Quick definition: An identity graph is a data structure that records links among identifiers—such as account IDs, devices, browser IDs, and email-derived keys—to support a governed view of related events.
What is Identity Graph?
Identity graphs help a measurement system decide which events may belong to the same account, device, or person. They can improve duplicate handling, cross-device analysis, and audience suppression, but a graph is a model of observed identifiers, not proof that every linked record represents one human. The stronger the linking method, the more important it is to preserve provenance, confidence, purpose restrictions, and a way to undo a bad link.
In practice, the useful question is not whether a term sounds familiar but what it permits a team to infer and what it leaves unresolved. Measurement choices shape the population that appears in a report, the events treated as outcomes, and the actions a team may take. A responsible implementation makes those choices explicit, minimizes unnecessary data, and separates operational reporting from causal claims.
Definitions and boundaries
A deterministic edge uses a known shared identifier or authenticated event. A probabilistic edge is inferred from signals and has a nonzero error rate. Nodes can represent identifiers, accounts, devices, or other scoped entities; edges should record direction, source, creation time, expiry, confidence where relevant, and permitted uses. Resolution rules must also address account merges, reused devices, shared accounts, and deleted identifiers.
For reliable reporting, document the measurement contract before reading the result. That contract should name the unit of analysis, population, outcome, event time, identifiers, inclusion and exclusion rules, refresh cadence, and owner. It should also state what the data cannot show. A clean dashboard is not evidence that its inputs are complete, representative, or suited to every decision.
Data minimization is a practical design discipline as well as a governance principle. Start with the decision and collect the least detailed information that can support it. Keep raw events, derived fields, and exports distinguishable. Limit access by role, log material transformations, and make retention and deletion behavior testable. These practices improve both reviewability and analytic reproducibility.
Data and experimentation applications
Measurement systems can support experiments when they preserve the randomized comparison rather than replacing it with a convenient segment or report. Define eligibility using information available before assignment; assign a stable unit; log assignment and actual exposure separately; and measure the pre-specified outcome consistently in every variant. Compare outcomes by assigned group first. This intent-to-treat comparison remains informative when delivery is imperfect, whereas filtering to people who received treatment can introduce selection bias.
Use data products as diagnostics. They can reveal missing events, uneven linkage, delivery failures, duplicated conversions, or changes in journey labels. Check these by variant and over time. A change in match rate, tracking availability, or attribution coverage may be a measurement change rather than a behavioral effect. The glossary guides to A/B testing, data validation, and event tracking describe the related controls.
Privacy-aware design does not require abandoning experimentation. It does require scoped identifiers, clear data flows, secure access, transparent definitions, and a plan for incomplete observation. When individual-level linkage is not needed, aggregate outcomes or privacy-preserving reporting may be a better fit. When it is needed, analyze only the approved population and disclose coverage limits. For measurement decisions with material business impact, pair descriptive reports with a randomized holdout, geo test, or another credible causal design.
Concrete scenario
A streaming service receives anonymous web events and authenticated account events. It links a browser to an account only after a successful login, records the observed timestamp and source, and keeps that deterministic edge separate from inferred similarities. When two household members share a television, the system does not assume that every viewing event identifies one individual. An onboarding experiment is randomized on the account ID and analyzed on that same unit, even if the graph later gains additional device links.
The scenario illustrates a useful separation of responsibilities. Product or marketing owners define the decision and operational context. Data owners define event contracts, quality checks, and access boundaries. Experiment owners pre-specify eligibility, assignment, outcomes, and analysis. Privacy, security, and legal functions review the relevant risks and requirements. No single dashboard should silently combine all of those decisions.
Limitations and common mistakes
Calling inferred matches certain; overwriting raw identifiers with a resolved profile ID; merging households or shared devices into a person-level graph; using post-treatment identity changes to redefine an experiment denominator; and exporting the full graph to every activation destination.
- Unclear definitions. A metric or label that changes by tool, team, or month cannot support a clean trend comparison.
- Hidden coverage gaps. Missing identifiers, delayed imports, blocked collection, and offline activity can change who appears in a report.
- Post-hoc analysis choices. Changing the segment, window, metric, or model after looking at outcomes increases the chance of a misleading conclusion.
- Uncontrolled reuse. A dataset collected for one workflow can become risky or analytically invalid when exported to a different purpose without review.
- Overconfident language. Report what the method estimates, its uncertainty, and the plausible alternatives it does not rule out.
Before acting on a result, ask four questions: What decision is this evidence intended to support? Which population and events are represented? What comparison or assumption turns the data into a conclusion? What would change the conclusion? Answers should be recorded with the report, not reconstructed after a disagreement. The articles on writing an A/B test hypothesis and primary and guardrail metrics offer practical decision discipline.
Operational checklist
Use a lightweight review before publishing or acting on a report. Name the owner and decision deadline; freeze the definition and relevant date range; verify event freshness and duplicate handling; compare coverage across meaningful groups; record any changes to identifiers, collection, or exports; and preserve a reproducible version of the query or model inputs. If the analysis supports an experiment, confirm that eligibility was determined before assignment, assignment is stable, exposure is logged separately from assignment, and the outcome is available under the same rules for every group.
Then communicate the result with calibrated language. State whether it is a descriptive allocation, a predictive model output, or a causal estimate; provide the unit, denominator, uncertainty where applicable, and material limitations. Do not turn an estimate into a promise. Where a conclusion could trigger a major budget, product, or data-use change, define the follow-up validation before the meeting rather than adding it after stakeholders have committed to a preferred narrative.
FAQ
Is this the same as causal impact?
Not necessarily. Descriptive measurement summarizes recorded data under stated rules. Causal impact requires a suitable comparison design and explicit assumptions about what would have happened otherwise.
Can we use this in an A/B test?
Usually, if the data flow preserves pre-treatment eligibility, stable assignment, comparable outcome collection, and a pre-specified analysis. Treat derived reporting as a diagnostic unless it is the planned outcome measure.
What should be documented?
Document the decision, population, unit, events, identifiers, transformations, windows, exclusions, access and retention controls, owners, and known coverage limits.
How do privacy limits affect measurement?
They can reduce event or identity coverage and alter which journeys are observable. Measure and disclose the limitation rather than silently compensating with uncontrolled collection.
When should a team seek specialist advice?
Seek the organization’s privacy, security, legal, and data-governance specialists when a design affects regulated data, sensitive categories, vendor sharing, retention, rights requests, or material risk. This page is educational and not legal advice.
Summary
Identity Graph should be used as a clearly scoped measurement concept, not as a shortcut to certainty. Define the data contract, preserve provenance and coverage limits, keep experimentation comparisons valid, and make high-stakes decisions with evidence that matches the causal question. Review privacy, security, and governance requirements with the appropriate specialists.