You Sound Like the Internet Average

The phrase came out of a script I was writing, and I've been unable to stop thinking about it since.

The scene is a buyer sitting through three vendor pitches in one day. By the third, she's finishing their sentences — not because anyone plagiarised anyone, but because all three decks were generated by the same handful of models, from the same kind of prompt, with roughly the same context. They didn't copy each other. They all independently converged on the mean.

I wrote it as a joke. It's the most serious idea in the campaign.

Convergence is not plagiarism, which is why it's worse

Plagiarism is a discrete, detectable, attributable event. Someone took something. You can find it, prove it, and fix it.

Convergence has none of those properties. Nobody took anything. Each team acted independently and in good faith. The output is original in every legal and moral sense. It is simply the same, because when you sample from a shared distribution with a generic prompt, you get the centre of that distribution — and the centre is where everyone else's sample landed too.

And here's the part that makes it genuinely dangerous: it reads well. A converged draft is fluent, well-structured, grammatical, on-message. Every quality check you'd normally apply passes. The one property it lacks — being unlike anyone else's — is invisible from inside your own document. You can only see it by standing where the buyer stands, which is in a room where they've just heard it twice already today.

This is a new failure mode. Historically, bad marketing copy was bad and looked bad. You could catch it by reading it. Converged copy is competent and looks competent. The tell isn't quality; it's overlap, and overlap is only visible in comparison.

The moat moved

For a while the assumption was that access to better models would be an advantage. It won't be, for the same reason access to a better word processor never was — the capability diffuses to everyone in the category within months, and the price falls to roughly nothing.

What doesn't diffuse is context. Specifically, the context that only exists inside your organisation:

• The exact words your customers use to describe the problem, before you translate them into category language
• The reasons you lost the last ten deals, in the losers' own framing
• The objection you hear so often that your team has an unwritten script for it
• The thing you tried that failed, and why

None of that is in the training corpus. All of it is in your call recordings, your CRM notes, and your team's heads. The differentiated output isn't produced by a better model; it's produced by feeding an ordinary model something nobody else has.

That reframes the work. The leverage is no longer in prompting well. It's in building the retrieval — in getting your actual organisational memory into the context window, reliably, every time. That's an engineering and knowledge-management problem wearing a marketing hat.

The test

One question, applied to any sentence in a pitch, a deck, or a landing page:

Could a competitor have generated this exact sentence from a public description of their own company?

If yes, it's doing no positioning work. It may be true, it may be well-written, and it is still worth deleting, because it occupies space that a differentiating sentence could have used. Every line that survives that test is a line only you could have written — and that, not fluency, is what the buyer in the third meeting of the day is actually able to notice.

The sharper version: the moat is what you've learned not to do

A month of arguing this out on camera has produced a tighter formulation than "context is the moat," and I think the tighter one is the real claim.

Capability is bought. Restraint is designed.

The list above — customer language, loss reasons, standing objections — is all positive knowledge: things that are true about your market. Useful, and still the right starting point. But the item that does the most work is the last one, and it's a different kind of thing. The thing you tried that failed, and why. Negative knowledge. The accumulated record of what doesn't work.

Why that's the durable asset: every capability you're excited about, your competitor has by Tuesday. Model access, tooling, the clever workflow you built last month — all of it diffuses on a timescale of weeks. What cannot diffuse is a record of your own failures, because it only exists if someone bothered to write it down, and almost nobody does. Failures are embarrassing, they're recorded nowhere in any system of record, and the person who learned the lesson experiences it as instinct rather than as information.

That also inverts how the output should be judged. A system that will attempt anything is not more capable than one that refuses; it's less designed. The refusals are the product. Knowing which requests to decline, which drafts to throw away, which market not to enter — those are the decisions a competitor cannot copy from your website, and they're the ones that came from losing something.

The drafting corollary

This has a direct, unglamorous consequence for how I actually write.

When I let a model scaffold a script, the structure comes back good — well-shaped argument, sensible order, clean transitions. And it is, precisely, the internet average of that argument. The fix I've settled on is to mark specific beats in the outline as mine and write them by hand: the opening line, the proof, the introduction of the product, the close. No exceptions and no drafting assistance on those four.

It's the same principle one level down. The model supplies the shape; the only things that differentiate the piece are the parts that could only have come from having done the thing. Skip those beats and you get a fluent, competent script that sounds like every other fluent, competent script — which is the failure this whole piece is about, arriving through the back door of my own process.