Matthew Prince posted one line about this: it was faster than expected, he said — he'd guessed end of 2027, then early 2027, but agentic traffic grew so fast that bots passed humans online for the first time in June 2026 already.

I read the line. Formed an opinion. Moved on.

A few days later I actually opened the dashboard, which is public and free, and had been sitting there the whole time. It also tells you something the one-liner doesn't: the measure is filtered to HTML responses — web page requests specifically, not every byte on the network. That's a real qualifier, and it took ten seconds to find.

So, I read an announcement about machines that read everything, and I read one sentence of it.

That's the whole post. A lot of subscription businesses are quietly built on the customer being tired. What's arriving is a reader who never gets tired.

The model isn't expiring. The tiredness is.

The tired reader

I've argued that the subscription model outlives every announced funeral, and that where the compute runs is not a pricing decision. Both still hold. The structure survives this too.

But there's a piece of that structure nobody puts in a deck.

Every pricing page I've shipped was designed for a person who is tired. The good stuff goes above the fold because a tired person reads the top. The limits live in a footnote because a tired person doesn't open footnotes. The real difference between two tiers takes three clicks, and a tired person doesn't click three times. He forms an impression and buys.

The tired reader also forgets to cancel. He means to. Some meaningful share of renewal revenue in every subscription business is exactly that, and everyone who has looked at the cohorts knows the number.

None of that was a trick. It's just who the reader was.

The second reader opens the footnote. It reads the terms page, the complaint thread, and every competitor's pricing, in nine seconds, because it has nothing else to do.

What that costs

Four things move, and they're one thing wearing four coats.

Your meter. Every price was calibrated against a human usage distribution. When an assistant does the checking instead of the person, usage stops tracking attention and starts tracking task frequency. One intention becomes hundreds of requests. The price stays flat. This is a slow bleed that looks like ordinary infrastructure creep for months.

Comparison. If you win on an honest side-by-side, this is the best thing that's happened to you in a decade. If you win partly because comparison is hard, that margin is gone and copywriting won't save it.

Access. If your value is "we look at a lot of data and tell you what matters," that's the job agents are getting good at. What they can't replicate is the data. So the question stops being is our summarization competitive and becomes on what terms does someone else's agent reach our data. Most of us answered that years ago, when the thing consuming the API was a person clicking through a partner app. It's a different trade when the consumer is making the purchase decision and you're not in the room.

Cancellation. "Cancel what I'm not using" is one sentence, and it's near the top of what people will actually delegate — tedious and saves money, the two properties that predict delegation. What's left after inertia is use. Engagement stops being a health metric and becomes the leading indicator of revenue.

Which sharpens the free-tier argument I've been having for years. Generosity was supposed to buy goodwill, and goodwill was supposed to buy retention. An agent has no goodwill. It has a baseline. What you gave away is the floor it negotiates up from next time.

The part I should be honest about

I've been on the receiving end of the tired reader for twenty years and never once called it that. I called it conversion. Some of it genuinely was. Some of it was a footnote nobody opened.

And requests are not dollars. A majority of HTML page requests being non-human doesn't mean a majority of purchasing intent is. It isn't close. Traffic is a leading indicator of attention, not of wallet, and that gap is where people are about to over-rotate. This is a twelve-to-twenty-four month architecture question wearing a headline's clothes.

There are probably categories where nobody delegates — where the purchase is emotional, or high-trust, or something people want to choose themselves. I'd believe that. I don't believe it for anything a person currently buys while tired.

What I'd do

Go look at Radar yourself rather than trusting my summary of it, which is exactly the behavior this post is about. Then do the same thing internally: segment your own traffic into human sessions, known integrations, and unidentified automation. One month, then the growth rate on each.

You'll learn either that this is already happening to you or that it isn't yet, and neither answer is on anyone else's slide.

Then ask the harder version of the question. Not what is the customer renewing for — I've asked that one. What would we lose if the customer read everything?

If the answer is nothing, you're fine, and you should be excited about all of this. If the answer is a number, that's the number.

As always, tell me where this breaks. Especially if you've run the cohorts and can tell me how much of renewal is really inertia.