What is a self-attributing network (SAN)?

How a SAN differs from an ordinary network

An ordinary network sends click and impression records to an MMP, which matches them against conversions and decides credit. The network does not decide; it supplies evidence.

A SAN does not supply the evidence. It performs the match internally and returns a result — this many installs, this many events. The MMP receives a claim and reconciles it against claims from other sources, but cannot independently verify the underlying match.

The largest platforms operate this way, which means a large share of most media budgets is measured by the party being paid.

Why SANs exist

The stated reason is user privacy, and it is genuine: exporting per-user interaction records to third parties is exactly what a decade of regulation has been narrowing.

The commercial reason is also genuine. Interaction data is the platform's most valuable asset, and exporting it would let competitors and advertisers model its audience. Both reasons point the same way, which is why the practice is stable regardless of which one is emphasised.

How to work with SAN numbers

• Treat them as claims, not counts. Useful for optimising inside the platform, unreliable for choosing between platforms.
• Reconcile against your own totals. Your order or install record is the denominator that constrains every claim.
• Verify windows. SANs default to generous windows, and comparing across unequal windows is meaningless.
• Hold out. Incrementality is the only measurement a SAN cannot influence, because it does not ask the SAN anything.

AI surfaces are self-attributing by default

Every AI advertising platform today is structurally a SAN: it owns the surface, decides what was delivered, and reports the outcome, with no independent counting layer in existence yet.

That argues for the same discipline — reconcile against your own records, verify window definitions, and run a holdout. It is worth noting the error currently runs the other way, though: because so much AI influence produces no click, self-reported AI numbers tend to under-count rather than over-claim. The remedy is identical.