The Wake Word Is the Moat: The EU Pried Open Android's Assistant Slot

On July 16, 2026, the European Commission did something more consequential than levy a fine, and it did it so quietly that most of the coverage missed what actually changed. It adopted two binding specification decisions under Article 6(7) of the Digital Markets Act, closing proceedings it had opened against Google back in January. The first orders Google to give competing AI assistants effective access to eleven Android features that today only Gemini can use fully. The second orders Google to share the anonymized data it collects to sharpen its own Search with rival chatbots that offer search, from January 2027, under a fixed pricing formula. Together the measures reach the roughly sixty percent of EU smartphone users who carry an Android device.

Read the headlines and you would think this was a story about privacy, or about Google getting punished again, or about European users eventually being allowed to say "Hey ChatGPT" instead of "Hey Google." Those framings are not wrong exactly, but they bury the thing that matters. What the Commission did on July 16 was declare, in the form of an enforceable order, that the competitively decisive asset in consumer AI is not the model. It is the slot the model gets to occupy on the device — the wake word that summons it, the context it is allowed to see, and the permission it has to act inside other apps on your behalf. Everything else in the AI industry has been arguing about capability. Brussels just regulated distribution.

That is the argument of this piece: the moat in consumer AI was never going to be the weights. It was always going to be the assistant slot, and the EU has become the first regulator on earth to say so out loud and attach a deadline to it.

What actually landed on July 16

Precision matters here, because the coverage overstated the timeline in ways that change the strategic read. So let me state the measures as the decisions actually frame them, not as the aggregators paraphrased them.

The Commission adopted two sets of binding specification measures under Article 6(7) of the DMA, the provision that lets it spell out, in detail, what a gatekeeper must do to comply with an existing obligation. This is not a fine and it is not a preliminary view. It is an order specifying concrete engineering and data-sharing behavior, backed by the DMA enforcement machinery, which can escalate to penalties of up to ten percent of global annual turnover, and up to twenty percent for repeat breaches.

On the Android side, Google must give third-party AI assistants effective access to eleven operating-system features that its own Gemini assistant already uses to hear, understand, and act for a user. These cover the full stack an assistant needs to be a real assistant rather than a chat window: voice invocation, so a rival can be triggered hands-free; App Functions and cross-app action, so it can book a taxi or draft a reply inside another app on your behalf; access to contextual data, so it knows what you were just looking at; background execution; and access to device resources. The Commission gave concrete examples of the end-user behavior it wants unlocked: booking a taxi by voice, getting suggested replies inside a chat app, and asking about a place you recently visited.

On the Search side, from January 2027 Google must share with eligible third parties — explicitly including OpenAI and other AI chatbots that offer search functionality — the same data it collects to optimize its own search results, subject to a multi-layered anonymization method and a fair, transparent pricing formula set out in the decision.

The gap between those two readings is the whole story. The reason earlier EU remedies underdelivered is that they operated at the wrong layer. A browser ballot changed which icon you tapped; it never touched the default that ran when you did nothing. This order is aimed squarely at the default-behavior layer. It is the first structural attempt to make the ambient assistant a contestable market rather than a property right that ships with the OS.

Why invocation, not intelligence, is the contested asset

To see why this is the important move, you have to be honest about where the value in a consumer AI assistant actually accrues, and it is not where the industry spends its marketing budget.

The last three years of consumer AI have been narrated as a capability race. Which model reasons better, which scores higher on the arena, which writes cleaner code. That race is real and it is fierce, but it has quietly commoditized the thing everyone assumed would be the moat. Open-weight models now land within a few months of the closed frontier as a matter of routine — a dynamic I traced in the durability discount and what it did to the AI trade, where a single strong open-weight coding release helped mark down the terminal value of the most profitable franchise in the industry. When capability converges, capability stops being the moat. What is left is distribution: who the user reaches for without thinking.

And on a phone, reaching without thinking has a precise technical meaning. It is the wake word, the long-press, the assistant that answers when you speak to the device rather than to an app. That surface has three properties that make it the real prize, and none of them are about how smart the model is.

The chart is a qualitative estimate, not a measured statistic, and I want to be clear about that. But the shape is the point. The layer where the incumbent held almost total control was never the model — capability is the one column where a challenger could already compete on merits. The incumbent advantage lived at invocation, context, and action. Those are exactly the three surfaces the eleven features address.

Consider each. Invocation is winner-take-most because a phone has one wake word and one long-press gesture, and the OS owner assigns them. If saying "Hey Google" launches Gemini and there is no supported way to bind a different phrase to a different assistant, then no matter how good a rival model is, the user has to leave their habit and go tap an icon to use it. Habit is the most durable moat in consumer software, and defaults manufacture habit. Context is decisive because an assistant that can see what you were just doing — the message you are replying to, the place you just looked up — can help in a way that a context-blind chat window cannot, and Google reserved the richest context signals for its own assistant. Action is the endgame: an assistant that can complete a task inside another app, not just describe how, is a different product category, and cross-app action required precisely the system hooks Google kept for Gemini.

Own those three and you own the assistant, regardless of whose model is underneath. That is why the Commission went after them, and it is why this order is a bigger deal than any fine Google has paid.

The wake-word nuance the headlines got wrong

Here is where precision earns its keep, because the single most-repeated line in the coverage — that Europeans will soon say "Hey ChatGPT" instead of "Hey Google" — attaches to the wrong year, and the gap between the years is itself the story.

The bulk of the Android interoperability work is tied to Android 18 and is due no later than August 1, 2027. But concurrent wake-word access — the specific capability that lets separate assistants be triggered by different wake words at the same time, which is what actually makes "Hey ChatGPT" work as a live alternative to "Hey Google" — carries a later deadline of August 1, 2028, tied to Android 19. The headline capability is the last one to arrive, roughly two years out, and a full year behind the rest of the package.

Why does the sequencing matter so much? Because a remedy that arrives in stages lets the incumbent bank the habit advantage during the gap. Between the bulk of the interoperability landing in August 2027 and concurrent wake words landing in August 2028, a rival assistant on Android can act inside apps and read context but still cannot own a hands-free trigger of its own. For a year, the most important surface — the reflexive, spoken invocation — stays effectively Google's. In platform competition, a year of retained default behavior is not a rounding error. It is enough time to convert a remedy-driven opening into a re-entrenched habit, which is exactly the risk a challenger has to price in before it invests in the European market at all.

This is the pattern to watch across the entire order. The Commission got the layer right — it is regulating invocation, not icons. But the value of getting the layer right is discounted by every month of sequencing delay, and the most valuable month is the last one.

The second decision is the one that scares Google more

Lost under the Android coverage was the Search data-sharing order, and if you ask which of the two decisions Google fought harder, the honest answer is the data one. The Android decision reallocates a surface. The data decision reallocates an input that a rival cannot manufacture at any price.

A search-capable AI assistant is only as good as the freshness and coverage of what it can retrieve. Google's twenty-plus-year advantage there is not really its ranking algorithm; it is the behavioral exhaust of billions of queries — what people clicked, what they refined, what satisfied them — that it feeds back to sharpen results. That data is not for sale and cannot be recreated by a challenger with more compute, because it is a function of already having the users. It is the closest thing in the industry to a genuine structural moat.

The Commission ordered Google to share exactly that category of data — the signals it uses to optimize its own Search — with eligible third parties including OpenAI, from January 2027, under anonymization safeguards and a mandated pricing formula.

If the Android order says the assistant slot must be contestable, the data order says the retrieval layer underneath every search-capable assistant must be contestable too. Taken together they are an attempt to open both the front door and the foundation of Google's position in one motion. That is why Google's objection, voiced by Kent Walker the same afternoon, leaned so hard on privacy and security rather than on competition: the strongest available argument against sharing query data is that doing so endangers users, and it is a real argument, but it is also the argument a moat-holder reaches for when the moat itself is the thing being ordered open.

Google's defense is real, and it is also a moat in disguise

I do not want to caricature Google's position, because the privacy and security concerns are legitimate and the Commission clearly took them seriously — the decisions build in explicit safeguards. Understanding those safeguards is also how you understand where Google will try to slow the whole thing down.

Walker's argument runs roughly like this: an AI assistant with system-level access to your microphone, your screen, your app data, and the ability to act on your behalf is one of the most sensitive pieces of software that can run on a phone. Forcing that level of access open to any third party that asks weakens the vetting that keeps malicious or sloppy assistants away from the crown-jewel permissions. Phone makers already play a gatekeeping role in deciding which assistants get deep access, the argument goes, and detaching sensitive permissions from that vetting trades user safety for a competition abstraction.

This is not a bad-faith argument. It is a genuinely hard tradeoff, and it is the same tension that runs through every attempt to make a powerful platform interoperable: the hooks that let a good assistant help you are the same hooks that let a bad one harm you. The Commission's answer is to require explicit user consent for covered access, and — critically — to let Google set objective, non-discriminatory criteria on five of the most sensitive capabilities, including screen automation, system integration, and centralized access to app data.

Watch that number. Five of the eleven features come with a carve-out that lets the incumbent define who qualifies, so long as the criteria are objective and non-discriminatory. That is a reasonable safeguard and a potential loophole at the same time. The entire enforcement battle from here forward will be fought on the word "objective." Google will propose criteria it can defend as neutral security requirements; rivals will argue those criteria are gerrymandered to keep them out of the crown-jewel permissions; the Commission will adjudicate. The remedy did not end the fight. It relocated it from "must Google open access" to "are Google's conditions on that access a fair fence or a disguised wall." That second fight is slower, more technical, and far easier for an incumbent to win by attrition.

Why this is the most consequential AI-competition action so far

Set this against everything else regulators have done in AI, and its distinctiveness comes into focus. Most AI regulation to date has aimed at models: the EU AI Act's risk tiers, disclosure and evaluation regimes, safety commitments, the question of what a frontier system is allowed to do. All of that governs the technology. Almost none of it governs the distribution — who gets to put the technology in front of a user by default. This order is the first major regulatory action aimed not at what an AI can do but at who gets to be the one you talk to.

That is a category shift, and it rhymes with a broader pattern I have been tracking: regulation increasingly shapes the AI market not by constraining capability but by allocating position. When China's approval process determined that Apple would ship Qwen because Qwen was on the approved list, that was the model layer being turned into a regionally licensed component, which I wrote up in the regional model layer. The EU's Android order is the consumer-distribution analogue: the assistant slot turned into a regulated, contestable surface rather than a property right that ships with the platform. In both cases the state, not the market, is redrawing where the value in AI is allowed to sit.

The revenue figures are approximate and rounded, and I include them only to make one point concrete: the DMA penalty ceiling of ten percent of global annual turnover, rising to twenty percent for repeat breaches, is measured against a number near four hundred billion dollars. That is what gives a specification decision its teeth. Google will litigate, delay, and shape the criteria — but it will do so inside a regime where non-compliance is priced in tens of billions, not in the low-single-billion fines it has absorbed as a cost of doing business before.

The prior fines, real and large as they were, shared a weakness: they punished past conduct without reliably reshaping future behavior. Google paid and, in the critical respects, continued. The DMA specification model is designed to fix exactly that. It does not ask a court to find abuse after the fact; it specifies the behavior in advance and escalates if the behavior does not appear. Whether that design actually works is the open question the next two years will answer, but the design is aimed at the thing the old fines could not touch: the ongoing default.

The Brussels effect, and its limits

The instinctive next question is whether this stays a European story. History says partly yes and partly no, and the shape of the partial spillover is what a strategist should actually plan around.

The optimistic case is the Brussels effect: the EU sets a rule, the compliance cost of maintaining two versions of Android exceeds the cost of shipping one, and the European standard becomes the global default. It has happened before, with privacy notices and with USB-C. If Google builds the interoperability hooks for Europe, the marginal cost of exposing them elsewhere is low, and rivals will pressure other regulators — the UK's DMCC regime, and antitrust authorities in the US and elsewhere — to demand the same. In that world, the assistant slot becomes contestable well beyond Europe on a lag.

The pessimistic case is fragmentation. Google has every incentive to scope the changes as narrowly as possible, to ship the open hooks only where legally compelled, and to lean on the "objective criteria" carve-out to keep the most valuable permissions gated everywhere it can. The concurrent wake word, the single most consumer-visible capability, is exactly the kind of feature a company confines to the jurisdiction that forces it. In that world, a European can say "Hey ChatGPT" in 2028 and an American still cannot, and the assistant market stays a Google default everywhere the DMA does not reach.

Which future arrives depends less on the text of the decision than on two things the decision cannot fully control: how aggressively Google uses the criteria carve-out, and whether rivals actually invest in the European opening or treat it as too small and too slow to bother. A remedy is only as strong as the competitor willing to walk through the door it opens. If OpenAI, Perplexity, and the rest decide the European assistant market is worth fighting for, the order becomes a genuine turning point. If they decide two years of criteria disputes for sixty percent of one region is not worth the engineering, Google's inertia wins by default, and the most important AI-competition order yet becomes a footnote.

What this means if you are not a regulator

Strip away the Brussels-versus-Washington framing and the order carries a lesson for anyone building in AI, including people who will never touch the European market. The lesson is about where to locate your defensibility.

For most of the current cycle, the assumed answer was the model — train or fine-tune something better and the users follow. This order is a regulatory acknowledgment that the market has already moved past that assumption. When the state has to force open the assistant slot to create competition, it is conceding that the slot, not the model, is where the durable advantage sits. The layer worth regulating is a reliable signal for the layer worth owning. Brussels just told you, in the form of an enforcement order, that the layer worth owning is invocation, context, and action.

If you are building a consumer AI product, that reframes the roadmap. Capability gets you considered; it does not get you reached-for. The reached-for position is a function of where you sit in the user's reflex loop — the wake word, the default, the surface that runs when the user does nothing deliberate. On platforms you do not control, that position is assigned by someone else, and this order is a reminder of how much that assignment is worth and how hard the assigner will fight to keep it. The parallel in the enterprise is the control plane: whoever owns the authorization and routing layer owns the account, which is the dynamic I traced in the agent gateway as the enterprise control plane. Consumer and enterprise are converging on the same truth from opposite ends — the model is the commodity, and the point of control is the surface that decides which model the user or the account actually reaches.

What would prove this read wrong

I have argued that this order matters because it regulates the right layer, and that its value is discounted by sequencing and by the criteria carve-out. Here is what would tell me I have overweighted it.

If, by the August 2027 deadline, rival assistants on Android in Europe demonstrably gain voice invocation and cross-app action and users start switching in measurable numbers, the "invocation is the moat" thesis is confirmed and the order is doing its work. If instead the deadline arrives and the practical experience is unchanged — because Google's objective criteria fenced out the serious challengers, or because the challengers never invested — then either the moat was never really at invocation, or a remedy at the invocation layer is too easy for an incumbent to neutralize through conditions. Both of those would be evidence against my read, and both are live possibilities.

The slot was always the prize

The story the industry has told itself for three years is that consumer AI is a capability race, and that whoever has the best model wins the user. The EU's July 16 order is the clearest signal yet that this story was incomplete. Capability has converged toward commodity; open weights land within months of the frontier; the differences between the best assistants are increasingly differences a user cannot feel. What has not converged, and what the whole edifice actually rests on, is distribution — the assistant slot on the device, the wake word that summons it, the context it sees, the apps it can act inside.

The Commission looked at Google's position and concluded, correctly, that the durable advantage was never the model. It was the eleven features that let one assistant be ambient while every rival stayed app-bound, and the query data that let one search be sharpened by twenty years of behavior no challenger could replicate. So it ordered both opened. Whether that order becomes a turning point or a regional footnote depends on sequencing it does not fully control, on criteria it will spend two years adjudicating, and on whether Google's rivals decide the opening is worth walking through.

But the framing is now permanent, and that is what makes July 16 matter more than any fine. A regulator has stated, in enforceable terms, that the contested asset in consumer AI is the slot, not the model. Every company in the industry spends its budget on the model. The one institution with no product to sell just told them where the moat actually is. For the day-by-day version of the decision and Google's response, see my news analysis of the two DMA specification decisions.