The categorizer in ai-bookkeeper has keyword rules for common merchants. They were first written while reading one set of Hack Club card-charge memos, so the 23.8% to 33.2% coverage reported there says little about new data. This run read a different 284 charges from transparent organisations on the same public feed, none looked at before.
Result: 17.6% suggested (50 of 284). Amazon alone was 109 charges and is left blank on purpose, because it could be parts, books or food. AliExpress was 15 more. Without labels this scores coverage, not accuracy, so the 21 distinct hits were read by hand. Two were wrong:
PayPal *LUMENS FTY.is a lighting shop. A rule meant for the carrier Lumen matched it and called it a phone and utilities bill.Solar power manager moduleis an electronics part. A rule for power bills matched the word "power".
Both are fixed with a word boundary and a short list of hardware nouns, and a test keeps Lumen the carrier and "City Power & Light" matching. openrouter was added as software after it showed up four times; that rule was written after looking, so it does not count as held out.
The takeaway is the usual one for rule-based categorizers: coverage on new merchants is low, wrong guesses come from words that mean two things, and the payee memory and review queue carry the long tail.