What changed
Two layers of accuracy for keyword tracking:
- Whole-word matching. A keyword now matches the word itself, not fragments buried inside longer words — Unicode-aware, so it holds in any language.
- AI mention verification. Before an alert fires, an AI pass checks each match: "this was tagged with your keyword — is it actually about it?" Coincidental hits, including foreign-language ones that merely contain your term, get filtered.
Keywords can now also carry a short description, so the verifier knows what your term means.
Why it matters
Every false positive is a notification you learn to ignore — and an alert you've learned to ignore is worthless. Substring hits and cross-language coincidences were the two biggest sources of noise. Verification also fails open: if the AI check is ever unavailable, you get a little noise rather than silence.
How to use it
Nothing to configure. For best results, add a one-line description to each keyword — in the dashboard, or over MCP with keyword_create / keyword_update — so the verifier knows that "syften" means the tool, not the Swedish word.
Limits
Verification runs within a fair-use AI allowance per workspace — a monthly pool with a weekly smoothing cap. Exhausting it never blocks anything: matches keep flowing, just delivered unverified until the weekly (Monday, UTC) or monthly reset, and the Watchers page says so while it lasts. Up to 5 keywords are verified per post. Your agent can check the remaining allowance any time with get_usage; in the dashboard it lives on the API access page.