Chatbots Are Search. Google's Logs Stay Closed

The designation is real. On 31 August the Commission named ChatGPT a Very Large Online Search Engine under the Digital Services Act, and named Reddit and Roblox Very Large Online Platforms, after each service declared at least 45 million average monthly users in the EU. ChatGPT is the first generative assistant in that search-engine tier: a hybrid that answers prompts, including by searching the web. Four months from notification — January 2027 — it must run the VLOSE pack of systemic-risk assessment, mitigation, and independent audit, with Ireland’s Coimisiún na Meán sharing supervision. That is safety law applied to retrieval at scale. It is not a grant of Google’s click logs.

The Digital Markets Act track is older and narrower. Article 6(11) already required Google, as a designated gatekeeper, to make search data available on fair, reasonable, and non-discriminatory terms. Google opened a licensing program in March 2024. On 16 July 2026 the Commission adopted binding specification measures because, in its own words, that offer had been ineffective. AI chatbots offering search functionalities are eligible. Subject to anonymisation, Google is to share the same ranking, query, click, and view data it uses to tune Search in the European Economic Area, starting in January 2027. Android AI interoperability is a separate decision, with user-facing changes from July 2027.

A third file, dated 23 July, fined Google €460 million for self-preferencing shopping, hotels, transport, and sports results, plus €430 million for Play steering, with sixty days to change presentation. Advertisers, on a parallel 2026 calendar, are living through Google’s shift from Dynamic Search Ads toward AI Max. Adjacent clocks. Not one launch.

The residual is blunt. If data sharing is meant to cut Google’s infrastructure advantage over assistants, why do the results stop at eligibility and deadlines?

Because the bottleneck is not the label. It is the pipe.

Formal eligibility is not a data feed. January 2027 is the first date the Commission treats as operational, not a scoreboard of who already trained on Google’s logs.

DMA sharing improves a rival only if the data is granular, timely, and usable through a process that actually completes. Google’s European Search Dataset Licensing Program now lists the gates: EEA-directed search engines, 50,000 average monthly EU users, two years of EU search or more than €50 million in capital, plus sanctions and non-EEA-state-control tests. Licensing agreements begin going out on 17 September 2026. Samples follow on 16 November. Full access still requires a Level 1 reasonable-assurance audit of segregation, access control, and logging; Level 2 reporting keeps the license alive. The Commission says Google must publish a list of recipients. There is no Commission or beneficiary notice that an eligible chatbot has received the specified Search Dataset and put it into production.

The DSA mechanism is equally specific, and equally distinct. A service becomes a VLOSE when it performs search-like retrieval and clears the user threshold. Designation then triggers transparency and risk duties. It does not entitle OpenAI, or anyone else, to Google’s query stream. Treating ChatGPT’s new status as competitive access to Search infrastructure confuses a duty with an input.

Separate remedies also explain the staggered dates. Self-preferencing is a ranking and layout problem on a page Google already operates. Data sharing requires anonymisation, FRAND pricing, and recipient-side audit architecture that the specification proceedings had to write down after the first program failed. Android invocation is an operating-system change with a later user-facing date. Implementation dependencies are not evidence that the competitive transfer has occurred.

What still gets mis-priced is the substitution story. Markets treat compute as the scarce AI input — the assumption China’s efficiency programs attack by squeezing more capability out of less silicon. Query, click, and view logs are a different scarce input: they tell a retrieval system what the web did when a human asked. Regulatory access was supposed to be the cheap substitute for having operated Google Search for twenty years. Until someone takes delivery, that substitute is an option with a strike date, not an asset. The same gap shows up in the utility test for AI apps: assistants need retrieval quality users will pay for. Generative systems are not a replay of the last platform boom in part because their data requirements are this specific.

The test is public. Watch Google’s recipient page and any beneficiary’s first operational-use notice after January 2027. If an eligible chatbot receives the specified dataset and independent performance evidence shows a material cut in the information disadvantage, the sharing theory holds. Until then, Europe has redesigned the paperwork around search. It has not recast the economics.