indeed.tech, as data

The newer unified single-sequence model kept the same operational pattern but improved downstream lift by letting the model attend across event types directly:* Click-based target: roughly +1.6% with legacy multi-sequence UBM, improving to about +3.5% with unified single-sequence UBM.* Apply-start rate: roughly +1.3% with legacy multi-sequence UBM, improving to about +2.3% with unified single-sequence UBM.These results support two intuitions:* More behavior channels matter because different actions contain complementary signals.* Long-tail history matters because the model can learn to denoise behavior instead of relying on hand-truncated windows.* Event order across behavior types matters because job search intent often emerges through transitions, not isolated action streams.### Impact on high-traffic job surfaces in productionThe more important question is what happens in production.Because the same embedding can be reused across many models, the impact compounds across surfaces. In production experiments, adding UBM features produced consistent gains in recommendation quality, apply efficiency, and monetization metrics.Jobseeker Email RecommendationsJobseeker Homepage recommendationsEmployer Candidate RecommendationsEmployer Acceptance rate +1.64%Across downstream models, the UBM embedding is often ranked as a top-1 or top-2 feature by importance, consistent with the observed business impact.## Extending beyond user behaviorOnce the infrastructure existed for user behavior, the same pattern became useful elsewhere in the marketplace.Employer behavior modeling applies analogous techniques to employer interactions with candidates. Those embeddings can feed sourcing, ranking, and bidding systems.