Why a Fixed Protocol
The typical “best AI tools” article asks you to accept its rankings on vibes. Where did the score come from? Which complaints did the writer read, and which did they skip? You cannot tell, and that is the point — untraceable praise sells affiliate clicks.
Mimo takes the opposite bet: reviews are more useful when you can audit them. So the site runs one written, versioned method for collecting evidence, applies it identically to every tool, and publishes enough of the trail that a skeptical reader can redo the work. When the method changes, the version number changes, and the change is recorded on this page.
The Source Hierarchy
Not every source deserves the same weight, and the ordering is decided in advance rather than per argument.
Peer-reviewed research carries the most weight: it has authors, a venue, and a review process behind it, so it can anchor a factual claim on its own.
Independent tests with a disclosed method come next. If a third party shows what they fed a tool, when they ran it, and what came out, their result is usable — attributed to them. A “we tested 50 tools” roundup that shows no inputs and no dates is not a test; it is a listicle.
First-person community reports — a dated Reddit thread or Quora answer where someone describes what a tool actually did for them — sit at the bottom of the usable range. They are evidence of one user’s experience, quoted and linked as exactly that.
Vendor materials (pricing pages, docs, marketing claims) are used freely but always tagged as the company speaking about itself: “the vendor states…”, never silently repeated as established fact.
The Qualifying Bar: Dated, Openable, Re-Checked
Three requirements, all mandatory, decide whether a source can be cited at all.
It carries a date. AI tools ship new models constantly; an undated complaint might describe software that no longer exists. Sources older than a tool’s last major update get a staleness note wherever they are used.
You can open it. Every citation is a public URL. If a claim can’t be linked, it doesn’t get made — an uncheckable citation is just an opinion in a lab coat.
It was re-fetched on capture day. Before an entry lands in the evidence log, the page is loaded again and the quoted passage confirmed word-for-word. Links rot; memories drift; this step catches both.
Two source types are rejected even when they pass the mechanics: affiliate rankings with no disclosed method (paid recommendations aren’t measurements), and coordinated fake-forum networks — clusters of near-identical “is X legit?” threads from fresh accounts that all funnel to one product. When we detect such a network, nothing from it is quoted, linked, or paraphrased.
How Claims Are Labeled
Each source that survives screening gets a registry entry with a short code — you’ll see markers like (EV-koala-03) in review text. The entry stores the claim, the URL, the source’s own date, the verbatim quote it rests on, the platform, and the date we captured it. The code travels with the claim so you can pull its file and judge it yourself.
Labels are fixed: research finding, independent test, community report, vendor claim. A community report never graduates into a statistic by being repeated, and a vendor claim never loses its label by being plausible.
Protocol Steps (v1.0)
For every reviewed tool, the sweep works like a census, not a highlight reel. The search queries are written down first and run exactly as documented. Every URL those queries surface is logged before anything is judged — the size of that raw list is the “identified” number you see on review pages.
Screening happens against written questions — is it dated, does it open, does it contain codable detail — and every cut is recorded with its reason: undated, unlinkable, duplicate, off-topic, astroturf, dead link, no-method affiliate content. Kept sources are then labeled in two independent passes; where the two passes disagree, a third resolves it, and anything still murky is published as ambiguous instead of being quietly dropped.
The output is the flow counter printed on each protocol page — identified → screened → excluded (by reason) → included — plus an evidence summary box (source count, platforms, date range, capture date) and a limitations paragraph. Those three artifacts are mandatory; a review missing any of them does not ship.
The Numbers Rule
Counts with a visible denominator are allowed: “12 of the 87 comments that passed screening (Mar 2025 – Jul 2026) describe billing surprises, list linked.” You can audit that sentence — open the list, count for yourself.
Population claims are banned. People who post in forums are not a random sample of a tool’s users; turning five angry threads into “most users report…” or a made-up percentage is fabrication, and it is the single most common lie in this genre. Mimo’s standing rule: a sample can only ever speak for itself.
Every protocol page also names its own limits: dissatisfied users write more than happy ones, each platform has its own skew, and every capture goes stale as tools update. Reading community evidence honestly means keeping those edges in view.
What We Won’t Do
- No invented test results — if we did not run it and cannot cite someone who documented running it, the number does not exist.
- No score without an evidence record behind it. Tools in the directory that haven’t been through the protocol are marked “not yet reviewed” — not given placeholder stars.
- No coverage of AI humanizers, AI detectors, or paraphrasing tools, ever. The founder builds products in those categories, so Mimo removes itself from them entirely rather than pretend to referee its own match (details on the about page).
- No ranking influenced by commissions. Tools with no affiliate program are scored on the same rubric, and when one of them wins its category, it is listed first anyway.
The Founder Conflict, Answered Directly
A fair question: why believe a review site whose founder also sells AI software? Two answers. First, the categories where my interest is real — humanizers, detectors, paraphrasers — are structurally excluded from Mimo; the conflict is removed, not managed. Second, for everything Mimo does cover, the protocol exists precisely so you don’t have to take my word: claims trace to sources that aren’t mine, and the exclusion log shows what was cut and why. Verify the trail, not the author.
Corrections
Evidence expires and mistakes happen; treating corrections as shameful just buries them. When an entry is wrong or superseded, the registry record is amended with a dated note (never silently deleted), the pages that cited it are updated, and the change is visible. Spot something off? Email hello@mimoaitools.com with the EV code and I’ll pull the entry.
This protocol governs every review and guide on the site — start with the latest reviews or browse the full directory.