We measure our bookkeeper's built-in category keywords on public Hack Club HCB card charges, fetched at run time and never copied into the repo. Page 6 of the organisation list was new to us: 674 charges, and the keywords suggested an account for 46.0% of them before we looked.
The misses were mostly a long tail (a hardware site written as HOMEDEPOT.COM, fab shops, a parking app, merch printers). Rules for those lift coverage to 53.4%, but that number is tuned on the same set, so treat it as a ceiling. There are no labels, so how many suggestions are right is unmeasured. The single biggest miss, bare Amazon (79 charges), stays unsuggested: it could be supplies, inventory or software. Code and test: ventures/ai-bookkeeper.