The learning model & smart rescue

Search Terms gets sharper the more you use it, but it's careful about when it trusts what it's learned.

How it learns

As your team reviews and approves recommendations, the learning model picks up which calls fit your accounts, and the rule defaults calibrate to match. Two safeguards keep this honest:

• A cold-start gate - until there's enough history to trust, it stays in plain rules-only mode rather than acting on thin signal.

• Brand terms are kept out of learning, so they can't distort what it picks up.

Smart rescue

Sometimes a term converts well but doesn't look related to your keywords - a local nickname, a synonym, a product's informal name. Plain relevance would wrongly skip it. Smart rescue lets the learning model keep that term when model confidence at least 15% and it has at least 10 conversions - proven, just oddly worded.

Example. "aircon" never matches your "air conditioning" keywords on wording, but it's converted 14 times. Smart rescue keeps it instead of discarding it as irrelevant.

Good to know. This is why honesty matters most for negative feedback - when you reject or reverse a recommendation, that's the strongest signal the learning model uses. Positive approvals on thin, single-account data stay advisory until the numbers back them up.

RELATED ARTICLES