How we test — the 14-day protocol, scoring and citations policy

The protocol: fourteen days, no shortcuts

Every app in the directory gets the same run: a fresh account, a consistent cast of characters, and fourteen consecutive days of ordinary use — chat, voice where it exists, memory probes, and a pricing check against the vendor’s own pages on day one and day fourteen. We do not score anything we have not used for the full stretch.

What we measure

Four things, weighted equally until something breaks: memory (does she remember Tuesday?), filters and their false positives, voice latency and naturalness where applicable, and total cost of a normal month — subscription plus the tokens the subscription quietly doesn’t cover. Anything that changes mid-test gets re-tested before it gets printed.

How scoring works

Each head-to-head scores the contenders on the dimensions printed in its table, and the winner is named in plain language with the reason attached. Scores are comparative, not absolute: an 8 here means “beats the field this quarter,” not “perfect.” When vendors ship material updates, the piece is re-run and the dateline moves.

The citations policy

Every price, limit and policy claim carries a numbered source line at the foot of the section that makes it. Sources are vendor pages first, vendor documentation second, and in-app measurements (screenshotted with the date in frame) when the docs are cagey. If we cannot source it, we do not print it.

Corrections

Wrong things get fixed in public. Corrections append to the footnotes block of the piece they correct, with the date of the change, and material ones get a note at the top. Silent edits are for typos only.

A note on FlirtAlly

FlirtAlly is made by the same team that runs The Exit Guide, which is exactly why it carries no score anywhere on this site. It appears in one clearly marked card — never in the rankings — and the apps we test are tested the same way whether they compete with it or not.