The programmatic online-advertising market is the only fully studied case of a multi-layered, vertically integrated intermediation stack in which one firm sets the allocation rules, appears on both sides of the market, and treats price formation itself as a design parameter. That makes it the closest available precedent for regulating the AI stack — not because AI changes ad markets, but because the structural problem is the same one.
Primary analysis product: dossiers/2026-08-21-adtech-als-regulatorischer-praezedenzfall.md
(German, 18 papers, six mechanism families). Sources in
sources/adtech/ — cited per claim below.
Requested by Martin Pompéry to inform policymakers on AI market regulation.
The actor taxonomy
Converged across CMA (2020, Appendix M), ACCC (2021, ch. 2) and Bundeskartellamt (2023): advertiser → media agency / trading desk → DSP → ad exchange / SSP → ad network → publisher ad server → publisher. Cross-cutting and outside the chain: DMP/CDP, identity providers, verification and brand-safety vendors.
The regulatorily decisive property is not the chain’s length but that one actor can occupy several non-adjacent positions at once — Alphabet is ad server, exchange, DSP and, through YouTube and Search, publisher. ⟨Inference: this is the structural fact the entire remedy debate rests on.⟩
The five transferable findings
- The rule-setting layer is the power layer, not the deepest layer. In adtech effective power sits with the ad server that sets auction rules and has “last look” — not with the largest data holder. The AI-stack analogue is the orchestration/routing layer, not the model and not compute. Regulation keyed to model size or training FLOPs addresses the wrong layer. ⟨Inference from Witte & Krämer 2025; Despotakis et al. 2021, Prop. 4; Latham et al. 2021.⟩ Corrected 2026-08-22: the identification of the layer holds, the conclusion does not. In adtech the rule-setting layer was also the stickiest, which is what let it hold a rent; in the AI stack stickiness (compute) and rule-setting (routing) sit on different layers, so the adtech pathology has no single home. See AI stack market structure, row 4.
- Opacity is a tunable parameter whose yield comes from the counterparty’s learning lag. Soft floors are revenue-neutral in equilibrium and profitable only in transition, because the bidders in the field experiment “do not reduce their bids when soft floors are introduced” (Zeithammer 2019, p. 3). A platform that changes mechanisms often enough harvests the transition rent permanently without ever running a mechanism that is harmful in equilibrium. The regulable quantity is therefore the change frequency behind a stable interface, not the disclosure of the logic. Sharpened 2026-08-22: literally documented in the AI stack (auto-versioning behind a stable model name) — but inverted, and that weakens the design-lever reading: adtech buyers knew what they bought and not what they paid; AI buyers know what they pay and not what they got. Vendor- documented opacity is disclosure, not concealment. See AI stack market structure, row 5.
- Cryptographic auditability stops at completeness. VEX and Addax prove that the rule was correctly applied to the submitted bids; neither can prove the committed bid set is the complete true set. Same limit as the “single-electricity-meter” side-step in Emission Impossible — see verifiable carbon accounting. ⟨Inference: any “auditable AI” mandate needs institutional completeness assurance; cryptography does not supply it. The SEC Consolidated Audit Trail is the only real-world pattern (Srinivasan 2020, Part IV).⟩
- Remedies must be binary-verifiable to be enforceable — see binary-verifiable remedies.
- Demand-side concentration can be countervailing power, not harm. +245.10 HHI points of agency concentration → −11.32 % platform revenue, via lower click prices (Decarolis & Rovigatti 2020, pp. 3–4). ⟨Inference: enterprise buying consortia and model brokers are candidate countervailing instruments in the AI stack — with the authors’ own warning that the platform responds via higher differentiated reserve prices, which hurt the smallest buyers.⟩
Sixth, on measurement: 2022) generate foreclosure with no data advantage — platforms are modelled as already omniscient (p. 12) — which makes the quantity to measure the user-level singlehoming rate at the distribution interface, not aggregate share or model count.
Where the analogy breaks
- Unit of trade. Adtech allocates a transaction (trillions/year, microseconds, fractions of a cent); the AI stack mostly allocates a capability under subscription. The auction results transfer only where allocation is genuinely auction-like: ad-funded AI surfaces, compute spot markets, agent marketplaces.
- Regulatory starting state. Adtech ran 15 years with no ex-ante regime; the AI stack starts with the AI Act, DMA and §19a GWB. “Conduct remedies are unenforceable” is a lesson from a regime that no longer applies — though Witte & Krämer’s own DMA finding (escalation too slow, p. 19) supports it, while Campbell, Goldfarb & Tucker (2014) cut the other way: an ex-ante obligation with fixed per-firm compliance cost concentrates the market toward the generalist even when large firms have no inherent advantage. The AI Act is structurally the same mechanism as the consent regime they studied.
- The binding constraint may be elsewhere. If 2022) are right that data has “the diminishing returns of capital, in the long run” (p. 30), the moat self-limits and the binding AI factor is compute — capital and energy, not information — which calls for procurement, energy and export-control instruments instead. Acemoglu et al. (2022) argue the opposite (data externalities are submodular, prices too low, and competition between platforms “may further reduce welfare”). Neither paper cites the other. ⟨Inference: the two most sophisticated economic claims about data as a production factor have never been tested against each other — anyone invoking “data power” is invoking an open question.⟩
Provenance is part of the finding
The technically most precise mechanism literature is written by employees of the firm under study (the autobidding survey has 25 of 25 authors at Google); the sharpest structural critiques come from consulting economists for complainants (CRA/News Corp) or INET-funded work; Prat & Valletti omit a co-author’s competition-authority role and Hovenkamp carries no conflict statement in either paper (interoperability remedies 2023; structural relief JLI 2024). Clean and explicit: Gilbert, Witte & Krämer (DFG, explicit independence statement), Bichler et al. (DFG), Zeithammer, Farboodi & Veldkamp. Acemoglu et al. disclose Google and Microsoft funding for a result that runs against the funders’ interest — which strengthens it.
⟨Inference: the structural point is worse than any single conflict — there is almost no disinterested empirics in this field because the data is not public. All three empirical papers depend on a commercial data provider or one firm’s cooperation (Redbooks + SEMrush; Xandr; a single publisher’s data). The same constellation is forming around AI evaluation and usage data; a research data access right (cf. DSA Art. 40) is the precondition for the AI debate not having adtech’s provenance structure in ten years.⟩
Open
Is AI allocation auction-like enough for families B and C to apply?Answered 2026-08-22, negatively. The AI stack has no auction anywhere: clearing runs on posted per-token list prices, subscriptions, multi-year take-or-pay capacity contracts, and rate limits (an allocation quota, not a price). Families B and C therefore have no target — bid shading, reserve prices, last look, header bidding and first-vs-second price transfer to nothing. The one forming exception is at compute, not at any ad-server analogue: CME/Silicon Data compute futures from 2026-10-05, “pending regulatory review” — a date now four weeks past, with no source here on whether the contract launched; see AI stack market structure, row 2 ⟨flagged 2026-11-02⟩. ⟨Inference: this is a momentary reading with a visible expiry date — posted prices are what scarce capacity looks like, and secondary markets appear under overcapacity.⟩- Acemoglu et al. vs. Farboodi & Veldkamp: unresolved, and transfer 5/6 hangs on it.
- Whether the agency taxonomy survives post-cookie, Privacy Sandbox and retail media. If not, the terminology imports a vocabulary for a dead market.
Related
- AI stack market structure — the 2026-08-22 comparison that tests these five transfers against the AI stack property by property, and corrects transfers 1, 2 and 4.
- Binary-verifiable remedies — transfer 4 in full, including the correction that such a remedy already exists in law, at the compute layer.