makefield

Week 4 · the second dimension

Accuracy: is what AI says about you true?

Once a model recognises you, the next question is whether it describes you correctly. Confident and wrong is the dangerous combination.

Makefield · Berlin

The second question

Recognised, then described — correctly or not

Once a model recognises you as a distinct entity, the next thing that matters is whether what it says about you is true. This is accuracy, the second dimension, and it's where the stakes get concrete: a model that knows you exist can still tell a buyer the wrong founding story, the wrong pricing, the wrong category, or a feature you don't have — in the same calm, fluent voice it uses for the facts.

That calm is the problem. A search result that's wrong looks like a link you can ignore. An AI answer that's wrong looks exactly like an AI answer that's right. Confident and wrong is far more dangerous than visibly uncertain.

How it goes wrong

Inaccuracy is usually a gap, filled

Most inaccuracy isn't malice or even a glitch. It's the model filling a gap with whatever was nearest. When the clear, current account of you is thin, the model reaches for a stale one, a third-party summary, or — when it has nothing solid — its own plausible-sounding guess. The failure modes repeat:

It varies by engine

Some engines fabricate; some refuse

Accuracy isn't uniform across assistants — which is why you can't check one and assume the rest. When I asked the same loaded question across six engines, five refused to invent an answer and one played along. And even when engines do cite sources, the sources don't always survive a click: I found cited links that didn't check out at all. So "it showed sources" is not the same as "it's accurate" — the citation can be as invented as the claim.

This is also why accuracy has to be measured per engine, with the same questions, and verified by a human against what's actually true — not taken on the assistant's word.

What it costs

The wrong answer with no second link

On a search results page, a wrong or unflattering result sits next to nine others, and the buyer can weigh them. In an AI answer, the wrong description often arrives as the conclusion, with no obvious second opinion — and the buyer acts on it without ever knowing it was wrong. A quietly inaccurate AI answer can cost you a deal you never saw enter the room.

What accuracy work is

Make the true account the easy one

You can't edit the model directly. What you can do is change what it has to work from — make the clear, current, true account of you the easiest one for a model to find, parse, and repeat, in the places it actually looks. That's the Correct stage of the method, and a later lesson is devoted to it. For now the point is the diagnosis: before you fix anything, you have to know exactly where the description is wrong, on which engine — which is what an audit is for.

Next — Competitor framing & share of voice: even when AI describes you accurately, how do you stack up against everyone else it names? New here? Start with the overview.

Common questions

Questions this lesson answers

Why does AI state wrong facts about my company so confidently?

Because the confidence and the accuracy are produced separately. A model writes in the same calm, fluent voice whether it's repeating a checked fact or filling a gap with the nearest plausible-sounding guess — there's no wobble in the tone to warn you. Most inaccuracy isn't malice or a glitch; it's a gap, filled. When the clear, current account of you is thin, the model reaches for a stale fact, a careless third-party summary, or an invention, and states it exactly the way it states the things it has right. That's why confident-and-wrong is more dangerous than visibly uncertain: the wrong answer doesn't look wrong.

Where does AI get the wrong information about my business?

Usually from one of three places. Outdated facts carried in the model's trained memory — old pricing, a former positioning, a product you've retired. Borrowed descriptions from a directory or comparison page that summarised you carelessly and are now repeated as truth. Or invented specifics — a statistic, a certification, a client name that simply isn't real; while studying this space I read engines attach impressive-sounding credentials to firms that, on checking, didn't hold up. Notice that only the first is about your own site being stale; the other two are the web's account of you, which is why accuracy work reaches beyond pages you control.

If an AI answer shows sources, does that mean it's accurate?

No. A shown source is not a checked source. I've found cited links that didn't survive a click — the citation was as invented as the claim it was meant to support. Accuracy also varies by engine: asked the same loaded question across six assistants, five refused to invent an answer and one played along, so checking one engine tells you nothing about the others. The honest bar is to verify per engine, with the same questions, against what's actually true — not to take "it showed sources" as proof.

How do I fix inaccurate AI answers about my company?

You can't edit the model directly, and no single tag or page forces a correction. What you can change is what the model has to work from: make the clear, current, true account of you the easiest one to find, parse, and repeat in the places engines actually look. But that's the second move. The first is diagnosis — knowing exactly which fact is wrong, on which engine, before you touch anything. Fixing a description you haven't precisely located just adds more noise for the model to average.