There is a number most media buyers never see: the rate at which the identity graph powering their campaign erodes between the day a file is onboarded and the day the last impression is served.
Onboarding reports present a snapshot. They tell you how many CRM records resolved to addressable identities at a specific moment, typically within 24 to 72 hours of file ingestion. That number becomes the foundation for audience sizing, budget allocation, and reach projections. It also becomes irrelevant almost immediately — because identity graphs are not static infrastructure. They are probabilistic, constantly refreshed systems built on signals that expire, rotate, and contradict each other in real time.
What Graph Degradation Actually Means in Practice
An identity graph connects persistent identifiers — hashed emails, device IDs, household address clusters, cookie adjacencies, mobile ad IDs — to individual-level profiles. The quality of any given connection is a function of the recency and frequency of the signals that created it. When a person switches devices, changes email providers, moves households, or simply goes dark on the platforms contributing signal to the graph, their profile in that graph weakens. Connections that were high-confidence at onboarding become low-confidence a few weeks later, and eventually break entirely.
For a 30-day campaign, this is a manageable but real problem. For a 90-day always-on program, it is a structural one. The audience a buyer paid to reach on day one is not the same audience the DSP is attempting to reach on day sixty — and no standard reporting surfaces that divergence.
The decay is not uniform. It is concentrated in the parts of the graph that are most dependent on volatile signals. Device IDs rotate on iOS after each app session for users who have limited ad tracking. Third-party cookie-based connections in graph maintenance pipelines are already thinning across browsers. Email-based connections are only as durable as the hashing logic and the freshness of the email record itself. The people most likely to fall out of the addressable graph mid-campaign are also, often, the people whose identities were resolved through probabilistic inference rather than deterministic match — meaning the weakest connections degrade fastest.
Why Standard Reporting Hides This
Most activation pipelines report audience size at the segment level, not the identity resolution level. The segment shows a stable count because it is refreshed with replacement matches as individual connections drop — the graph operator continuously backfills decayed profiles with newly resolved ones to maintain segment integrity. From the buyer's perspective, the audience looks healthy. What they cannot see is whether the people being reached on day forty are the people from their CRM file, or increasingly diluted approximations of them.
This distinction matters enormously for campaigns built around suppression, customer exclusion, or lookalike seeding. If a suppression list was built against a high-confidence graph state and the graph has since decayed for a portion of those records, the suppression logic is operating on stale linkage. People intended to be excluded re-enter the reachable pool not because the list failed but because the identity connection that enabled suppression no longer resolves. The buyer never learns this happened.
The Measurement Consequence
Graph decay introduces a specific kind of attribution noise that is nearly impossible to isolate after the fact. When a conversion is credited to an exposed individual whose identity connection degraded mid-campaign, there is no reliable way to confirm that the impression driving that conversion actually reached the intended person versus a graph-adjacent match who happened to share a household IP or a probabilistically linked device.
This is not a hypothetical edge case. In households with multiple adults, shared devices, or frequent network switching, the graph is doing significant inferential work to maintain identity continuity. That inferential work becomes less accurate over time as contributing signals age. The buyer who runs a post-campaign match-back analysis is measuring against a graph state that may look structurally similar to the original but has replaced a non-trivial share of individual connections underneath.
What Buyers Should Demand Instead
The practical response is not to stop using identity-resolved audiences. It is to require transparency at a different point in the pipeline than buyers currently ask for it.
First, request identity connection confidence scores at segment ingestion, not just match rate totals. A segment where 85% of resolved profiles carry a high-confidence deterministic signal behaves very differently in mid-campaign decay than one where 60% of matches were probabilistically inferred. Both can show identical match rate headlines.
Second, build campaign flights with explicit re-resolution checkpoints. For programs longer than 45 days, require the graph operator or onboarding partner to confirm identity connection refresh rates and report what percentage of the original onboarded file remains actively resolved at the midpoint. This is operationally achievable and rarely requested.
Third, separate your suppression logic from your targeting logic in terms of how often each is revalidated. Suppression lists need more frequent identity revalidation than targeting segments, because the consequence of a false positive — reaching someone you intended to exclude — is more damaging than the consequence of a decayed targeting connection dropping from the pool.
Finally, when evaluating identity graph vendors or onboarding partners, ask specifically about maintenance signal sources and refresh cadence. A graph that relies heavily on third-party data cooperative feeds will degrade differently than one anchored to first-party publisher log-ins. Neither is automatically superior, but the decay profile is different, and it should be factored into how you structure campaign measurement windows.
The Underlying Discipline
Identity resolution is most commonly evaluated as a procurement question — which vendor resolves the most records at the lowest cost. It should also be evaluated as a temporal question: how long does that resolution remain actionable, and what happens to campaign logic when it stops being so?
Buyers who treat the onboarding report as the end of identity due diligence are measuring a single frame of a process that runs for the duration of every campaign they activate. The gap between that frame and the reality of what the graph looks like sixty days later is where a significant share of performance variance lives — and where most post-campaign analysis never looks.