Device Graph Decay: How Fast MAID and HEM Freshness Degrades

Every identity graph carries an expiration date that vendors rarely print on the catalog sheet. A MAID graph marketed at 250 million historical identifiers may carry 175 million observable-today devices after Apple ATT, partner churn, and consent-lane restructuring. HEM graphs inflate with dormant addresses that no longer read mail. Buyers licensing MAID Feed, Core Email File, or audience targeting programs should model decay explicitly: refresh cadence, observable-cohort diagnostics, and rate-card denomination, or renewal math surprises the team that signed on headline cohort size.

Decay is not a product defect: it is the nature of identifiers in 2026. The wrong response is denial: quoting activation programs against peak historical MAIDs forever. The right response is operational: observable-cohort reporting, refresh SLAs matched to campaign windows, and model adjustments when composition shifts. Read this alongside MAID graph economics and data refresh cadence.

MAID Decay Under ATT and Panel Churn

MAID churn accelerates when users reset advertising identifiers, uninstall apps, or opt out via ATT prompts. iOS-heavy panels decay faster than Android-utility panels; gaming and hyper-casual sourcing often shows steeper curves than news or utility apps. Graphs exposed to FTC 2024 location enforcement lost partner lanes in step-functions, not smooth monthly drift. The FTC order banning InMarket from selling precise location data and the finalized FTC order against Mobilewalla moved other supply lanes the same way. Require vendors to report observable-today MAIDs with last-seen timestamps, not only cumulative historical IDs.

Match-rate decay is distinct from panel decay. A graph may retain identifiers but lose matchability to your seed file as sourcing mix changes. Seed match tests on quarterly cadence catch both problems. See seed match testing and 5 questions before licensing a MAID feed.

HEM Decay and the Dormant Tail

Email addresses persist in databases long after users stop engaging. Headline HEM counts include mailboxes that bounce, role accounts, and addresses still valid but unread. Honest freshness metrics count observable-engagement HEM: recent opens, clicks, authenticated sessions, or CRM interactions within a defined window. Core Email File buyers should require engagement-tier fields and validation status, not only syntax-valid addresses.

• Active engagement: opens, clicks, or site auth within 90 days (vendor-defined windows vary).
• Valid but dormant: deliverable but no recent engagement; risky for outreach.
• Invalid or suppressed: bounces, complaints, role accounts; must not enter activation.

Pricing Against the Right Denominator

Re-denominate rate cards against observable-today graph size. A vendor at $X per historical MAID with thirty percent decay gap is materially more expensive than a vendor at $1.2X per observable-today MAID. Finance teams that flip the denominator often reorder finalists. IAB Tech Lab identity guidance treats observable-reachable scale as the relevant measure for activation buyers. Contract for monthly observable-cohort reports and price true-ups tied to documented composition shifts, not surprise renewals.

Refresh Cadence by Use-Case

Programmatic activation and fraud/identity verification need real-time to weekly identifier refresh with monthly cohort diagnostics. Campaign measurement with four-week windows tolerates weekly graph refresh if attribution models document decay adjustment. Audience planning may tolerate monthly refresh; quarterly refresh is stale for forward spend decisions. CDP enrichment should be weekly minimum: quarterly graphs ship dead joins to downstream campaigns.

Operational Response: Monitoring and Remediation

Set amber and red bands from pilot statistics: observable MAIDs, match rate to seed, last-seen distribution, iOS/Android mix. Alert when bands breach before campaigns underperform. Remediation paths include refresh frequency upgrades, seed expansion, graph vendor swaps, or model retraining, not only bid increases on a decaying audience. Document decay assumptions in media mix models and incrementality studies so results remain comparable quarter to quarter.

Identity programs that ignore decay optimize for slide-deck reach while activation teams pay for ghosts. Treat freshness diagnostics with the same rigor as POI refresh and panel QA. They are the same class of operational risk.

Decay Is Not Uniform Across the File

A single decay rate quoted for a graph is an average over populations that behave very differently. Identifiers attached to heavily used devices persist and refresh differently from those seen rarely, and the segments a buyer most wants to activate often sit in the fastest-decaying part of that distribution. The average therefore flatters exactly the cohorts a campaign depends on.

Ask for decay broken out by observation frequency and by the age of the link rather than as one headline figure, then compute it on your own matched population once a delivery lands. The measurement is simple: take a matched cohort, re-match it after a period, and count how many links survive unchanged. That number, computed on your records rather than the vendor's, is the one that predicts what your next campaign will actually reach.

Telling Decay Apart From Rotation in Your Own Deliveries

An observable-today count that holds flat month over month is consistent with two very different panels. One is stable. The other is losing members and replacing them at roughly the same rate, which is a much shorter shelf life for anything you build on it. A net figure cannot separate those, and neither can a vendor attestation. Cohort survival can, and you can measure it from files you already receive. Freeze the identifier list from one delivery, keep only its hashes, then test membership in each later delivery and report survival by months since the freeze.

That test produces two curves worth reading separately. Survival tells you how long an audience built today stays reachable, which is the number that should drive activation windows and refresh SLAs. Gross additions tell you whether losses are being backfilled with genuinely new supply or with identifiers that appeared in an earlier snapshot, dropped out, and returned. Split additions into never-before-seen and returning. A heavy returning share usually means an intermittent partner lane rather than growth, and it inflates the headline in exactly the months a lane comes back online.

Audit the provenance of the last-seen field before you trust any of it. If that timestamp is written at build or delivery time rather than at observation time, a static extract re-shipped on schedule looks perfectly fresh, and every freshness metric derived from the field reads clean while the panel underneath is frozen. Two checks settle it. Look at whether last-seen values cluster on delivery dates, and look for identifiers whose last-seen advances across deliveries with no corresponding activity. Then ask the vendor which process writes the field. The last_seen field in the MAID-to-HEM identity graph is meant for decay-aware filtering, so its observation-time meaning matters more than the file's delivery date.

Run the same freeze on the subset that actually matched your seed, not only on the full graph. Whole-graph decay is a vendor property. Matched-cohort decay is what your spend rides on, and the two diverge whenever the sourcing mix shifts away from your customer base. See why vendor match rates are not comparable.