Skip to main content
Core concept · AI Search Infrastructure

​
Definition

Embedding drift is the movement of a passage’s embedding vector away from a target retrieval cluster caused by the introduction of off-topic content. A section that starts answering one question and pivots to address a second ends up positioned between clusters rather than inside either one. Embedding drift is why mixed-topic passages underperform structurally clean ones even when both contain good information. The drift is not a gradual weakening — it is a geometric repositioning. The vector ends up in a low-density region between clusters, far from the dense retrieval targets of both topics. The practical fix is the same as the structural advice: one question per section.

Semantic center of gravity

Semantic density

Topic coherence

Embedding

Cluster

​
Relevant Plate Lunch Collective Services

Citation-Ready Content AI SEO