standard set of scenarios (flirtation, romantic escalation, explicit language, roleplay themes) and note both what triggers refusals and what the refusal patterns look like. "Unfiltered" claims get tested against actual behaviour.
3. Persona quality and consistency
How well companions hold character across long sessions. We check for persona drift (companion breaking character), memory bleed between characters (details from one companion appearing in another's replies), and tone-matching (whether the model mirrors user tone in practice or defaults to a generic register).
4. Memory behaviour
How much history the platform holds within a session and across sessions. We time-tag specific details early in a chat and check whether they resurface correctly at message 50, 100, and 200. We also test cross-session memory on paid tiers where offered.
5. Privacy and data handling
What the platform collects, how it stores it, and what "delete" really does. We test account deletion, single-conversation deletion, and check whether deleted data reappears anywhere. We read the privacy policy against observed network behaviour.
6. Access friction
Time from landing on the site to first message sent. Login requirements. Age verification friction. Payment flow if reaching a paid tier. Cancellation friction if closing an account.
7. Performance and reliability
Response times measured across three time windows: peak (evening hours in North America), off-peak (early morning UTC), and mid-day. Uptime tracked across the two-week testing window. Queue behaviour and cooldown behaviour under sustained load.
What we verify vs. what we take on faith
Claims we verify with reproducible steps: memory length, response times, message caps, feature availability, moderation triggers, deletion behaviour, uptime during testing, price transparency.
Claims we take on faith and note as such: back-end architecture we can't inspect, promises about data retention beyond what we can test, revenue and user counts, roadmap timelines, and anything a platform tells us privately that we can't independently confirm.
Sources
Reviews are based on hands-on testing first, product documentation second, and third-party sources last. When we quote a platform's own claim, we attribute it directly. When we cite an external source, it appears with a link. We do not paraphrase competitor reviews or rely on aggregators.
Update cycle
Every review is re-verified at minimum every 90 days. Product features move fast in this category, and a review from six months ago is often factually wrong today. When we re-verify, the dateModified stamp updates, the "Last reviewed" date at the bottom of the page updates, and any material changes get called out in a changelog entry at the bottom of the review.
Reviews older than 180 days without re-verification are either updated or marked as such. We don't leave stale material sitting unmarked at the top of search results.
Corrections
If a factual claim on any review is wrong, we want to know. Corrections raised through the contact channel on the about page or directly to Melissa on LinkedIn get a response within seven business days, with acknowledgements logged on the editorial policy page under the corrections section.
Author
Reviews are written and fact-checked by Melissa Blake, senior editor covering AI companion platforms. Editorial oversight of the methodology itself is documented on the editorial policy page.
Last reviewed: July 8, 2026 · by Melissa Blake