How the lab works

1. 01
   Intake. Makers submit their products on the submit page; the lab also picks launches readers will want to know about from open sources and the old Broadwise forum.
2. 02
   Rubric. Each product is judged on the job its users actually hire it for: an AI builder on whether its AI gives you a site, a launch directory on whether a launch there gets makers seen. Data safety is always on the list.
3. 03
   Assay. Agents sign up, use the product like a customer, run the core flows, measure speed and read pricing, terms and privacy pages. Screenshots of everything they make.
4. 04
   Background. Founder, current owner, acquisitions, outages, breaches and controversies go into a dossier and a timeline.
5. 05
   Report. Scores per criterion, a verdict, pros and cons, screenshots, plus what the run cost at market token prices. Reports go live one by one as they are finished.

There is no one-size-fits-all checklist. An AI website builder is judged mostly on its AI, a launch directory on whether launching there gets makers seen and visitors new tools worth trying. Each product gets 5–8 criteria with weights that add up to 100; the report lists every criterion and its weight. Two rules hold for every rubric: data safety is always scored (export, deletion, backups, uptime, breaches, ownership changes), and products of the same kind share a rubric template, so they can be compared criterion by criterion. Compare pages exist only between products of the same kind.

Grades follow the verdict bands: A must-try (90–100), B worth-it (75–89), C meh (50–74), D skip (25–49), F avoid (0–24), with + and − for the top and bottom third of each band. The court measures against the ideal version of each product, so a solid real product meets about half of it; the score shown here is the court's, moved up a fixed ladder so that half the ideal reads 60, not 40. The order never changes, and each report shows the product's place among products of the same kind. The score starts from the headline promise: a blind judge writes what the ideal version of the product would do and counts how much of it is real (20, 40, 60, 80 or 95; 5 only for a scam). Then a court moves it up to 10 points (never below 10 unless it is a scam): one agent argues against the product, one for it. Three practical judges, working blind and apart, rule on every point by what an ordinary user actually gets, and the median of their three scores wins.

Every report shows the tokens its runs consumed and what they would cost at API list prices, even when the run used a subscription. So far the lab has used 14.42M tokens and $105.82 of compute across 22 runs.

Scores can't be bought: makers pay for the test, not the grade, and can get a refund at any time. They can email the lab to correct a fact, but they can't edit the report. Outbound links carry ?ref=broadwise.org so makers can see the traffic.