Attention Is Won in the Answer Now

On 10 September we made a prediction in public: new work on this site would keep being taken by machines within about forty-eight hours of publishing. Four days later our own dashboard said we had beaten it by a distance. The two newest posts had waited three hours and five hours.

Neither number survived reading the rows behind it. The three-hour fetch was MJ12bot, which crawls for an SEO backlink index, followed by a curl request and a phone building a link preview. The five-hour one was a run of curl requests and more previews from a single home connection — almost certainly us, checking the page after it went up. The first fetch on the newer post from an AI crawler was GPTBot, about twenty-five hours after publishing. On the older one it was ChatGPT-User, twenty-two hours in: not a crawler at all, but a model fetching the page because somebody had just asked it something.

The prediction held. The dashboard was counting the wrong thing.

Read the rows before you believe the number

The figure was called waited, and it measured the time until any machine took a page. It was built to measure appetite — how quickly a model comes for something new — and it was quietly measuring how quickly we look at our own work.

It was not the only one. The check that separates a real crawler from an impostor trusted only reverse DNS, and some vendors verify a different way: Perplexity publishes the addresses it crawls from, so its real fetches were being marked as forgeries. And more than a third of a week’s log was not anybody reading us at all — SEO tools, our own checks, and one scanner on three cloud addresses that claimed twenty-seven different crawler names while it looked for .env files.

We keep this log because standard analytics cannot see a model reading your site. Building the instrument yourself does not make it honest. A number on a dashboard is a query somebody wrote, and it answers the question they typed rather than the one they meant. Before a figure goes into a decision, or a post, open the ten rows behind it. We nearly published three hours.

A feed pays for reaction, not for usefulness

The first post argued that a model reading you is the funnel now. It did not ask why that channel is worth more than the one everybody already posts to.

Picture the person on the other end: somebody who runs a welding shop and wants to know how to grow it. In our judgment they were never going to find that answer scrolling TikTok, and if they work in a trade, probably not on LinkedIn either. That is our read, not a finding. What the research on ranked feeds does find is why a useful post loses there.

• On Twitter, each moral-emotional word in a tweet raised its expected retweet rate by about 19%, across 563,312 tweets on three political topics (Brady and colleagues, 2017).
• Across 2.7 million Facebook posts and tweets from news outlets and members of Congress, each word about the political out-group raised the odds of a share by about 67% (Rathje, Van Bavel and van der Linden, 2021).
• When researchers compared Twitter’s engagement algorithm with each user’s own chronological feed, the algorithm amplified anger by 0.47 standard deviations — and users rated the political posts it chose lower than the ones it passed over (Milli and colleagues, 2025).
• Facebook’s 2018 ranking change counted a like as one point and a reaction or reshare as five, and its data scientists later wrote that the approach had “unhealthy side effects” on politics and news, according to internal documents published by the Wall Street Journal.
• TikTok says a signal like watching a longer video to the end carries more weight than a weak one. The ranking runs on watch time.

The limit belongs here too. When Meta moved users to a chronological feed for three months, their political attitudes did not measurably change. Ranking changes what people see more than what they believe. For anybody publishing something useful that is the point: the question is not whether a feed will persuade anyone, it is whether your work is shown at all, next to whatever provokes a stronger reaction.

We argued the attention economy in January without sources. These are the sources.

An answer arrives with attention already given

A person asking a model a question is not scrolling. They came with something they want solved, and the answer is assembled for that question alone. The competition for their attention happened before they typed.

It will not stay untouched. ChatGPT has shown ads to logged-in adults on its free and Go tiers in the US since February 2026, chosen by the topic of the conversation. OpenAI says they are labelled, kept separate from the answer, and do not influence it. So the durable position is not a channel without ads. It is this: the answer is still built from sources, and the ad sits beside it.

Being the specific, checkable source a model quotes is a different place to stand from buying the slot next to the answer — and the only one of the two a small business with real expertise can win on merit.

Write down what your work already answers

A post answering one welding-shop owner’s question was never worth commissioning. The audience is small, and none of it is on the feed where the post would have had to compete.

What changes that is where the writing comes from. Here every non-obvious decision is written down as it is made, and that record is the raw material: this week it became the reasoning behind investing in FabWise, drawn from a customer interview and the codebase, and a piece on testing a design rather than its code, drawn from the harness that checks every Kepler mission. Neither was a separate marketing effort. Each was the work, written down.

That makes everybody who builds something of a producer. The decision an engineer records on Tuesday is a draft by Wednesday, agents on several platforms can draft from it, and a person decides what goes out. Long-tail writing becomes affordable when it stops competing with the work for time.

What we expect by the next entry

Measured by the first fetch from a named AI fetcher rather than any machine, new pieces keep being taken inside two days. The figure worth watching after that is the user-triggered fetch — ChatGPT-User, Claude-User, Perplexity-User — because each one is a person asking a model about something we wrote, at the moment they asked. There were four genuine ones in the week before this was written.

What this changes

For a business with real expertise and a small audience, the work is not to get louder in a feed. It is to write down the specific answers the work already produces, keep them where a model can read them, and measure what gets taken — from the rows, not the headline figure.

If your team makes decisions worth writing down and nobody outside the building is reading them, tell us what you’re working on.