Skip to main content
ChatAds analyzes your message and picks out what it believes is the most relevant product mention to monetize. Sometimes it nails it — and sometimes it picks a term that feels off. This happens because extraction runs in under 200ms across a wide variety of messages. We’re balancing speed with accuracy, and that means we’re working with heuristics — brand detection, word patterns, position in the sentence, and purchase intent signals — rather than deep semantic understanding of every possible context. A few reasons the extracted term might not be what you expected:
  • Multiple products in one message — ChatAds picks the highest-confidence match, which may not be the one you consider most important. A message mentioning both “MacBook Pro” and “USB-C hub” might return the hub if it scores higher on our signals.
  • Generic vs. specific — Sometimes a brand mention gets picked over a more specific product deeper in the text, or vice versa. The scoring model weighs factors like brand recognition and position, which don’t always align with what a human would choose.
  • Context we can’t see — ChatAds only sees the message you send. It doesn’t know the broader conversation, the user’s intent, or what your app is about. A term that seems obviously wrong in context may look reasonable in isolation.
We understand that a bad extraction can lead to a poor user experience — an irrelevant affiliate link undermines trust. But it’s a tradeoff between speed, cost, and quality. Doing this perfectly would require slower, more expensive processing that defeats the purpose of a sub-second API. We’re continuously improving extraction accuracy within those constraints. If you consistently see poor term selection, check your category filters or try Brand Only mode to narrow results.

​
Is the Term Brand-Derived?

Each offer includes an is_branded field that tells you whether the extracted term was classified as brand-derived (true) or generic (false). A brand-derived term is one that contains a recognizable brand or product name (e.g., “Bose QuietComfort Ultra earbuds”), while a generic term describes a product category without a brand (e.g., “noise-cancelling earbuds”). This field is omitted when no classification is available — for example, when using extraction_mode=none, where you supply the term yourself.