The DMA’s Self-Preferencing Rules and AI Answer Engines

When a search engine answers the question itself instead of listing pages that answer it, the site that would have ranked first loses the click. Whether that is prohibited self-preferencing under the Digital Markets Act turns on a question the Regulation was not drafted with in mind: whether a generated answer is a ranking of the gatekeeper’s own service at all.

The three limbs of Article 6(5)

Article 6(5) of Regulation (EU) 2022/1925 prohibits a gatekeeper from treating more favourably, in ranking and the related indexing and crawling, services and products offered by the gatekeeper itself than similar services or products of a third party, and requires it to apply transparent, fair and non-discriminatory conditions to such ranking. The text is at EUR-Lex. Article 2 defines ranking broadly enough to include the relative prominence given to results, and to cover results presented in any format, including voice — a detail that matters here, because it shows the drafters were already thinking beyond the ten blue links.

Three limbs, then, and they are not interchangeable. Ranking is about ordering and prominence at query time. Indexing is about what is in the corpus at all. Crawling is about what is fetched to build the corpus. Most commentary on AI answers argues the first limb only, which is the weakest of the three for reasons the next section sets out.

Is a generated answer a ranking decision?

The self-preferencing argument runs: the generated answer is a “service or product offered by the gatekeeper itself”; it is placed above third-party results; the third-party results are “similar services or products” because they answer the same query; therefore more favourable treatment in ranking. Each of those three steps has a serious counterargument.

• Is the answer a separate service? The gatekeeper will say it is a feature of the search engine, not a distinct service competing with the results — the same argument made about spelling correction and knowledge panels. Article 6(5) was aimed at vertical services that plainly exist as businesses, such as shopping or travel comparison. A generated summary is less obviously one.
• Are the results similar products? A publisher’s article and a synthesised paragraph are substitutes for the user and not obviously similar products in the sense of the provision. The further the generated answer is from a commercial vertical, the weaker the similarity limb gets.
• What is the counterfactual? Self-preferencing analysis needs a comparison. If the gatekeeper does not offer any third party the ability to place a generated answer in that position, it is not treating its own service more favourably than a comparable third-party one; it is occupying a position no third party can occupy. That is arguably a different problem, closer to Article 6(12)’s fair-access duty or to ordinary competition law under Article 102 TFEU than to Article 6(5).

None of that means the conduct is lawful. It means the ranking limb is contested, and a page that asserts the answer confidently is doing the reader a disservice. What would settle it is a Commission decision on the point, or a preliminary reference reaching the Court of Justice, or an amendment following the Article 53 review. None of those had produced an answer on generated answers specifically at the time of writing.

The other half: the data the answer is built from

The stronger arguments are often about the second and third limbs and about a separate provision entirely. A generated answer is built from crawled content. If a publisher cannot decline to have its content used to generate answers without also declining to be indexed for ordinary search, then the crawling and indexing limbs are engaged directly: the gatekeeper is conditioning presence in the index on contributing to the gatekeeper’s own generated product. That is a cleaner Article 6(5) argument than prominence, because the coupling is concrete rather than a question of pixel real estate.

Article 6(11) is the other provision worth knowing here. It requires a gatekeeper providing an online search engine to give third-party search engines, on their request, access to ranking, query, click and view data on fair, reasonable and non-discriminatory terms, with personal data in query, click and view data anonymised. That is an input-parity duty rather than an output-prominence duty, and it is the provision a competing answer engine would invoke to argue that it cannot build a comparable product without the same signals.

Copyright runs alongside all of this and is not the same question. The text and data mining opt-out under Article 4 of the CDSM Directive governs whether content may be used for training at all, and it is covered in its own page. A publisher usually has both arguments available and they are decided by different authorities.

Enforcement posture, not prediction

The Commission has open Article 6(5) work in the search context. It opened non-compliance proceedings against Alphabet in March 2024 concerning self-preferencing in Google Search, and issued preliminary findings in 2025 that Google treats its own vertical services more favourably than rivals’. Preliminary findings are exactly that: a statement of the Commission’s provisional view, which the addressee replies to, and which may or may not become a non-compliance decision. They decide nothing.

It is worth being explicit about what that proceeding does not determine, because this is the paragraph most coverage omits. It does not determine whether a generated answer is a service for the purposes of Article 6(5). It does not determine whether placing one above organic results is more favourable treatment. It does not create any obligation about AI answers for other gatekeepers. And even a non-compliance decision would be appealable to the General Court, so it would not be a settled statement of the law on the day it issued. The current status of every case is on the Commission’s DMA case register, which is the only source worth checking for posture.

What a publisher can actually do

The DMA has no private enforcement mechanism of its own: the Commission is the enforcer, and Article 27 lets any third party, including business users and competitors, inform the Commission of practices falling within the Regulation. That is a lower bar than a formal complaint and it is how most evidence arrives.

The evidence that helps is comparative and quantitative: query samples showing the interface, the position and area given to the generated answer against organic results, and traffic data before and after a feature was rolled out in a given Member State. What helps least is a general assertion that AI answers reduce clicks. Separately, and in parallel, keep the crawling question clean: document whether the gatekeeper offers a control that separates training or answer-generation use from ordinary indexing, and what happened when you used it. That record is the raw material for the indexing-and-crawling argument, which is the one least likely to be answered by an argument about what counts as a service.