Elegant Integration of Ads in Search: An Analysis of Baidu's Mobius Approach

Search is fundamentally a tool for users to actively retrieve information, and its core starts with a user‑issued query. Compared with traditional portal navigation, search reduces noise by directly matching queries to relevant results, while recommendation systems further lower the entry barrier by pushing content without explicit input.

The article contrasts three information‑acquisition models—portal indexing, general search, and recommendation—highlighting the strengths and limitations of each, and explains why search advertising occupies a unique middle ground where ads must be tightly related to the query.

Typical ad formats such as native feed ads, alliance ads, and form‑submission ads are illustrated, showing how search ads differ by requiring strong query‑ad relevance. The discussion then introduces Baidu's Mobius paper, which proposes a unified pipeline that simultaneously satisfies relevance thresholds and maximizes eCPM (CTR × bid).

Mobius’s workflow includes: (1) constructing query‑ad pairs and augmenting data via cross‑product; (2) using a relevance judge (teacher) to label relevance; (3) filtering low‑relevance pairs; (4) feeding the filtered set into a neural click model (student) to obtain pCTR; (5) selecting high‑pCTR, low‑relevance samples as “bad cases”; (6) merging bad cases with normal training data to train a three‑class ranking model (click, unclick, bad case); and (7) jointly optimizing relevance and value in a single ranking stage.

The article critiques the traditional three‑stage approach, arguing that Mobius’s three‑class design effectively incorporates relevance awareness into CTR prediction, allowing high‑value ads that may have lower relevance to be retained while still respecting user intent.

Finally, the author reflects on the broader challenge of understanding user intent (purpose) beyond mere relevance, suggesting that future systems should classify intent types and adapt relevance constraints accordingly, while always keeping commercial value maximization as the ultimate goal.