A digital marketplace can report hundreds of participants, intense transaction volume and stable market shares without knowing how independently its participants form decisions. A bidder count measures presence. An auction outcome records a transaction. Neither, by itself, establishes how many materially different economic judgments contributed to the result.
Decision diversity matters because auction outcomes determine who receives inventory, what advertisers pay, what publishers earn, and which intermediaries collect fees. The structure of independent bidding judgments is therefore an economic variable, not a speculative concern. The direction and size of any change in a specific market’s prices or revenue still depend on auction rules, buyer valuations, and observed outcomes.
What Effective Decision Diversity is
Effective Decision Diversity (EDD) asks how many materially different decision processes a body of marketplace evidence can support behind a set of nominal participants, within a defined market, opportunity population and observation regime. A result is a bounded estimate with its uncertainty, assumptions and competing explanations stated—or an explicit “unresolved” when the records cannot support one. EDD complements bidder counts, market shares, spend and transaction reporting; it does not replace them.
What EDD does and does not claim
A measured behavioral pattern, a causal theory of shared control, an economic impact estimate and a legal conclusion are four separate propositions. Passing one does not authorize the others. Specifically:
- EDD is not a measure of welfare, unlawful coordination or financial damage; those are distinct economic and legal questions.
- A change in EDD alone does not identify a counterfactual price, a net publisher loss, advertiser savings or exchange revenue impact, and it does not identify the vendor, owner or intent behind the behavior.
- Exchange Optics does not claim to have demonstrated programmed collusion in live exchanges, a universal ROI, certified client savings or a validated pricing premium.
- Independence here is a governance practice, not a claim of accreditation.
The missing denominator can reverse the interpretation
An exchange does not necessarily invite every bidder to every opportunity. Consent status, geography, deal eligibility, buyer seat permissions, floors, inventory quality, throttling, timeout policies, pretargeting and routing may determine who is allowed to bid and which bids become visible. OpenRTB explicitly allows deal-specific terms, buyer-seat restrictions, no-bid responses and different notice events. [S01]
If an analyst sees only winners, an apparent collapse in the diversity of winners may reflect a change in selection—not a change in the underlying bidding policies. Even a complete accepted-bid ledger can miss bidders that were not asked, excluded, timed out or declined. An observer can also mistake shared reactions to a supply-quality shock for shared decision control.
A serious investigation therefore needs to specify its population, opportunity denominator, observation level and alternative explanations before interpreting a diversity signal. Useful source categories include eligible impression opportunities; bidder invitations; timeouts and no-bids; bid values and auction conditions; reserves and deal rules; quality, currency and fee transformations; decision and change timestamps; and reconciled delivery, cost and publisher receipts. Bidder and vendor metadata can improve explanation when lawfully available, but common vendor identity is neither necessary nor sufficient to establish common decision control.
If the data lack those records, a bounded conclusion may still be possible under narrower assumptions. Sometimes the defensible result is unresolved. Treating an unresolved result as a failure misunderstands what independent assessment is for.
Why auction design changes the economic meaning
Programmatic advertising is a natural place to begin because transactions are algorithmic, repeated and dependent on bidding and routing infrastructure. The IAB Tech Lab’s OpenRTB specifications describe exchanges sending requests, receiving seat-level bids or no-bids, and issuing win and billing notices. They also permit exchange-specific policies and different auction terms. [S01]
These details are economically consequential. Under a first-price auction, a winning buyer’s own effective bid is central to the clearing price; under a second-price-type design, another bidder’s offer may influence the price paid. Floors, deal priority, quality adjustments, fees and routing alter what counts as a valid competitive opportunity. Google’s documented 2019 transition of Ad Exchange inventory to a unified first-price auction illustrates that the market rules themselves are not fixed. [S01, S02]
Suppose many bidders begin bidding similarly. One explanation is a genuine narrowing of independently generated willingness-to-pay estimates. Another is that competing systems make efficient, independent use of the same conversion signal. A third is that the exchange routed a different mix of opportunities to the observed bidders. Those cases can produce related bid patterns yet call for very different responses.
Auction intuition does not run in one direction. The clearing price does not rise or fall simply with how similar bids are: the upper tail and distribution of bids determine who wins, and the price effect depends on the pricing rule, reserves, quality and the underlying valuations. More behavioral variety is not automatically better if it reflects noisy valuation, inconsistent objectives or poor information. Less variety is not automatically worse if better common information brings bids closer to true economic value. Any claim of impaired price discovery therefore requires a clear definition of the relevant information and a credible comparison with an alternative market state.
There is direct empirical evidence that auction-rule choices affect advertising revenue. Ostrovsky and Schwarz studied reserve prices through a large field experiment in sponsored-search advertising and found substantial revenue gains from changed reserves. A separate field experiment by Choi and Mela, published in Marketing Science in 2026, found that setting reserve prices raised a major publisher’s revenue by 35% relative to a zero reserve in that tested display environment. Neither study measures EDD or establishes the financial effect of shared bidding policies. Their relevance is narrower, but important: economic conclusions must be tied to specific market mechanisms and interventions, not inferred from participant counts. [S03, S04]
The advertiser, publisher and exchange do not share one payoff
For an advertiser, the relevant question is not simply whether the average clearing price rises. Price per qualified impression, reach, conversion probability, return on advertising spend, pacing, budget displacement and availability of alternative inventory all matter.
More intense bidding can raise media cost while producing better allocation to an advertiser that values the impression more. Conversely, convergent underbidding could lower observed prices but cause missed delivery or lost valuable reach. A lower CPM is not necessarily a better outcome.
For a publisher, the question is often net yield and the quality of demand. Reduced independent participation could weaken price pressure on an impression, particularly where bidder valuations are otherwise heterogeneous, yet the realized effect depends on reserve prices, fill, floors, deal constraints, fees and traffic composition. A publisher can also benefit from more precise and reliable automated bidding even where decisions grow more similar. A reported change in gross clearing prices cannot be silently converted into a change in publisher revenue.
For an exchange, revenue may depend on transacted volume, fee structure, deal mix, retained demand and partner economics. An intervention that improves publisher yield might affect fee revenue differently depending on pricing and take rates; an intervention that increases near-term transaction volume may create longer-run dependence on concentrated demand. Commercial risk can therefore include disputed auction fairness, erroneous vendor deactivation, revenue volatility, weakened partner confidence and costly investment in the wrong controls—not just one theoretical price direction.
The Association of National Advertisers’ Q1 2026 programmatic transparency benchmark shows material differences across advertisers in the share of spend converted into qualified impressions, and the UK’s Competition and Markets Authority documented serious questions about digital advertising intermediation and fees. Those findings justify looking closely at media quality, fees and supply-chain accountability. They do not measure EDD. [S05, S06]
Algorithms change the coordination problem—not the legal standard
The warning is credible enough to study seriously. In a widely cited 2020 American Economic Review simulation, Calvano and coauthors found that independently learning Q-pricing algorithms could sustain supracompetitive outcomes without direct communication in a specified repeated-price-competition model. That was a controlled model, not a field finding about advertising bidders. [S07]
Other scholarship reinforces the need to separate mechanisms. Brown and MacKay model how automated responses and differing pricing frequencies can change competitive prices; Assad and coauthors report observational and instrumental-variable evidence relating algorithmic adoption to margins in Germany’s gasoline market. Neither study establishes that a reduction in EDD caused a loss in an advertising auction. And contrary evidence matters: a 2024 e-commerce study by Hanspach, Sapi and Wieting finds no general increase in Buy Box prices associated with algorithm use across the sampled products, including instances of lower prices. Automated decision-making is not a one-direction economic treatment. [S08, S09, S10]
The mechanisms also differ legally and causally:
- Explicit algorithm-assisted collusion: firms coordinate through or around software. Evidence concerns agreement, communications, instructions and conduct, not just bid resemblance.
- Common-service or hub-mediated coordination: separate firms may use a central provider in ways that require scrutiny of information flows, decision rights, instructions and the actual operation of the service.
- Autonomous strategic coordination: algorithms may learn or adapt to competitors in a repeated setting without an explicit agreement; the economic phenomenon does not itself settle the applicable legal classification.
- Common-input dependence: systems respond to the same data, forecasts, cost shocks or update schedules; dependence may be real without coordination.
- Independent parallel behavior: distinct firms independently reach similar bids because the commercial facts and incentives are similar.
The OECD’s 2025 review describes both the efficiency potential of pricing algorithms and competition concerns across G7 jurisdictions. The U.S. DOJ and FTC’s 2024 statement of interest in a hotel algorithmic-pricing case states an agency legal position about potential algorithm-facilitated price fixing. Courts have not simply adopted it: in August 2025 the Ninth Circuit affirmed dismissal of that case, Gibson v. Cendyn, holding that competing hotels’ independent adoption of the same pricing software did not, without more, establish a Sherman Act violation, and distinguishing shared use of a tool from an agreement among competitors or the pooling of their confidential data. Neither the agency position nor the court’s holding concerns ad exchanges or any Exchange Optics prospect, but together they show why shared tooling and shared control must be kept apart. [S11, S12, S14]
For deeper treatment of that distinction, see Research 05: Shared infrastructure does not mean shared control.
Concentration can be operational as well as strategic
Economic exposure need not involve collusion at all. If different firms depend on the same model release, data service or cloud provider, they may encounter common errors or experience synchronized responses to a shock. A widespread model error could simultaneously distort pacing, values or eligibility decisions. A shared infrastructure failure could constrain bidding or impair allocation. These are distinct from conduct that restrains competition; in some cases, common infrastructure can increase reliability and lower costs.
The Financial Stability Board has identified third-party concentration, correlated behavior and model risk as possible financial-sector AI vulnerabilities. Its assessment is evidence of a broadly recognized dependency mechanism—not evidence that advertising exchanges currently experience comparable failures or losses. The correct next question is operational: which dependencies are material, what alternative pathways exist, how do errors propagate and what is the consequence if one component changes? [S13]
Decision diversity and system resilience overlap but are not identical. A high diversity measure cannot certify uptime or provider independence; a low measure does not prove a single point of failure. Provenance, update timing, operational exposure and incident records may be needed in addition to bidding behavior.
From a signal to a financial claim
A business should demand three increasingly stringent answers.
First: Is the observation credible? Define the unit and scope; check timestamps and coverage; distinguish participants, seats and legal entities; test missingness, selection, bidder identity changes and dependence among repeated observations. Compare like-for-like opportunities and preserve uncertainty rather than turning highly repeated events into fictitious independent samples.
Second: Does the observed pattern move with an economically relevant outcome? With adequate records, estimate conditional associations with net clearing prices, advertiser quality-adjusted costs, publisher yield, fill, allocation, fees, latency or reliability. Segment by auction type, inventory, demand, floor, deal and period. A correlation is a useful question, not yet an estimate of money lost.
Third: What would have happened otherwise? To assign causal economic impact, identify an intervention or credible natural experiment: for example, a prospectively planned randomized rollout, appropriately controlled update sequencing or a defensible quasi- experiment. Verify comparable opportunity populations and changes in demand, consent, fees, reserves, optimization and selection. Consider spillovers between treatment and control; advertisers can reallocate budgets and competing bidders can adapt. Show uncertainty, sensitivity to assumptions, effects by market side and the limitations of extrapolation.
The resulting estimate could be positive, negative, negligible or not identifiable. For executives, this is the difference between a statistic to display and a result on which capital, partner relationships or auction controls can responsibly depend.
Why independent verification is a commercial purchase
Exchanges already have essential analytics. Advertisers have supply-path optimization and campaign-quality tools. Publishers have yield diagnostics. Industry standards and independent verification services address fraud, measurement, transaction transparency and authorized selling. Those disciplines should be strengthened, not dismissed. They generally do not answer the identical question of how much independent decision-making a given observer’s marketplace evidence can support.
An operator may nevertheless have to choose among consequential actions: change routing or floors, adjust a shared service, make a claim about auction competitiveness, commission a larger integration, investigate partner concentration, or decide not to intervene. False reassurance can defer a necessary investigation. A false alarm can trigger unnecessary engineering changes, damage partner relationships or move spend without improving outcomes. Both have costs even when no financial effect of EDD itself is established.
A separate assessor offers value only if it actually challenges customer records and interpretations, manages conflicts, publishes limitations and can return a negative or unresolved conclusion without pressure to manufacture a positive one. The relevant commercial proposition is therefore decision quality under uncertainty: a rational purchaser compares the cost and decision value of an assessment with the potential cost of acting on incomplete evidence. That expected value varies by organization and decision. For how independence is protected when the customer pays, see Research 03: Independent assurance in AI marketplaces.
An evidence-gated Founding Design Partner sequence
Stage 1 — Source-readiness and claim assessment (initial, bounded engagement). The operator receives:
- a map of available data and legal permissions against the questions the operator wants answered;
- reconciliation of provenance and denominators, where the scoped access allows;
- a register of missing observations and of independence and selection risks;
- a test of whether any existing claim is supportable; and
- a written disposition: proceed, repair the source, or unresolved/no-go, with limitations stated.
A valuable outcome may be learning that a proposed measurement would be misleading or premature. Stage 1 is not live-market EDD certification.
Stage 2 — Qualified field baseline. Only if data access, observability, validation and independent review requirements are met, assess bounded behavioral decision diversity, uncertainty, competing explanations and appropriate negative findings. Customer records do not automatically become independently verified simply because a third party processes them.
Stage 3 — Economic evaluation and change verification. Where outcome feeds, rights and a defensible comparison exist, examine associated market outcomes and, if identification supports it, effects of a defined intervention. Subsequent re-measurement evaluates whether conditions changed; it does not promise that a particular operational change improves welfare or revenue.
Ongoing assessment, if warranted by validated operational capabilities, is a later engagement rather than a certification implied by the initial review. Real marketplace data access, independent field qualification and causal economic demonstration remain separate requirements.
The business question worth asking
The issue is not whether algorithms are dangerous or whether every similar bidder is secretly controlled by one system. It is whether the marketplace can distinguish its sources of apparent competition from the independent decisions that actually shape outcomes—and whether it has enough evidence to act.
A bidder count is a starting point. An independently defensible explanation of the decisions behind it can be a better basis for allocating engineering effort, managing dependencies, evaluating auction changes and making promises to partners. An independent assessment should identify supported concerns, credible negative findings, and gaps in available evidence. Market integrity depends on defensible answers—not reassuring participant counts or predetermined verdicts.
Seeking Founding Design Partners. Exchange Optics is working with a small number of exchanges, publishers and advertisers that need to know what their marketplace evidence can and cannot support before they act. The starting point is a bounded Stage 1 conversation.
Source notes
Identifiers below map to the dated evidence classification and verification notes in SOURCE_REGISTER_2026-10-26.md . These sources support general economic mechanisms, industry context and case-specific results, not a demonstration of Exchange Optics field performance or economic benefit.
[S01] IAB Tech Lab, OpenRTB 2.6 specification (https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md).
[S02] Google Authorized Buyers, first-price auction transition announcement (https://support.google.com/authorizedbuyers/answer/9612058?hl=en), October 2019.
[S03] Ostrovsky & Schwarz (2023), Reserve Prices in Internet Advertising Auctions: A Field Experiment (https://www.journals.uchicago.edu/doi/10.1086/725702), Journal of Political Economy.
[S04] Choi & Mela (2026), Optimizing Reserve Prices in Display Advertising Auctions (https://pubsonline.informs.org/doi/abs/10.1287/mksc.2023.0002), Marketing Science.
[S05] Association of National Advertisers, Q1 2026 Programmatic Transparency Benchmark (https://www.ana.net/content/show/id/pr-2026-05-programmatic).
[S06] UK Competition and Markets Authority, Online Platforms and Digital Advertising Market Study (https://www.gov.uk/cma-cases/online-platforms-and-digital-advertising-market-study), 2020.
[S07] Calvano et al. (2020), Artificial Intelligence, Algorithmic Pricing, and Collusion (https://www.aeaweb.org/articles?id=10.1257/aer.20190623), American Economic Review.
[S08] Brown & MacKay (2023), Competition in Pricing Algorithms (https://www.aeaweb.org/articles?id=10.1257/mic.20210158), American Economic Journal: Microeconomics.
[S09] Assad et al. (2024), Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market (https://www.journals.uchicago.edu/doi/10.1086/726906), Journal of Political Economy.
[S10] Hanspach, Sapi & Wieting (2024), Algorithms in the Marketplace: An Empirical Analysis of Automated Pricing in E-Commerce (https://doi.org/10.1016/j.infoecopol.2024.101111), Information Economics and Policy.
[S11] OECD (2025), Algorithmic Pricing and Competition in G7 Jurisdictions (https://www.oecd.org/en/publications/algorithmic-pricing-and-competition-in-g7-jurisdictions_f36dacf8-en.html).
[S12] U.S. Department of Justice (2024), DOJ/FTC statement of interest in hotel algorithmic-pricing litigation (https://www.justice.gov/archives/opa/pr/justice-department-and-federal-trade-commission-file-statement-interest-hotel-room).
[S13] Financial Stability Board (2024), The Financial Stability Implications of Artificial Intelligence (https://www.fsb.org/2024/11/fsb-assesses-the-financial-stability-implications-of-artificial-intelligence/).
[S14] Gibson v. Cendyn Group, LLC, 148 F.4th 1069 (9th Cir. Aug. 15, 2025).
This analysis explains the economic and evidentiary questions without disclosing proprietary Exchange Optics estimator parameters, code, thresholds, private marketplace records or unreleased tests. Legal classifications and causal economic attribution depend on separate case-specific evidence.