A transparent auction can show why one bid won without showing whether the bidders made independent decisions. These are two different questions. The first concerns auction rules, supply paths and recorded outcomes. The second concerns the decision processes behind apparently distinct participants, and whether the available evidence can distinguish independent responses from common influences.
For an exchange, publisher or agency, both questions matter. Neither substitutes for the other.
What transparency work already establishes
Auction transparency has made substantial progress. In January 2026 the Media Rating Council (MRC) issued its final Digital Advertising Auction Transparency Standards. They address auction administration, participant qualification, pricing and winner determination, reporting, change disclosures and auditing, and they explicitly discuss complex AI- and machine-learning-driven auctions. Adoption is voluntary: operators opt in to an independent audit. The reporting framework permits disclosed, representative sampling and some aggregation; it does not guarantee that every participant receives a complete record of every eligible decision opportunity. [1]
These are meaningful protections. A buyer who understands pricing adjustments, quality scores and win/loss explanations can challenge a disputed charge or a misunderstood rule. An independent audit against MRC requirements can test the specified auction-transparency controls. That work should not be dismissed because it answers a different question.
Other tools cover adjacent territory. IAB Tech Lab’s ads.txt, sellers.json and SupplyChain object help buyers identify authorized sellers, intermediaries and supply paths. [2] OpenRTB specifies bid requests and responses, explicit no-bid signaling and distinct win, loss and billing notices. It is a protocol, not proof that a particular operator retained every relevant event. [3] Prebid.js exposes useful auction events, including bid requests, responses and timeouts, but browser-side events are not automatically a complete account of an exchange’s upstream selection process. [4]
Supply-path optimization (SPO) adds genuine value. Jounce Media, for example, describes evaluation of seller authorization, path directness and inventory quality. [5] Those questions influence buying efficiency and accountability, but they address a different question from whether bidder policies are independent.
Fraud and invalid-traffic detection, event reconciliation and ordinary analytics likewise matter. They help establish whether inventory is genuine, who sold it, what occurred in the recorded auction and which delivery or financial events join correctly. Passing those checks is not, by itself, a finding that bidders decided independently.
The selection problem: who was allowed to decide?
Imagine six independently owned bidders eligible to compete for a publisher’s impressions. For one audience segment the exchange invites all six. For another it routes requests to just one, perhaps because of latency constraints, policy or expected response quality.
A winner-only report might show that the same bidder wins the second segment repeatedly. A bid-level report might show that the invited bidder responds consistently. Neither reveals how the five excluded bidders would have acted had they been invited.
Now imagine two underlying markets. In one, the excluded bidders would use genuinely different strategies. In the other, they would all make nearly identical decisions. If the router never exposes those policies, both markets can produce exactly the same invitation and bid records. More observations drawn from the same restricted invitations cannot identify the missing counterfactual behavior.
Exchange Optics tested this in a controlled synthetic experiment: six fixed bidder policies, forty independent market tapes and 1,024 publisher opportunities per run. In one world the six policies were genuinely distinct; in the other they had converged. Both used the same selective router, which invited bidders on only one audience segment. The invitation, response and winner records were byte-identical across the two worlds. Among the bidders actually invited, effective decision diversity (how many independent decision-makers the bids behave like) measured 1 in both worlds, which is the correct answer for that population. Across all six bidders on all opportunities, it would have been about 5.5 in the distinct world and 1 in the converged world. Records cannot reveal actions that were never taken. When we added a 25% random probe of the excluded segment, the two worlds became distinguishable.
This is a mechanism test on stylized bidders, not an allegation or a finding about any real exchange. Critically, the behavioral target for actually invited bidders can still be properly defined; it must not silently be relabeled as the diversity of all eligible bidders.
The missing distinctions accumulate through the event lifecycle: original publisher opportunity; admission; eligibility; routing or invitation; confirmed request receipt; positive bid, explicit no-bid, timeout or unknown response; accepted bid; winner; and delivery or settlement. A timeout is not a deliberate no-bid. An uninvited participant did not make an observed response. A winning bid does not necessarily prove that an impression was delivered or billable. [3]
Three practical observation tiers
| Evidence | Useful answer | Cannot show by itself | Next information |
|---|---|---|---|
| 1. Winner or accepted-positive records | Recorded winner mix, accepted bids and price patterns | Losing responses, true no-bids, invitation denominator or market-wide policy independence | Losing bids, source IDs, retention and sampling definitions |
| 2. Reconciled invitations and responses | Conditional behavior among actually invited participants, when dispatch, receipt, explicit no-bids and missing states are independently sound | Policies of never-invited candidates; latent vendor or controller identity | Candidate census, exclusions, pre-decision context, independent stage counts |
| 3. Independently reconciled opportunities, eligibility, routing and provenance | Better-defined invited population, selection diagnostics and potentially qualified conditional measurement | Counterfactual policy behavior where invitation probability is zero; automatic proof of controller independence or financial harm | Permissioned tests that give excluded bidders a nonzero chance of being invited, authenticated controller lineage if needed, and separate settlement/causal evidence |
These are evidence tiers, not certifications. Even the third may leave the answer unresolved. To infer uninvited actions, an operator might need lawful, governed exploration with known nonzero selection probabilities and valid comparisons. Statistical weighting cannot reconstruct a segment of the market that is never observed.
Behavioral resemblance has several explanations
Suppose three automated buying systems respond similarly to a public change in conversion expectations. They may independently react to the same information. Alternatively, nominal accounts might rely on a shared optimization service. Or an exchange router might repeatedly expose them to the same narrow opportunities.
Those explanations have different implications, but bid correlation alone cannot separate them. Context matters: audience, floors, budgets, time, market shocks, pacing and changes in algorithms or auction rules can shape observed actions. Conditioning on the wrong information, including variables determined after selection, can introduce distortion.
Even complete behavioral observations do not necessarily reveal how many distinct software controllers, vendors or owners operate behind the bids. One shared platform can support distinct client policies. Separate systems can independently make similar choices. A measure of marginal behavioral resemblance is not proof of joint independence, controller provenance, common ownership, coordination or legal collusion.
That is why effective decision diversity is better treated as a family of carefully specified measurements, not a universal headcount converter. The actor, population, context, decision regime and measured behavior must be stated before interpreting any estimate.
Some data environments justify only narrow descriptive comparisons. Others may support controlled prospective tests and independent provenance checks.
More transparency can still lead to an honest limit
A sound assessment begins with what the records prove. Event reconciliation can locate broken joins. Source audits can reveal sampling gaps or repeated physical events. Invitation censuses help distinguish router selection from observed bidder behavior. Regime-aware comparisons can challenge some common-shock explanations.
But a signed log authenticates recorded bytes, not necessarily an entire upstream universe. Likewise, an apparent decrease in decision diversity does not automatically establish higher prices, publisher losses or economic damage. Fees, budgets, auction design, billing and settlements may change the direction of effects. Those claims require separately specified outcomes and defensible comparisons.
A usable diligence checklist
Before making a decision-independence claim, ask:
Auction transparency, supply-path assurance and valid-traffic verification establish critical facts about transactions. Effective Decision Diversity adds a distinct economic question: how independently do participants behave within the opportunities the marketplace actually makes available? Where records omit that opportunity population, the missing evidence becomes a concrete instrumentation and verification priority.
- Who is the actor? Legal advertiser, DSP seat, bidder endpoint, exchange adapter or independently authenticated controller?
- What is the population? Winners, accepted bids, responders, actual invitees or all eligible opportunities?
- Where does selection occur? Which admission, eligibility and routing rules exclude potential respondents?
- Are response states complete? Can the data distinguish no-bid, timeout, failed dispatch, missing response and never invited?
- Who verifies the source? Are pre-retention counts, sampling, physical clusters and loss rates checked beyond self-report?
- Which conditions changed? Are pre-decision context, shared shocks, policy updates and regimes documented?
- What is actually claimed? Behavioral resemblance, common infrastructure, controller identity, coordination and financial impact each require different evidence.
- What would force refusal? Is “not identifiable from these records” an acceptable finding?
What this means in practice
The useful question is not whether a marketplace is “diverse.” It is which of the three evidence tiers its records can support today, and what it would take to reach the next one.
Exchange Optics is developing its Effective Decision Diversity instrument with Founding Design Partners. A first engagement is deliberately narrow:
- Map the lifecycle. Identify which stages (opportunity, eligibility, invitation, receipt, response state, winner, delivery) are recorded, and which are missing or self-reported.
- Place the records on the tiers. Reconcile invitations and responses to show what can be measured now: winner patterns, conditional behavior among invited bidders, or neither.
- State the defensible finding. That may be a scoped measurement. It may also be “not identifiable from these records,” with a specific list of the instrumentation that would change the answer.
If you operate or buy through an exchange and want to know which tier your records support, contact Exchange Optics.
Primary sources
- Media Rating Council, Digital Advertising Auction Transparency Standards, January 2026, sections 1, 3–6: https://www.mediaratingcouncil.org/sites/default/files/Standards/MRC_Digital_Advertising_Auction_Transparency_Standards_FINAL.pdf
- IAB Tech Lab, ads.txt and sellers.json / SupplyChain: https://iabtechlab.com/ads-txt/ ; https://iabtechlab.com/sellers-json/
- IAB Tech Lab, OpenRTB 2.6 specification, sections 2–4: https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md
- Prebid.org, pbjs.getEvents() documentation: https://docs.prebid.org/dev-docs/publisher-api-reference/getEvents.html
- Jounce Media, The Marketer’s Guide to SPO: https://jouncemedia.com/resources/marketers-guide-to-spo
The examples in this article are conceptual or drawn from controlled synthetic research. They illustrate measurement boundaries, not a finding about any named exchange.