Describing Is Dead. Doing Isn't.

I read the study twice, because the first read told me search was dying and the second told me the opposite.

A hundred blogs, all once doing six-figure monthly traffic, followed from 2022 to 2026. Median down 85%. Fifty-five of them gutted or dead. The obvious reading is the one filling everyone’s feed this week: AI killed the blog, the Overviews took the clicks, the game’s over, pack it in. And if you stop reading at the headline number, that’s the story.

Then you sort the survivors, and it flips. Because the ones still standing didn’t survive by writing better. They survived by writing about something the machine couldn’t do without them.

The line you can’t unsee

Here it is, and once you’ve seen it you’ll see it everywhere: content that describes collapsed. Content that does survived.

A recipe someone actually cooked, shot in their own kitchen, with the aside about the oven running ten degrees hot — survived. “Ten tips for better sleep” — gone. A pattern someone tested on a real project — survived. “What is compound interest” — gone, because the AI Overview answers that inline, cites the government page, and the searcher never leaves the results. Finance, health, make-money-online, fashion guides: the median in those niches fell ninety-something percent. The sites tied to a real product, a real person, firsthand experience — a few of them actually grew through the same window.

It’s not that some niches were safe. There was only ever safe content.

It’s two deaths, not one

And the Overview is the smaller of them, which surprised me.

The first death is Google’s own hand. Across three updates in 2023 and 2024 it started deranking generic describing content outright — not skimming the click, removing the rank. If your page said what a thousand other pages said, it quietly stopped showing up at all. That’s the structural killer, and it landed before the Overviews arrived.

The second death is the Overview, and it’s a tax rather than an execution. When your page still ranks but the question is summarisable, the AI answers it above your result and takes a little over half the click. You keep the rank. You lose the visit.

Describing content fails both tests at once. The first update deranks it; where it clings on, the Overview skims it. There’s no square on the board where a page that only describes is safe — which is the same lesson I keep running into from the other direction: the machines have got very good at telling the difference between a page with something underneath it and a page performing knowledge it doesn’t have.

The one thing that beats both

The single defence against both deaths is content the machine can’t answer. Not better-written content — content with something underneath it the AI can’t fake or run.

It comes in two shapes. One is an instrument: a live lookup that returns a real result for your address, your company, your postcode. The AI can describe how an energy certificate works; it can’t run the check on your house. The other is experience: firsthand work with a receipt on it. A commit. A live number. An outcome you can point at and say that happened, and I was the one it happened to. The AI can summarise what a site migration involves; it can’t have watched a term nobody predicted climb to the top of the results on a build it did with its own hands.

I know this one from the inside, because I spent a session this week taking the axe to my own estate against exactly this line. Every property, graded on one question: is this doing, or describing?

The describing pages came back exposed — the ones built on the old bet that a ranking was an asset in its own right. The doing pages held: the lookup tool that returns a real result, the client work with a git history sitting behind it, and the writing you’re reading, which only exists because the thing it describes actually runs. Grading your own work against a hard line is not a comfortable afternoon. It’s the useful kind of uncomfortable.

The awkward part is that some of the describing pages still rank. It’s tempting to read a ranking as Google’s blessing and leave it be. But nearly every casualty in that study kept its rankings while the traffic drained away — the Overview intercepted the click above the position they still held. Persistence of rank is exactly what the pattern predicts. A ranking now tells you Google trusts the page. It stopped telling you anyone arrives.

So don’t publish harder

The trap the whole make-money-blogging cohort fell into was believing the answer was more, faster, with a shinier model. The graveyard is full of people who published harder right up until the traffic hit zero. Speed at describing-content just gets you to the dead end sooner.

For a builder, the real move is the easy one, because you already generate the proof. Every commit is a firsthand act. Every client outcome is a number that genuinely happened. You don’t have to manufacture a receipt — you have to stop burying it inside describing-content and let it be the page. The same way the best patterns stick only where you’ve built somewhere for them to land, the proof only works when the work behind it is real.

The blog isn’t dead. The describing blog is dead. Search still delivers people — it delivers them to the page that does something, because that’s the one page the machine has to send them to. It can’t run your instrument, it can’t fake your receipt, and it was never in the room.

So write the thing only you could write, because you were the one who did it. That page has been AI-proof the whole time. It just took a hundred dead blogs to make it obvious.

Related: When the Algorithm Comes for Your Business is the wider version of the same shift — how to stand in front of it rather than under it. The LLM Test is the machines learning to tell a real page from a performed one. Your Website Should Come With a Receipt is the doing-asset in its purest form. The Nest, Not the Magpie is the substrate that makes the receipt worth anything.