For the last two years, the most valuable piece of real estate in consumer AI hasn't been a model. It's been the wake word.
On July 16, 2026, the European Commission moved to take that real estate away from Google. In two binding specification decisions under the Digital Markets Act, Brussels ordered the company to give competing AI assistants the same access to Android that it has reserved for Gemini — and, separately, to share anonymised Google Search data with rival search providers and AI chatbots that offer search. It is the first time any regulator has treated "being the assistant your phone actually listens to" as a market position that can be opened by law.
Google says the decisions threaten privacy and security. Its competitors say this is the only reason they will ever get a fair shot. Both, in their way, are describing the same shift: the assistant layer of the phone is being unbundled from the company that makes the operating system.
What Google Actually Has to Do
The Android decision rests on Article 6(7) of the DMA — the interoperability provision that requires designated gatekeepers to give third parties access to the same hardware and software features their own services use. Until now, that article has mostly been argued over in the context of smartwatches and notifications. The Commission has now pointed it squarely at AI assistants.
Analysts reading the decision have grouped the required access into four layers, and it is worth understanding them separately, because each one removes a different structural advantage:
| Layer | What Gemini has today | What rivals get |
|---|---|---|
| Invocation | System-wide wake word and long-press launch — the assistant that answers when you don't open an app. | Voice activation and system entry points for the assistant a user actually chooses. |
| Context | Screen contents, sensor data, and on-device app data that let it answer "what am I looking at?" | The same contextual signals, under privacy safeguards written into the decision. |
| Actions | Screen automation and cross-app control — booking the taxi, not just describing how. | The ability to act inside other apps on the user's behalf. |
| Resources | On-device models and privileged background execution. | Access to the same on-device resources and background capability. |
The Commission's framing of the problem is blunt: as things stand, roughly 60% of EU Android users cannot meaningfully use an alternative AI assistant, because the assistant they might prefer is locked out of the functions that make an assistant useful. You can install it. You just can't summon it with your voice, and it can't see or touch anything.
Reporting on the decision indicates Google must open 11 Android AI features to rival assistants by August 2027, with the technical work phasing in before then. That is a long runway by consumer-tech standards and a short one by platform-engineering standards, which is roughly where the fight over implementation will happen.
The Second Decision: Search Data
The companion decision is arguably the more consequential one for anyone building a search-capable chatbot. Google must share the same anonymised data it uses to optimise its own search — query, click, ranking and view signals — with eligible third-party providers. Critically, the Commission defined AI chatbots that offer search functionality as eligible recipients, not just traditional search engines. That is a deliberate acknowledgment that the competitive set for Google Search in 2026 includes assistants, not just other crawlers.
The mechanics come with guardrails in both directions. The Commission worked with privacy experts on a multi-layered anonymisation method, and Google is permitted to assess cybersecurity and data-protection risk before handing data to a given recipient — a carve-out that will be watched closely for how liberally it gets used. On the commercial side, the decision requires a fair pricing formula and a transparent access process, with the search-data measures phasing in across the remainder of 2026 and Google expected to finalise pricing by January 2027.
Query and click data is the flywheel argument in miniature. Ranking quality comes from watching what billions of people click, which produces better rankings, which attracts more users, which produces more clicks. No amount of capital buys a competitor out of that loop. Forcing a share of the signal is the Commission's attempt to break the circularity — and it is the part of the package Google has fought hardest at every prior stage.
Google's Objection
Google pushed back publicly the same day. Its statement centred on three claims: that opening these interfaces weakens privacy and security safeguards for European users, that mandated data sharing risks exposing business secrets, and that some of the material touches national-security-sensitive ground.
The privacy argument deserves to be taken seriously rather than waved off, because the capabilities in question are genuinely invasive. An assistant with screen-reading and cross-app action rights can see your banking app and tap buttons inside it. Granting that to Gemini is a decision users make once, inside a trust relationship with the OS vendor. Granting it to an arbitrary third-party assistant is a materially different risk surface, and the quality of the consent flow and permission model Google builds will determine whether this ends well.
The counter-argument is equally straightforward: Google is not a neutral party in assessing that risk. It is the incumbent whose product currently holds the exclusive privileges at issue, and "this would be unsafe" is the standard defence of every platform asked to open an interface. The Commission's answer was to write the safeguards into the decision itself — device integrity, security, and privacy conditions attached to the access — rather than leave the question to Google's judgment.
If Google under-delivers, the Commission can open a separate non-compliance case, and the DMA's penalty ceiling is up to 10% of annual worldwide revenue. For a company of Alphabet's size, that is the rare fine large enough that compliance is cheaper than litigation.
Who Actually Benefits
The obvious winners are the assistant makers with real models and no distribution: OpenAI, Anthropic, Perplexity, Mistral, and the wave of European AI startups that have spent two years discovering that being one tap away from your user is not the same as being zero taps away.
But the more interesting beneficiaries may be the ones this article can't name yet. The reason there is no thriving market in specialised assistants — one tuned for accessibility, one for a specific profession, one that runs entirely on-device for privacy reasons — is that none of them could reach the invocation layer. Lower that barrier and the category stops being a three-horse race between companies large enough to negotiate with Google directly.
There's a real caveat worth stating plainly: this applies in the European Union. A US Android user gets none of it unless Google decides that maintaining two divergent assistant architectures is more expensive than shipping the open one everywhere. That calculation has historically favoured convergence — the GDPR cookie banner is a global artefact of a European law — but Google has also shown a willingness to ship EU-only variants when the alternative is giving up an advantage worldwide. Which way this one goes is the single biggest open question about the decision's real-world impact.
The Bigger Pattern
Read alongside the rest of July's regulatory docket — the Commerce Department lifting export controls on Anthropic's frontier models after safeguard commitments, newspapers moving to sanction OpenAI over discovery in the copyright litigation — a pattern is visible. Regulators have stopped asking whether to touch AI and started picking specific leverage points: export licences, evidentiary duties, interface access.
The Android decision is the most structural of the three, because it doesn't regulate what a model may do. It regulates who gets to plug in. In a market where every serious lab now ships a competent frontier model, the question of who reaches the user was quietly becoming the whole competition. Brussels just made that question a matter of law.
The Takeaway
Model quality was never going to be the moat. Distribution was — and on a phone, distribution means the wake word, the screen, and the right to act inside other apps. The Commission's July 16 decisions are the first serious attempt to treat that bundle as infrastructure rather than as a product feature Google owns.
What to watch between now and the August 2027 deadline: how Google's permission and consent flow is designed (a deliberately alarming set of warning screens can neuter an interoperability mandate without technically violating it), how liberally the security carve-out on search-data sharing gets applied, what pricing lands in January 2027, and whether any of it ships outside the EU.
For everyone else, the practical version is simpler. If this holds, the assistant on your phone in a couple of years may be one you picked — not the one that came with the operating system. That has not been true of a default on a major platform in a very long time.