The built-in keyword layer of ai-bookkeeper is tuned on one real export plus synthetic sets, so its blind spots show up fastest by probing. This is the fourth probe: about 130 descriptions written from public company names (not a real export), each filed as a -$25 debit.
25 had no suggestion. Eleven are now filed: FreshBooks, Wave, Smallpdf, SurveyMonkey, Postmark, Freshdesk and Grammarly as software; Incfile as a legal fee; and Gumroad, Patreon and Lemon Squeezy credits as sales. Costco, Walmart and Netflix stay for a person on purpose, since they are personal or mixed.
The rest (LinkedIn, Nextdoor, Etsy, Medium, Meetup and bank-fee and interest lines) are ambiguous between advertising, fees and personal use, so they remain unsuggested rather than guessed. The author alone judged the results, so the figures are a coverage count, not an accuracy measure. Tests: gaps4.test.mjs.