Field note

The ghost in the logged-in answer

I asked an AI what my own project was. Logged in, it knew me. Logged out, it had never heard of me.

Denes Csaszar · Makefield · Berlin · not yet published

I asked an AI assistant what my project was — while signed into the account I use every day. It answered confidently and accurately, describing the thing I'm building as if the world already knew about it. For a second, it was thrilling.

Then I opened a clean, signed-out window and asked the exact same question. The confident answer was gone. What came back was the real picture: a name the model didn't actually recognise, surrounded by a few unrelated namesakes.

Here's what happened. Signed in, the assistant was drawing on my own past chats and saved memory — it was repeating my own idea back to me. That isn't visibility. It's a mirror. The flattering version existed for exactly one person in the world: me.

This is the single most common mistake people make when they go to "check what AI says about us." They check it logged in, with memory and personalisation on, and the engine reflects their own information back at them — their site, their phrasing, their framing. It feels like proof. It's the opposite: it's the one reading you can't trust, because no buyer is inside your account.

The fix is one line: test your AI visibility the way a stranger would. Signed out, no history, a fresh window, every time. If you have to stay signed in because an engine limits what it'll do otherwise, turn memory and custom instructions off first — and write down that you did.

The gap between the two answers is the whole point. The logged-in version is who you wish you were to the AI. The clean version is who you actually are when a buyer asks. Only one of those is worth measuring.

This is why an audit runs clean by design — see also ask the same question twice, get two answers.