What should an identity graph SLA include?

At minimum: the match rate measured against your full file, the precision behind it including the misidentification rate, whether each guarantee is deterministic or probabilistic, and a refresh cadence with a decay curve. A match rate without those four is a marketing number, not a service level.

Why does the denominator matter for a match rate?

Match rate is matched records divided by a reference population. Measured against a vendor's own addressable panel it looks excellent, but measured against your full customer file, including stale and partial records, it is lower and more honest. Always require the rate computed on your data.

Is a higher match rate always better?

No. Coverage and precision trade off. Probabilistic matching raises the match rate but also mislinks more records through household bleed and device churn. A higher rate with an unstated misidentification rate can be worse than a lower, cleaner one.

How do I account for identity graph decay?

Ask for a decay curve showing how much of the match rate survives at 30, 60, and 90 days, and put the refresh cadence in the contract as a hard term. Identifiers reset and devices churn, so a signing-day match rate degrades unless the graph is rebuilt on a stated schedule.

What is the best way to verify a vendor's match rate?

Run a seed-match test on a representative slice of your own file before signing. You hand the vendor the records, they return matches, and you measure both the rate and the accuracy yourself. It is the only match rate that describes your data rather than the vendor's panel.

Identity Graph SLAs: Match Rate and Decay Terms

What an Identity Graph SLA Should Actually Specify

The phrase on most identity decks is a single number: a 90 percent match rate. On its own it guarantees nothing, because it hides three questions the buyer needs answered. Against what population was it measured? How many of those matches are correct rather than merely returned? And how long does the number hold before device churn erodes it? A real service level answers all three, in writing, before signing. The strongest identity vendors have already moved this direction, shifting from asserting scale to documenting method against shared industry standards and guidelines.

Coverage is also per population, not a single global figure. GSDSI's identity graph resolves roughly 200M+ US MAID-HEM pairs and about 700M+ international device identifiers across 150+ countries, and a match rate against a US CRM file and a match rate against a global mobile audience are two different SLAs with two different denominators. A contract that quotes one number for both is quoting the more flattering one. For the product-level view of what the graph covers, see the MAID feed.

Match Rate Is a Ratio, So Name the Denominator

A match rate is matched records divided by a reference population. Change the population and the same graph reports a very different number. Measured against the addressable subset the vendor already knows it can hit, the rate looks excellent. Measured against your full customer file, including the stale and partial records that make identity hard, it drops to something you can plan against. The number you want in the SLA is the second one, computed on your data.

What a single '90% match rate' can mean under different denominators

Denominator used	What is being measured	Buyer risk
Vendor's own addressable panel	Coverage within records the vendor already resolves	Flatters the vendor, tells you nothing about your file
Clean subset of your file	Match rate on records that were easy to begin with	Excludes the hard records that drive the real gap
Your full customer file	Coverage across all records you actually need matched	The honest number, and usually the lowest
The addressable universe	Share of a market the graph can reach at all	A reach claim, not a per-file guarantee

This is why a seed-match test on your own data beats any quoted figure. You hand the vendor a representative slice of your file, they return matches, and you measure the rate and the accuracy yourself. For the coverage and diligence framing behind that test, see MAID and identity graph coverage, match rates, and compliance diligence.

Precision, Not Just Coverage

A higher match rate is not automatically better, because coverage and precision trade off. Deterministic links, built on shared logged-in identifiers, are fewer but reliable. Probabilistic links, inferred from signals like IP and behavior, extend coverage but introduce mislinks: two people merged into one identity, or one person split across several. A privacy-preserving record-linkage study measured this trade directly, reporting sensitivity, specificity, and misidentification rates rather than a single hit rate, because the misidentification rate is where probabilistic matching quietly costs you.

The practical failure modes have names. Household bleed assigns one member's behavior to another under the same roof. Device churn attaches a reset advertising identifier to the wrong person. Both inflate coverage while degrading precision, and neither shows up in a match rate that counts links without checking them. An SLA worth signing states the precision target alongside the coverage target, and says which link type, deterministic or probabilistic, each guarantee applies to. This is the discipline that makes identity resolution usable for targeting and measurement rather than a source of quiet error.

Decay Curves and Refresh Cadence

A match rate is a snapshot, and the snapshot ages. Mobile advertising identifiers get reset, cookies expire, people change devices and email addresses, and every one of those events breaks a link the graph was counting. A rate measured on signing day is the best it will ever be under a static file. That is why the refresh cadence belongs in the contract as a hard term, not a footnote: how often the graph is rebuilt, how quickly a reset identifier is retired, and what the expected decay looks like between refreshes. GSDSI covers the mechanics in how fast MAID and HEM freshness degrades.

Ask the vendor for a decay curve, not just a refresh promise. A curve tells you how much of the match rate survives at 30, 60, and 90 days, which lets you plan campaign windows and measurement periods around real freshness instead of the signing-day peak. A vendor who cannot produce one is telling you they do not measure their own decay.

Contract Terms to Negotiate

Translate all of this into clauses. The following belong in an identity graph SLA, and their absence is itself a finding:

Match rate stated against your full file, measured by a seed test you control, not the vendor's panel.
Precision target alongside coverage, with the misidentification rate named, not just the hit rate.
Link type disclosed per guarantee: which numbers are deterministic and which are probabilistic.
Refresh cadence as a hard term, plus a decay curve at 30, 60, and 90 days.
Coverage stated per population, so a US figure and an international figure are not blended into one.
Provenance and permitted use for every identifier key, so the match is compliant as well as accurate.

A graph that can commit to these terms is one you can build audience targeting and cross-channel measurement on with a known error budget, instead of a match rate you cannot audit. To scope a seed test against your own file and coverage needs, contact GSDSI.

Frequently Asked Questions

What should an identity graph SLA include?: At minimum: the match rate measured against your full file, the precision behind it including the misidentification rate, whether each guarantee is deterministic or probabilistic, and a refresh cadence with a decay curve. A match rate without those four is a marketing number, not a service level.
Why does the denominator matter for a match rate?: Match rate is matched records divided by a reference population. Measured against a vendor's own addressable panel it looks excellent, but measured against your full customer file, including stale and partial records, it is lower and more honest. Always require the rate computed on your data.
Is a higher match rate always better?: No. Coverage and precision trade off. Probabilistic matching raises the match rate but also mislinks more records through household bleed and device churn. A higher rate with an unstated misidentification rate can be worse than a lower, cleaner one.
How do I account for identity graph decay?: Ask for a decay curve showing how much of the match rate survives at 30, 60, and 90 days, and put the refresh cadence in the contract as a hard term. Identifiers reset and devices churn, so a signing-day match rate degrades unless the graph is rebuilt on a stated schedule.
What is the best way to verify a vendor's match rate?: Run a seed-match test on a representative slice of your own file before signing. You hand the vendor the records, they return matches, and you measure both the rate and the accuracy yourself. It is the only match rate that describes your data rather than the vendor's panel.