Insights · September 11, 2026 · Agent-led advertising
Agent-led advertising needs receipts, not theatre
Birch Reserve Editorial
Agent-led advertising needs a receipt
An advertising agent can plan quickly, compare options, and prepare a recommendation. None of that proves that a transaction should happen. The meaningful question is whether an outside reviewer can reconstruct what the agent knew, what it was allowed to do, what it committed, and what was actually delivered.
That distinction matters because adoption and performance are different claims. A product announcement can establish that a protocol was released. It cannot, on its own, establish that the protocol is broadly adopted. A dashboard can display clicks or conversions. It cannot, on its own, establish incremental value. Evidence standards should make those boundaries difficult to blur.
What AAMP 2.3 actually gives us
The IAB Tech Lab's July 30, 2026 release describes AAMP 2.3 as adding enterprise deployment, privacy diligence, pricing integrity, and standards for agentic advertising [1]. The concrete details are more useful than the headline. The release describes a privacy approval gate, enhanced pricing guardrails, deterministic and provable paths for spend commits with human approvals outside value-based thresholds, and inventory availability based on real rather than derived numbers [1].
Those are control requirements and protocol capabilities. They are not evidence that every agent uses them, that every implementation is correct, or that a campaign produces a return. The release itself is the right source for what AAMP 2.3 says it published; it is not a market census or an independent ROI study.
Five receipts a serious agent should leave
First, authorization: show the buyer's budget, permitted inventory, thresholds, and the exact point at which a person approved the spend. “The agent decided” is not an approval record.
Second, inventory identity: record the placement, seller, time window, unit, availability basis, and any exclusions. If an agent says inventory is available, the record should distinguish a live count from a modelled estimate. A number without provenance is a sales assertion.
Third, price and money movement: preserve the price input, fee treatment, currency, budget ledger, and final insertion order. Guardrails matter because an agent can be directionally right while committing to the wrong commercial terms. A deterministic commit should be reproducible from the record.
Fourth, diligence: retain the privacy, brand-safety, and policy checks that allowed the recommendation to proceed, including the version of each rule. A green status without the rule, timestamp, and input is not an audit trail.
Fifth, delivery: reconcile the authorized plan to what was served. Keep corrections, no-fill events, creative versions, and measurement windows visible. Delivery evidence should be separable from outcome evidence; a delivered impression is not an incremental sale.
Adoption evidence versus performance evidence
To substantiate adoption, ask for independently verifiable implementations, dated deployment records, conformance evidence, or a denominator that explains how many eligible buyers and sellers use the capability. The existence of a release, contributor quote, or integration option is not that denominator.
To substantiate performance, ask a different question: compared with what? A pre/post chart can move because demand, seasonality, pricing, or selection moved. Better designs use randomized holdouts where feasible, or a transparent quasi-experimental comparison with pre-specified outcomes and a stated observation window. Report eligible opportunities, exclusions, missing data, uncertainty, and adverse results. Do not convert a click rate into causal customer value.
For Birch Reserve, this means a buyer should receive an understandable insertion order and delivery record before hearing a performance story. Agentic tooling can reduce friction, but it should increase the quality of the receipt. The standard is not autonomy for its own sake; it is accountable execution that a buyer, host, and reviewer can all inspect.
Source limits
The only external source here is a standards-body release. It supports a precise description of AAMP 2.3's published controls. It does not demonstrate broad adoption, independent implementation quality, campaign performance, or Birch Reserve results. Those claims require separate evidence.