A design rule for structural remedies in digital markets: prefer obligations whose compliance can be checked as a yes/no observation over obligations that require an ongoing judgement about intent, quality or fairness.
Formulated most explicitly by Witte & Krämer on Google’s adtech stack
(sources/adtech/2026-08-21-witte-kraemer-google-adtech-divestiture-telpol2025.pdf,
Telecommunications Policy 2025). Their proposal is deliberately narrow: divest
DV360 entirely, keep Google Ads but strip its access to third-party display
inventory, ending its role as a general-purpose DSP. They reject full buy-side
divestiture as needlessly costly. The reason the narrow cut is preferred is
purely one of enforceability: “verifying that no requests are sent should be a
straightforward technical check” (p. 18).
Why the rule matters
The remedies it rules out are the ones regulators keep choosing:
- Non-discrimination / self-preferencing prohibitions require an unending counterfactual — what would a neutral ad server have done with this bid request. Gilbert concedes that separation does not remove the discrimination incentive, and notes the Google Shopping remedy “accomplished little to restore competition.”
- Functional separation fails on “wooden dollars”: internal transfer prices between separated units are not real prices, so the incentive survives the org chart (Witte & Krämer).
- Data siloing is unverifiable because ML pipelines “reuse partial data or inferred signals” — there is no observable event corresponding to the violation (Witte & Krämer).
- Log-level data access and universal transaction IDs improve detection but “would not eliminate the underlying incentive to discriminate” (p. 17).
- Ex-ante escalation regimes can be sound in design and still fail on clock speed: their DMA finding is that escalation “may lag behind a rapidly evolving ad tech sector” (p. 19).
The counterweight
Hovenkamp (Structural Antitrust Relief Against Digital Platforms, JLI 7(1),
2024,
sources/adtech/2026-08-21-hovenkamp-structural-antitrust-relief-digital-platforms-jli2024.pdf)
argues remedy effectiveness follows an inverted U — injunction, then
quasi-structural, then divestiture last — and that in the AT&T case “the latter
[interconnection obligations] were much more effective … than were the physical
divestitures” (p. 43), with a “strong presumption against breaking into a network
unless the decree also provides for effective interoperability” (p. 18). But
interoperability is exactly a non-binary obligation: terms and price must be set
(p. 43), and it “may require long-term or permanent oversight or even
re-assignment to a regulatory agency” (pp. 18–19).
⟨Inference: the two positions are compatible if read as a sequencing rule rather than a contradiction — cut where the cut is binary-checkable, and accept an ongoing-judgement obligation only where you are also willing to fund a permanent regulator for it. The adtech literature’s real finding is that regulators routinely choose the judgement-based obligation without funding the regulator.⟩
Transfer to the AI stack
The candidate binary prohibitions are connection prohibitions: an orchestration or agent layer must not route to its own affiliated model; a model provider must not operate an agent marketplace that ranks its own model. Both are, in principle, observable as network events.
The strongest objection: in adtech the checkable object is a discrete bid request on the wire. In the AI stack the boundary between “own client” and “third-party model access” is not a discrete network event — a provider serving its own model through its own inference path produces no request to observe. ⟨Inference: the rule transfers only where the affiliated relationship crosses an API boundary someone else can see. Where it does not, the AI stack has no binary-verifiable remedy available and falls back to the judgement-based class Witte & Krämer reject.⟩
Correction (2026-08-22): the remedy already exists, at the compute layer
The inference above was wrong about availability, and wrong about the layer. A
binary-verifiable remedy of exactly this class is already law in France, and it
sits at compute, not at orchestration. Per the Autorité’s own account of the
SREN Law (Law 2024-449 of 21 May 2024,
sources/ai-market/2026-08-22-autorite-concurrence-generative-ai-opinion-2024.pdf,
¶ 252):
- Art. 26 — cloud credits are “limited to one year and cannot include an exclusivity clause”; conditioning the sale of a product on the simultaneous conclusion of a cloud contract is prohibited; and the Autorité is statutorily obliged to report on the practice of “self-preferencing.”
- Art. 27 — egress fees above cost are prohibited when a customer moves data to its own or another provider’s infrastructure.
- Art. 28 — interoperability and portability requirements.
All three are yes/no checks against a contract or a price list. Sunset clause: 12 January 2027, for compatibility with the Data Act.
⟨Inference: the correct statement is neither “no analogue” nor “only at the routing layer.” The remedy type transfers, has already been legislated, and the measurement gap is therefore localisable: a statutory self-preferencing reporting duty exists at compute and does not exist at the routing layer — which is the one layer where self-preferencing would be structurally built in (the integrated first-party router). See AI stack market structure, row 11.⟩
Related
- Adtech as a regulatory precedent for the AI stack
- Verifiable carbon accounting — same structure, different domain: cryptography proves correct application of a rule to submitted inputs, never the completeness of the input set.