Advertising auctions count bidders and transactions. Those counts don’t tell you how many distinct decision processes are actually shaping competition. Effective Decision Diversity is our term for that missing dimension of market structure.
Corporate identities and bid counts describe participation, not the diversity of the decisions underneath. Measuring effective diversity requires comparable opportunities and a close look at rational responses, shared dependencies, and how the auction selects who gets to bid. The practical task is to separate those possibilities, and to say plainly when the evidence can’t.
A hypothetical auction with two explanations
Imagine a marketplace with 24 separately represented bidding seats. Over a month, winning shares stay dispersed and the number of active seats barely changes. An analyst nevertheless notices that many bids begin moving together.
One explanation is a genuine common dependency: several independent companies use the same optimization service, rely on similar data, or update their systems at about the same time. Their decisions become more alike without any change in ownership.
A second explanation: all 24 seats keep using different policies. A supply-quality change shifts the value of the same inventory for everyone. Meanwhile, the exchange’s routing increasingly invites only the seats likely to respond to certain opportunities. The analyst sees a narrower sample of bids with stronger apparent similarity, even though the underlying policies haven’t converged.
Both scenarios are hypothetical. The point is that participation counts, winner shares, and bid similarity look the same under either one.
Four things that get mixed up
1. Participation: who showed up. The count of firms, seats, or accounts, at a stated level. A DSP seat is not necessarily a separate advertiser, controller, or model.
2. Observed behavior: what was actually recorded. Bids can move together because of shared software, shared market conditions, or because the exchange only sent certain opportunities to certain bidders.
3. Decision diversity: how different the underlying bidding policies are across a defined set of opportunities and market conditions. Two bidders can follow the same policy independently, and one shared service can run very different policies for different clients. So this is neither a count of controllers nor proof that actions are linked. The answer depends on how you define the actor, the decision context, the weights, and the time window.
4. Lineage and coordination: who controls what, and whether anyone agreed to anything. Bid data alone generally can’t answer this. It takes authenticated provenance and, for any question of misconduct, entirely different evidence and legal analysis.
Each question needs its own evidence. A good report says which one it measured.
The biggest missing denominator: who had a chance to bid?
A winner log records the end of a process, not its beginning. It may omit losing submissions, explicit no-bids, timeouts, bidders that were never sent a request, and eligible bidders excluded by routing. Even an all-submitted-bids file doesn’t show which participants were invited but stayed silent.
Exchange logs are valuable, but they record what was submitted and what happened next, not what never entered the record. Index Exchange’s Client Audit Logs, introduced in 2019, give publishers the top five submitted bids per auction and give buyers all of their submitted bids (source). Its current documentation describes a bid event as a DSP submitting a bid response, notes that files arrive within 36 hours, and says they are purged after 14 days (source). Those logs are a strong foundation. They do not show which eligible bidders were never invited, timed out, or chose not to respond.
A useful investigation traces opportunity, admission, eligibility, routing, dispatch, receipt, action, winner, and settlement, and reconciles identities, sampling, and missing records.
A timeout is not a no-bid. A bidder that was never invited is not a bidder that chose not to compete.
Similarity is evidence to investigate, not a diagnosis
Two bidding policies may react similarly to a change in inventory quality, and that can be perfectly rational, independent behavior. Adjusting for visible variables such as format, floor, geography, or time helps describe what similarity remains, but it can’t remove unmeasured common shocks. Conditioning on variables shaped by eligibility or selection can even create misleading relationships.
If an eligible bidder had zero chance of invitation, no amount of winner data reveals its policy. A statistic about observed winners can’t quietly stand in for all eligible bidders.
Useful tests include comparing like-for-like contexts, examining changes around documented model or routing updates, reconciling opportunity counts across independent records, repeating the analysis by venue and regime, and asking whether the same evidence supports more than one explanation. Where lawful and feasible, prospective tests with known invitation probabilities can answer what passive logs cannot, with assumptions, disruption, and permissions disclosed. Short log-retention windows add to the case for designing tests in advance rather than reconstructing events afterward.
Sometimes the honest finding is “the evidence can’t tell.” That is a result: not a clean bill of health, and not an accusation.
Why it matters
Auctions work because bidders form their own judgments about what an impression is worth. Competition between those judgments is what discovers price.
Return to the 24-seat example. If those seats increasingly acted on the same few decision processes, the exchange would still report 24 participants and dispersed winner shares. The dashboard would look identical, but fewer independent judgments would be setting the price. That could mean less competitive pressure, and more exposure if a shared dependency changes or fails everywhere at once.
Whether that is happening in any given market is an empirical question. Today’s standard metrics can’t answer it, and answering it requires a defined comparison, reconciled settlement data, and a design that separates routing and demand changes from bidder behavior.
A complementary question, not a replacement for existing transparency
The industry is already working on real gaps. The Media Rating Council’s January 2026 auction transparency standards (cover marked “final draft”) set requirements for disclosure, reporting, and winner determination, and they encourage voluntary independent audits (source). The same document notes that AI and ML auction systems are hard to describe in simple terms, and says MRC plans further work on them. Supply-path optimization, invalid-traffic verification, and auction reconstruction address other questions that matter, such as path quality, impression validity, and how a price was computed.
The question Exchange Optics asks is narrower: within a defined set of opportunities, how much independent decision-making does the evidence actually support, and what explanations remain? A path can be efficient, an impression valid, and a price correctly computed, and that question can still be open. None of those checks answers it, and it doesn’t replace any of them.
Industry voices are also asking for visibility into AI-driven decisions. Goodway Group’s Andrea Kwiatek told AdExchanger in August that her agency is testing buyer and seller agents and wants verification on both sides (source). That describes demand for transparency. It does not say any exchange has lost decision diversity.
What makes an independent answer credible?
Independent assessment is more than a third-party logo. The assessor needs lawful access, evidence lineage, an appropriate source census, defensible methods, and the freedom to report adverse or unresolved results.
A paying sponsor creates a potential conflict, so the engagement has to manage it: pay for scoped work and documented findings whether or not they’re favorable, keep scientific control outside the sponsor, allow factual corrections without a veto over conclusions, and protect sensitive data and disclosure rights by contract. Qualified external review is added once reviewers are appointed.
Contracting for measurement is getting more attention too. In September 2026 the IAB released a draft Measurement Services Addendum, a standardized contract framework for measurement providers and their clients, with comments closed October 22 (source). It is a draft framework, not a mandate and not a validation of any method.
Exchange Optics has built and tested its method on controlled synthetic markets. The next step is validation on real exchange data under an agreed scope, with findings reported whether favorable or not. That is what our design-partner engagements are for.
Where to start
The starting question isn’t “which bidder is colluding?” It’s: what population of decisions can your authorized records actually describe, and what additional evidence would separate shared dependence from common conditions and selective visibility?
If you operate an exchange, run a bidder seat, or buy through one, we’re taking on a small number of founding design partners for a scoped diagnostic on a defined opportunity population. Contact Exchange Optics via the inquiry form.
References
- Media Rating Council, January 2026: Digital Advertising Auction Transparency Standards (marked final draft).
- Index Exchange, September 3, 2019 and May 4, 2026: Client Audit Logs introduction and CAL API documentation.
- Alyssa Boyle, AdExchanger, August 14, 2026: interview with Andrea Kwiatek, Goodway Group.
- IAB, September 22, 2026: Measurement Services Addendum v1.0 public comment release; October 20, 2026: Cross-Channel Measurement Decision Guide consultation.
Editorial note: Sources describe independent industry developments; none is an endorsement, customer relationship, or validation of Exchange Optics.