Contextual targeting hasn't been a thing because it's not very easy to accurately extract structured targeting signals from unstructured documents. At best for

Contextual targeting hasn't been a thing because it's not very easy to accurately extract structured targeting signals from unstructured documents. At best for now, we may extract a set of word embedding per page which likely misses precise contexts as well as subtle nuances. Now we have LLM which has objectively better capabilities to do that job and everyone is now forced to invest more into contextual targeting, I guess there will be significant developments on targeting systems. The question would be how to make super cheap models which can handle both pre-processed offline web page context as well as online usage contexts.