The Attention Deficit

In 2017, eight Google engineers published a paper titled "Attention Is All You Need."

In defining Transformers (the 'T' in GPT…) it became the architectural foundation for ChatGPT, Claude, Gemini, and every large language model that followed. Seven years later, the market decided Google risked being obsolete and the stock was oversold in 2024.

The market had an attention deficit.

Not about AI. About what Google was quietly building while everyone watched the race it had supposedly lost. While investors sold the stock and analysts wrote the obituaries and the technology press declared search dead, Google was embedding the architecture from that 2017 paper into every layer of the most vertically integrated marketing platform ever assembled.

The paper described a mechanism for directing attention and deciding, mathematically, what matters and what doesn't. The irony is almost too neat. The market couldn't direct its own attention correctly. It watched query volumes wobble on Google Search and concluded the student was about to destroy the teacher.

Two AI races. One result.

The real competition in marketing isn't between agencies. The holding company soap opera with WPP restructuring and Omnicom absorbing IPG is a sideshow. The structural fight is between the four layers of the value chain, and the question is which layer ends up owning the economics.

For the purposes of comparing scale, it takes Google just thirteen days to generate the combined annual earnings of the top five holding companies. Put another way, Google earns the entire holding company layer twenty-eight times over every year.

This is Act Three of the Four Layer Battle. Act One introduced the frame. Act Two, The Transparency Gambit, showed how agencies are trying to audit the independent ad tech layer. Act Three belongs to the platform layer.

For those joining the series here: the four layers are (1) clients, who are increasingly building marketing capability in-house; (2) agencies, the holding companies that sit between client and media; (3) the middle technology layer, independent vendors like The Trade Desk and Adobe who position themselves as neutral infrastructure; and at the base, (4) the platforms (Google, Meta, Amazon) who own the inventory and the audiences. Every layer is fighting to capture more of the same economics.

From the platform layer, Google is currently winning that fight decisively. But to understand what Google did, you first have to understand why the market misread the AI transition so badly.

The first race is about intelligence, reasoning, and enterprise workflow. Who can build the most capable model? Who can integrate it most usefully into document workflows, code generation, customer service, and knowledge management? This is the race Anthropic is running. It's the race OpenAI is running. It's the race that produces comparisons of GPT-4o versus Claude versus Gemini on reasoning benchmarks.

Google is competing in that race. But it is absolutely dominating the second one.

The second race is about commerce and marketing. Not commerce as a category on an analyst slide, but commerce as a specific set of activities: reaching the right person with the right message at the right moment, at the lowest possible cost, and being able to prove it worked. This race is decided not by benchmark scores but by distribution, intent data, closed-loop attribution, and the ability to generate revenue for advertisers at scale.

In that race, there is no contest.

On March 23rd, Google held its NewFront event in New York. The headline was "The Gemini Advantage." The substance was something more significant: a complete, closed-loop AI marketing platform that no competitor can replicate in the near term.

Let me make this concrete, because the jargon tends to obscure how remarkable the product suite actually is.

Pomelli is a free tool, now available in over 170 countries, that converts a business's website URL into a complete brand identity profile. It reads your fonts, your colour palette, your tone. It then generates social media campaigns, ad assets, and marketing copy styled exactly to that brand. A small business owner with no agency, no designer, and no marketing team can now produce professional, on-brand advertising from their phone. Google Labs built it. It is free. It is explicitly designed to replace the creative agency function for the bottom of the market.

Veo 3.1 is a video generation model that produces 4K footage with synchronised sound, music, and voiceover from static product images. Google embedded it directly in Google Ads. An advertiser uploads three photos of a product. Veo produces a 10-second YouTube ad with natural motion and professional production quality. The same week this became globally available, OpenAI quietly shelved Sora.

Nano Banana Pro is Google's image generation model. Studio-quality imagery with legible text, conversational editing, and consistency across multiple subjects. Already embedded in Google Ads. Already used to edit 5 billion images in 2025.

Then there is the enterprise layer. At the NewFront, Google announced that its DV360 programmatic platform now offers biddable access to live sports inventory of NBCUniversal, the UFC, MLB, NFL Sunday Ticket on YouTube. Live sports was the last major holdout of traditional TV buying, the segment that kept brands writing cheques to broadcast networks rather than running programmatic. Google just made it available through a single, AI-optimised bidding system.

Gemini is now woven through all of it. Natural language campaign creation. Automated report generation. Creative rejection diagnosis. The system that plans, creates, buys, and measures advertising is now operated, in large part, by Google's own AI.

Model quality is only one variable. Distribution is destiny.

Google processes 8.5 billion searches every day. Those searches are not data points. They are moments of human attention, voluntarily directed, with commercial intent attached. No AI assistant can generate that signal. They can answer questions. They cannot tell you, at this precise moment, that 47,000 people in the UK are searching for running shoes. Google can, and it charges advertisers to be there when it happens.

YouTube generated over £47 billion in revenue in 2025, making it larger than Netflix. It is the number one streaming platform in the United States for the third consecutive year. Its connected TV advertising reach covers 96% of ad-supported US households. DV360, Google's programmatic buying platform, holds 47% of the DSP market by spend, more than double The Trade Desk (and we all know the trajectory of The Trade Desk).

This is the part OpenAI has no answer to. ChatGPT has 600 million users. Impressive. Google's ecosystem has 8.5 billion daily search interactions, 3 billion Android devices, 2.5 billion Gmail accounts, and 200 billion daily Shorts views. You cannot build distribution of that scale. You can only inherit it. And Google inherited it from the two decades before generative AI existed.

The antitrust cases that might have forced structural separation are moving slowly. A December ruling rejected Chrome divestiture. The DOJ's antitrust chief resigned in February. The ad tech remedies ruling, expected mid-2026, may produce behavioural constraints rather than the structural breakup that would genuinely threaten Google's integrated position. Even if forced divestiture is eventually ordered, the appeals process would take years. Meanwhile, Google is compounding.

Meta's ad machine is powerful and its Advantage+ AI suite is genuinely impressive with more than one million advertisers used Meta's generative AI tools to create over fifteen million ads in a single month last year. And yet Meta spent the better part of three years and somewhere north of $50 billion not building the foundation model that would make it competitive in the intelligence race, focusing instead on a metaverse where legless avatars can meet other legless avatars. Reality Labs burned capital and leadership attention at precisely the moment the architecture for everything that followed was being settled. Zuckerberg corrected course, but the correction cost time that compounded in Google's favour.

The deeper structural point is this: Meta and Google are not actually competing for the same thing. Meta is a demand creation platform. It reaches people who are not yet looking for your product and makes them want it. Google is a demand capture platform. It reaches people at the exact moment they are looking. Both are valuable. But in marketing, the moment of intent, when someone types a query into a search bar or watches a relevant YouTube ad, is where conversion happens. Google owns that moment at a scale no one else can match. Meta feeds the funnel. Google closes it.

Which means that the more effectively Meta drives awareness, the more Google benefits. Every brand campaign running on Instagram is building the brand equity that eventually converts on Google Search. They are not rivals for the same budget. They are structurally complementary and Google sits at the more commercially valuable end of the relationship.

A balanced assessment requires honesty about the gaps.

Performance Max, Google's AI-driven campaign format that now drives an estimated 62% of all Google ad clicks, remains a black box in the ways that matter most. Advertisers can see where their money went. They cannot redirect it. A recent study of 503 accounts found that 91% had keyword overlap between Search and Performance Max, with Search outperforming on conversion quality nearly twice as often. The suspicion, shared privately by many experienced practitioners, is that Performance Max cannibalises conversions that would have happened anyway and presents them as incremental wins. Reporting transparency has improved. Control has not.

There is also the creative quality question. Google generated 70 million ad assets using AI in the final quarter of 2025. That is an astonishing number. It is not, by any stretch, an astonishing standard. The internet is already saturated with AI-generated advertising that is technically competent and emotionally inert. The efficiency gains from AI creative production are real. The brand-building consequences of flooding your category with algorithmically optimised mediocrity are a different matter entirely and one I'll return to in a future piece.

And ChatGPT advertising, though nascent, moved faster than almost anyone expected. From launch in February to $100 million in annualised revenue within six weeks. OpenAI is at the very early stages of building an advertising model, but the pace is notable.

None of this changes the structural conclusion. It contextualises it.

When I introduced the Four-Layer Battle framework in March, I wrote that the holding companies' most dangerous competitors were not each other. The real pressure was structural and coming from clients building in-house capability, platforms automating execution, and technology vendors being squeezed from both directions simultaneously.

Google's NewFront announcements confirm that framing in full. The platform layer is not competing for a larger share of the existing value chain. It is collapsing the value chain itself. Every agency function it absorbs reduces the case for the agency layer. Every independent ad tech vendor it displaces reduces the case for the middle layer. Every CMO who automates media buying through Performance Max and creative production through Nano Banana and Veo reduces their dependence on every layer that sits between brief and platform.

Google's business is not search. It is not AI. It is not advertising technology. Google's business is human attention. The capture, curation, and sale of the moments when a human being is actively interested in something.

Eight and a half billion of those moments happen on Google every single day. That is what the 2017 paper was ultimately about. Not language. Not reasoning. Attention itself or the mathematical problem of deciding, from everything available, what matters.

The market spent 2024 with an attention deficit about the company that solved that problem before anyone else knew it was a problem to solve. The engineers who wrote "Attention Is All You Need" gave their field the tool that was supposed to unseat them. The market spent a year deciding they had lost. What actually happened is that Google used the time to build something no model alone can match: a closed loop from creative production to media buying to measurement, sitting on top of the most valuable commercial attention data ever assembled.

The inheritance turned out to be the throne. And the throne, it turns out, was always about attention.

The Four-Layer Battle continues. Act Four will examine what's left for the agency layer once the platform has automated the execution. Watch this space.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 2 April 2026 in the Brandflow newsletter on LinkedIn.