Most identity graph evaluations happen once. A buyer or their agency runs a match-rate test, reviews methodology documentation, maybe stress-tests a seed file against a known customer set, and then signs. After that, the graph becomes infrastructure—trusted, invisible, and assumed stable.
It is not stable. Identity graphs decay, and they decay faster than most procurement cycles, measurement reviews, or vendor check-ins are designed to catch.
What Decay Actually Looks Like
Identity graph accuracy depends on a chain of signals—device IDs, hashed emails, postal records, login data, probabilistic bridges—each of which has its own half-life. Mobile advertising IDs turn over when users reset them for privacy, when they upgrade devices, or when apps are uninstalled. Hashed email linkages break when people switch providers, abandon addresses, or change behavior patterns that probabilistic models used to infer the connection in the first place. Postal records drift with household moves, which the U.S. Census Bureau has historically measured at roughly 12 percent of the population annually.
None of these events are broadcast to the graph. They accumulate silently. The record that mapped Device A to Email B to Household C six months ago may now be partially or entirely wrong, but the graph has no mechanism to flag that record as degraded—it simply continues to serve it until a refresh cycle catches the inconsistency, if one runs at all.
The practical result: a graph you evaluated for accuracy in Q1 may have meaningfully different accuracy characteristics by Q3, with no visible signal to the buyer that anything has changed.
Why Audit Timing Compounds the Problem
The standard vendor evaluation process is front-loaded. Accuracy audits, match-rate reviews, and methodology assessments happen before contract signature. After that, buyers typically receive quarterly reporting on reach and scale metrics—numbers that measure how many IDs are in the graph, not how many of those IDs are still correctly resolved to real, reachable, in-market people.
This creates a structural gap between when quality was measured and when it matters. A buyer running a Q4 acquisition campaign is activating against accuracy assumptions established in Q1, filtered through whatever the vendor's internal refresh cadence has caught in between—which varies significantly by graph provider and is rarely disclosed in detail in standard reporting packages.
The problem compounds when measurement is also run through the same graph. If your targeting, frequency management, and conversion attribution all route through a single identity layer, decay in that layer doesn't just affect who you reach—it corrupts the measurement of whether reaching them worked. You're not just spending against a degraded map; you're grading your own performance against that same degraded map.
The Signals That Surface Decay—If You Know to Look
Decay rarely announces itself, but it does leave traces in campaign data that most buyers don't connect back to graph quality.
Auction win rates that hold steady while conversion rates fall can indicate that you're reaching the IDs you intended but those IDs no longer belong to the people your model expected. Frequency cap behavior that looks correct in reporting but produces user-level complaints or brand safety flags can indicate multiple real people being collapsed under a single degraded identity. Retargeting pools that shrink faster than your known customer churn rate would explain can indicate that device-to-person bridges are breaking faster than new ones are being built.
These are not proof of graph decay on their own—each has other explanations. But when they appear together, the first diagnostic question should be: when was this graph's accuracy last independently validated, and against what population?
What a Decay-Aware Evaluation Process Looks Like
The practical fix is not to distrust identity graphs—it's to treat them like any other data asset with a known degradation rate, which means building validation into ongoing operations rather than treating it as a one-time pre-contract exercise.
A few structural adjustments matter here.
Require refresh cadence disclosure at contract stage, not just methodology disclosure. How often the vendor reconciles device IDs against current signals, re-validates email-to-household bridges, and runs household mobility corrections is as material to your accuracy assumptions as the methodology used to build those bridges in the first place. Vendors that can't answer this question with specifics are, by default, asking you to trust a static snapshot.
Run lightweight accuracy probes on a scheduled basis, not just at onboarding. This doesn't require a full audit. A subset of your known CRM records—ones where you have high-confidence ground truth on current email, device, and address—can be run against the live graph on a quarterly basis to check whether resolution accuracy has drifted. A meaningful degradation in that probe population is an early signal worth catching before it propagates through a full campaign.
Separate the graph powering your targeting from the graph powering your measurement wherever operationally feasible. When both functions run through the same identity layer, graph decay is self-concealing—errors in targeting and errors in attribution cancel each other in ways that can make degraded campaigns look stable. Using an independent measurement graph, even for a subset of campaigns, gives you a cross-check that pure same-graph attribution cannot provide.
Build contract language around accuracy floors, not just scale guarantees. Vendors routinely commit to audience size and match volume. Fewer commit to maintaining a minimum accuracy rate—defined as correct person-to-ID resolution—over the contract term. Pushing for that commitment, even if the floor is negotiated conservatively, changes the vendor's incentive structure around refresh investment.
The Broader Point
Identity infrastructure gets evaluated like a software license and managed like a utility. Neither frame fits. Identity graphs are closer to perishable data—accurate when fresh, increasingly unreliable as time and real-world change accumulate between the last reconciliation and the current activation.
Buyers who treat graph quality as a one-time procurement question and then activate against that assumption indefinitely are not making a technical error. They're making a timing error. The graph they're trusting today is not the graph they audited, and in a campaign where identity accuracy determines who you reach, what you exclude, how you pace, and how you measure—that gap has budget consequences that no amount of creative optimization or bid strategy adjustment can fully offset.