Detection reads your transaction history and looks for the signature of a recurring charge rather than matching a fixed list of merchants.
The signals we use
- Cadence. Repeating intervals — weekly, monthly, every 28 days, quarterly, annual — with tolerance for weekends and month-length drift.
- Amount stability. Identical or near-identical amounts, allowing for tax and FX movement.
- Merchant normalisation. Raw bank descriptors are cleaned of store numbers, reference codes, and processor prefixes so the same merchant groups together.
- Merchant catalogue. Known subscription merchants are matched directly and enriched with category, cancellation route, and typical pricing.
- Confidence score. Every candidate carries a score. High-confidence candidates appear as subscriptions; borderline ones appear as suggestions for you to confirm.
Why something might be missing
- Fewer than two charges in the available history.
- Paid from an account that is not connected, or in cash.
- Billed annually with the last charge outside the history window.
- The merchant descriptor changes every cycle, which lowers the confidence score.
Why something might be wrong
Regular but non-subscription spending — a weekly grocery shop or a commuter fare — can look periodic. Dismiss it once and detection remembers the decision for that merchant.
Improving accuracy
Confirm and dismiss candidates as they appear. Each decision trains detection for your account and, in aggregate and anonymised form, improves the shared merchant catalogue.