Record Revenue Is What Peaking Looks Like

Google just posted some of the best quarters in its history. Search revenue up 15 percent year over year.¹ Meta up 33.² If you wanted proof the attention economy is fine, there it is.

I think it’s the wrong number, and if your business depends on being found, chosen, or bought, here’s my argument as to why.

The money is the last to know

Ad revenue doesn’t measure what people are doing. It measures what advertisers believe people are doing, and advertisers budget on habits formed years ago. Newspapers billed healthy ad revenue for a decade after their readers left. The money is always the last thing to move.

So skip the earnings calls and ask a simpler question: where do people go now when they have a question that matters?

Where people actually go

ChatGPT reached a billion monthly users faster than any application in history.³ 37 percent of consumers now start their searches with an AI tool, up from basically zero two years ago,⁴ and 60 percent say the answers are better.⁵ Among people under 35, more than 70 percent use these tools.⁶ That isn’t a default yet. It’s how defaults start.

Google still processes more raw queries a day than every AI tool combined,⁷ which sounds like the whole argument until you look at which queries are leaving. Search is keeping the reflexive stuff: store hours, weather, navigation. AI is taking the multi-step research, the comparisons, the “which one should I actually buy” sessions. Pew found that long, conversational, question-shaped queries are exactly the ones being routed to AI now.⁸

Search kept the exhaust. AI took the intent. To be precise about it: the split is a direction, not a finished fact. Google still fields an enormous volume of commercial queries today, and careful cohort studies show something the doom takes skip over, which is that people who adopt AI don’t actually search less.⁹ Total information-seeking is expanding. Both things are true, and neither rescues the ad model, because the ad business doesn’t monetize curiosity. It monetizes the moment before a purchase, and that’s the slice migrating fastest. A platform can grow total volume, keep 90 percent of it, and still lose the queries that carried the money.

The stat nobody pulls out

A growing share of “searches” aren’t people at all. Cloudflare reported this summer that automated traffic crossed 57 percent of the web, a machine majority their own CEO didn’t expect until 2027¹⁰… and to be clear, a meaningful chunk of that is still traditional crawling. Bots have fed search indexes since the beginning; that was the old bargain, and publishers got a click out of it. What’s new is the fastest-growing slice: retrieval on behalf of a person who never arrives. When an AI answers your question, it usually does it by querying a search index, and that query gets counted as a search.

So a meaningful part of the disruption is being recorded as loyalty. You skip the search box, ask an AI instead, and the search engine’s volume goes up anyway… no human saw a page, saw an ad, or formed a habit. Strip the machines out and the human exodus is further along than any published chart shows.

Google can read its own logs, of course. That’s why it’s racing to ship browsing agents and agent-to-merchant checkout protocols.¹¹ You don’t build the thing that replaces your front door unless you can see the front door closing. Losing a platform war never looks like collapse from the outside. It looks like record revenue and a quiet scramble to rebuild on the far side of your own moat.

The write-only web

Old programmer joke: write-only memory. You can store anything, you just can’t read it back.

That’s the internet the machine majority is building. Somewhere right now an AI is writing an article to rank for a keyword. Another AI will crawl it. A third will summarize it, maybe for a person who skims two sentences, more likely for an agent that needs one number out of it. Most of that text will never touch a human. Real electricity, real GPUs, real invoices, spent generating and consuming words nobody reads. Dark economy is the right name for it, dark the way dark matter is dark: enormous mass, no light. Ad impressions served to bots. Articles written by machines to persuade other machines. Economic activity that never touches a life.

Vanity compute

Underneath the dark economy there’s a stranger layer, and I’ve started calling it vanity compute: tokens burned on output that nothing consumes at all. The app you generated at midnight and never deployed. The essay nobody read, including, if you’re honest, you. The dark economy at least has a consumer somewhere in the loop. Vanity compute doesn’t. It isn’t an economy. It’s a bonfire with a dashboard.

Some of that is play, and play is how you learn a tool. Fine. But nobody is measuring vanity compute at industrial scale, which is part of the problem, and the content industry is about to pour real budgets into it, because every trick in the old playbook aims that direction.

The reader with no eyes

The old playbook was engineered for eyeballs. Keywords, clickbait, pop-up timing, headline formulas. The new reader has no eyes. An LLM pulling your page doesn’t see the banner, can’t be retargeted, doesn’t care that your title promised a secret. It reads everything and can’t be dazzled.

What still gets through is almost embarrassingly old fashioned. A reputation for being right over years. Answers that are genuinely better than the other answers. Things that haven’t already been said a thousand times. The kind of content only people can write. The machines are drowning in machine-made content, and the way above the noise floor is being worth citing.

Who pays the agent

Follow the intent, because it didn’t evaporate. The most valuable moment on the internet, a person deciding what to buy, is moving into agents: software that knows you, remembers your constraints, and increasingly buys on your behalf. The position Google held for twenty years is being handed off right now. I’ll be straight that this is the most forward-looking claim in the piece: today’s agents mostly recommend, and only a small share actually transact. But checkout protocols are beginning to ship, and funding models harden early, long before anyone is paying attention to them.

The thing that decides whether the next era is better or just concentrated is who funds that software. An agent paid by the merchants it chooses between, through commissions and placement fees, isn’t your agent, it’s a lobbyist wearing your colors, and we’ll have rebuilt the old machine with a friendlier voice. An agent paid by you has your incentives, and unlike a feed it leaves receipts: if it bought you the worse thing, you lived with the worse thing, and you can go back and check. A feed never gave you that.

One question before you generate

Stop watching the revenue, it’s a rearview mirror, and watch where people take their real questions instead. But the practical takeaway is smaller and more personal. Before you generate the next thing, whether it’s a landing page, a content calendar, or an app at midnight, ask whether anyone, human or machine, will ever actually consume it. If yes, make it worth citing. If no, you’re just flaring tokens.

The write-only web is getting built either way. You don’t have to live there.