Your About page isn’t copy anymore. It’s training data.

Type your company's name into ChatGPT. Then Perplexity. Then whatever AI summary Google now staples to the top of the page. Read what comes back.

You'll get three confident sentences about what you do, who you're for, and which competitor you're basically a worse version of. Nobody on your team wrote them. A model did — from your site, a review, a comparison page, a directory listing from 2023 — and it's handing that paragraph to your next buyer before they ever reach your homepage.

That paragraph has a name now. Call it the machine read: the description an AI writes about you and repeats to everyone who asks. For a growing slice of your pipeline, the machine read is your positioning. And you've probably never read it. Your competitors have one too, and most of them haven't read theirs either. The first company in a category to write its own answer gets described on its own terms. The rest get described by whatever the model managed to scrape.

The obvious fix is the wrong one

The reflex is technical. Publish an llms.txt — the little file that hands AI a clean version of your site. Add schema. Chunk your pages. A year of trade-press posts told you this is how you control the machine.

Here's the uncomfortable part: the file isn't the lever. We took this apart in an earlier post, and Google says it in its own docs — no special files, no llms.txt, no chunking needed to show up in AI answers. Models read plain language on purpose; that's the whole point of them. Keep your llms.txt. We keep ours. It costs nothing, and it changes nothing about the sentence the model writes about you.

So what writes the machine read?

Your ordinary pages, read in plain English. So the machine read is only as sharp as what your pages plainly say — and most pages plainly say nothing.

The homepage hero is the worst offender. "The OS for human-agent teams." "Unleash your human+AI superteam." "Where people and agents drive results together." A human skims past these and forgives you. A model can't. It won't repeat a sentence that means nothing, so it does one of two things:

• It drops your line and grabs a competitor's clearer one.
• It fills the blank with a review, an old listing, a guess.

Every vague line you write hands a sentence of your description to someone else.

The fix isn't louder. It's checkable.

Write what you are so a machine can repeat it and a buyer can catch you if it's wrong:

• What it is — concrete, one line. The category a stranger would use, not the one you invented.
• Who it's for — the real user, named. "Teams everywhere" is not an answer.
• What it does — the job, in plain words.
• What it won't do — the thing you're deliberately not.

That last line earns the trust, because it's the one nobody fakes. You can see it in the wild. Cohere opens its homepage with a refusal — "Your data. Your infrastructure. Cohere keeps it that way." — a promise a customer can hold them to. PostHog skips the hero games and publishes its whole company handbook in the open: the code, the roadmap, the values, all checkable against the company that wrote them. A model reading either one doesn't have to guess. Neither does a buyer.

Run a real line through it. "We help teams do their best work" answers none of the four, so the model pads the gap from elsewhere. "A crash-reporting tool for mobile game studios; files and triages the reports automatically; not general project management" answers all four. The second is harder to write. That's the point — it commits to something a competitor's copywriter can't lift.

"My buyers don't ask a chatbot what we are"

Maybe not yet. But the AI summary already sits on top of the search they were going to run anyway. The analyst skimming your category pastes your name into a model to save an hour. The junior on the buying committee asks it for a shortlist before anyone books a call. You don't have to believe AI is the front door to notice it's standing in the hallway. The machine read doesn't replace your homepage — it gets there first, and it sets the expectation your homepage then has to live up to.

The honest catch

Checkable cuts both ways. Claim what you can't back, and the model repeats it — confidently, at scale, to everyone who asks — until a customer checks and catches you. A vague tagline is too blurry to be wrong. A specific claim can be held against the product. That's exactly what makes it worth more, and exactly what makes it dangerous if it's a lie. The discipline isn't writing the line. It's only publishing the lines you'd survive someone checking.

You're not writing copy. You're writing training data.

This is bigger than your copy deck. You're not writing taglines for humans to skim. You're writing the source text a machine uses to tell humans what you are.

An AI is going to answer "what is this company" whether you write the answer or not. It's answering right now, for people you'll never meet, in words you didn't choose. Put your answer where the model reads first — the homepage line, the first sentence of the about page, the top of the product page. What you are, who it's for, what you'll never do.

The description is getting written either way. The only question is whether you're the one writing it.