About 60% of the pages an AI cites about your industry are different the next day.

WHAT WE ASKED

what sources do AI models read about my industry

About 60% of the pages an AI cites about your industry are different the next day.

When you try to track what an AI model reads about your industry, you might expect a steady shelf of regular reference sites, trade publications, and core company pages. What you actually see on the screen is constant turnover.

We tracked this directly across 26 question-and-model pairs, run on two different days. About 60% of the pages an AI cites change between one day and the next.

That means most of the reading material an AI pulls together to answer an industry question on one day disappears from the source list on the next. The system does not return to a settled index. A specific page cited on Monday is frequently gone on Tuesday, replaced by an entirely different URL answering the exact same prompt.

Why this source list moves so fast remains an open question. We did not alter the models or inspect their retrieval pipelines, so we cannot say what drives the replacement rate. We only observed what the models returned.

This has a practical consequence for anyone acting on a single reading. If you check once and treat that list as the model's view of your sector, you are working from a snapshot that has already changed. In our data the 60% figure was the median across the pairs we observed, meaning half of the runs saw even higher turnover between the two dates.

The pages move, but the sites do not. Across the same corpus, 85% of cited pages appeared exactly once, while 78% of citations landed on sites the models return to. The individual article is the layer that churns. The domain underneath it is the layer that holds.

That changes what is worth chasing. Getting one page cited is a result that may not survive the week. Being present on a site a model reads repeatedly is a position that does. The first is a reading on a given day; the second is the thing that produces readings.

Measured: 26 question-and-model pairs, run on two different days

Not covered: Two dates, not a daily rate. This is the median; half churned more. Page and site figures are from one corpus, not measured per category.

Metrisque · Competitor's Citations · September 2026