What a Model Does When Two Sources Disagree

01Not the newest, not the most official

Suppose your company changed its price, its name or what it does, and the internet now holds two versions of you. The natural hope is that an assistant will prefer the newer one, or the one on your own site, because that is where the truth lives. What I observe, in answer after answer, is less flattering. The version that tends to come back is the one that appears in more places. A model does not adjudicate. It counts.

Why counting is the default

I am not claiming to know how any model weighs its sources, and the observed patterns are enough to explain this. Most answers are built from a handful of retrieved pages or from memory formed by reading many pages. If seven pages say one thing and two say another, a search that reads five of them will usually find the first version more often, and memory built from all nine leans the same way. No judgement about which is true is needed for the majority to win. A wrong fact repeated in seven places beats a right fact stated once, even when the one place is your own homepage.

02Most of the old copies are yours

The places you forgot you wrote

When I go looking for the old version of something I changed, the first copies I find are usually mine. One of my products carried an outdated claim about where its data was processed in five separate places on a single privacy page, including the structured data underneath it. Correcting four of them would have left the fifth to keep telling machines the old story. I have also found my own agency described in a file I wrote specifically for machines as an ad intelligence platform, which it has never been. The most persistent wrong description of you is often one you wrote yourself and forgot.

The copies you do not own

After your own pages come the ones you only influence: directory listings, profiles, old press mentions, review sites, forum threads, and the encyclopedic sources covered in the entity you cannot edit. Each of these was accurate when written and has quietly aged since. None of them will update themselves, and a model reading them has no way to know they are stale unless something newer outnumbers them.

03Change the tally, not the argument

Order of work

The practical consequence is that correcting an AI answer is a counting exercise, not a persuasion exercise. Start with every place you control: every page, every piece of structured data, every instruction file, every profile you can log into. Search your own site for the old phrasing and change all of it, not most of it. Then work outward through the sources you can ask to update, starting with the ones that rank or get read most. Then publish something new and dated that states the current version plainly, so the count grows on the right side. You rarely win an AI answer by being right. You win it by being right in more places than you are wrong.

How long it takes

Answers that come from live search can move within weeks, as soon as enough retrieved pages carry the new version. Answers that come from memory move only when a newer model is trained, which is why the old version can survive long after every page you control has changed. That is the same slow layer described in entity hygiene, and it is why consistency has to come before any correction campaign. Keep asking in fresh sessions, with browsing on and off, and record the wording each month. When the new version wins with browsing on, the count has flipped in retrieval. When it wins with browsing off, it has reached memory.

One honest exception. Sometimes the majority version is the true one and you are the outlier, because you are the one who changed the story. If that is the case, the fix is not to argue with the count but to explain the change, with a date, in enough places that the change itself becomes the most repeated fact about you. A change nobody has written about does not exist yet, as far as a machine can tell.

Tomorrow: what happens after a wrong number goes out in public, including one this blog had to fix.