01

Define what each node and edge means

A useful graph begins with a narrow universe. Nodes might represent query classes, domains, product categories, content topics, or journey stages. Edges record an approved relationship such as a click, transition, shared research path, or repeated co-occurrence.

The schema should state the unit, direction, observation window, and aggregation rule. A line between two nodes has little meaning until those definitions are clear.

  • Query class to clicked domain
  • Domain to destination category
  • Journey stage to next stage
  • Category to recurring adjacent category
02

Weights need time and coverage context

An edge weight can represent count, share, transition rate, lift, recency, or a normalized score. Those measures answer different questions and should stay separate in the schema.

Time windows expose whether a relationship is stable, seasonal, or changing. Coverage and suppression notes show where a small or uneven group could distort the weight.

  • Observation period
  • Edge-weight definition
  • Minimum cohort rule
  • Market and source coverage
  • Stability or quality flag
03

Match the graph to the decision

Search teams can use query-to-domain edges to study retrieval paths. Recommendation teams can examine category transitions. Market researchers can map attention between brands, publishers, products, and topics.

The same graph should not be reused blindly across tasks. Entity definitions, negative examples, time splits, and evaluation metrics should follow the model or business question.

  • Ranking and retrieval context
  • Recommendation candidates
  • Market and competitor maps
  • Journey-stage analysis
04

Observed behavior is evidence, not certainty

A visit or transition may reflect curiosity, work, comparison, accidental navigation, or another context. An intent label is an inference tied to a definition and a population.

BGraph standard outputs use approved fields, aggregation controls, and documented limits. Sensitive classes, small groups, individual profiling, and high-impact uses require exclusion or separate review.