Events become useful when order is preserved
A list of page views loses the path between discovery and action. A trajectory keeps the order of approved events, the transition between them, and the time allowed between steps.
The sequence can use domains, categories, channels, or journey stages. The right level depends on source rights, model needs, cohort size, and the risk of exposing a distinctive path.
- Ordered domain or category visits
- Query class to destination path
- Channel and journey-stage transitions
- Time between approved events
Trajectory, session, and cohort are different units
A session usually follows a defined inactivity rule. A trajectory may span a shorter task or a longer approved research window. A cohort groups similar paths so teams can compare patterns without carrying a personal history into the output.
The data dictionary should name the unit directly. It should also state how paths are truncated, grouped, suppressed, and refreshed.
- Session rule or observation window
- Path depth and truncation
- Trajectory-class definition
- Cohort and suppression threshold
Use sequence context where it changes the answer
Retrieval teams can study which destination follows a query class. Recommendation teams can model category transitions. Forecasting teams can monitor changes in research depth and repeat activity before a market outcome appears.
Market researchers may prefer trajectory aggregates because the path remains explainable. Model teams may request encoded sequences or transition features after the purpose and evaluation plan are set.
- Search and retrieval evaluation
- Recommendation context
- Demand and trend features
- Journey and market research
Keep human evidence separate from inferred labels
An observed browsing sequence can support a behavioral hypothesis. It cannot establish identity, attention, motivation, or whether every event was produced by the same real-world person without additional evidence.
BGraph does not currently present a standard human-versus-synthetic web classifier. A future product in that area would need a published labeling method, validation data, error measures, and clear limits before a public claim.