Which sources AI models cite, and how concentrated they are
Written by
Odin Lund
Category
AI search
Date
Share this article
Most of the answer comes from somewhere else
When a model answers a question about a business, it assembles the answer from sources it retrieved. We traced 4,222 answers in one category back to those sources and counted where the retrievals came from.
The third-party sources carrying the client produced 18,742 retrievals. The client’s own website produced 4,319 across 63 cited pages. The nearest competitor’s site produced 1,296.
Put plainly: roughly four retrievals in five came from somewhere the client did not own. A business that works only on its own pages is working on about a fifth of what the model reads.
The citations are concentrated, not spread
Forty-six domains carried the client. That number matters more than it looks, because it is small. A category does not have a long tail of equally weighted sources; it has a short list the models return to.
This is good news for the work. If citations were spread evenly across thousands of sites, earning them would be hopeless. Because they concentrate, the useful question is narrow and answerable: which domains did the models actually retrieve here, and is the business present on each with the same facts?
It also means the audit has a definite end. You can enumerate the sources that decide a category.
Your own site still has a job
A fifth of retrievals is not nothing, and the owned share is the part that can be corrected immediately rather than earned slowly.
Sixty-three pages on the client’s site were cited. Those pages do two things: they answer questions directly, and they give the third-party sources something consistent to agree with. A directory listing and a review profile that both match the business’s own stated facts produce a claim the model repeats with confidence.
So the owned work is not separate from the earned work. It is what makes the earned work resolve to one answer instead of a hedge.
What this changes about where effort goes
Start by finding the short list. Run the questions your customers ask, record the citation behind every claim, and count the domains. In most categories a small number will account for most of it.
Then split the work honestly: fix the facts you own, because conflicts there poison everything downstream, and pursue the domains on the list, because that is where four retrievals in five are coming from.
Both at once. Earning a citation on a source that then disagrees with your own site buys you a hedge rather than a recommendation.
Share Blog
