Radiology and imaging software, measured across four AI engines

Radiology and imaging is a category with real third-party infrastructure. Analyst firms rate it, health systems publish selection criteria, professional bodies issue guidance. That is more external validation than most B2B software categories have, and it makes the citation structure worth reading closely.

A broad source list with a very flat distribution underneath it.

Six hundred and twenty-two citations across 187 domains looks like breadth. Ninety-four of those domains were cited exactly once, which means half the source list contributed a single reference and never came back. The tail is long and it carries almost nothing.

KLAS is the reference point buyers in this category name without being asked, and it was the largest single third-party source in the sweep. It took 9 of the 622 citations, under two percent.

That is not a criticism of KLAS. It reflects how retrieval works: an engine composing an answer opens whatever it can read, and a great deal of analyst material sits behind registration or subscription where a retrieval system cannot reach it. Coverage that a buyer values highly can be nearly absent from the machine that now writes the shortlist.

The practical reading for a vendor is that analyst relations and AI search visibility are not substitutes. A strong analyst position does not automatically produce engine visibility, because the engine may never have read the report.

It is worth setting this beside our own agency category, which we measured with the same four-engine method and published in full.

Similar totals, very different source composition. The agency category had no third-party source above the vendor domains at all.

Two categories with almost the same citation volume need different work. In one, the engines have independent material to reach for and a vendor competes to be described accurately in it. In the other, the engines read the vendors and nothing else, so the vendor's own published surface is the entire input. Spending on the wrong one of those is the common and expensive mistake.

Three things follow from a 622-citation, 187-domain, 94-once structure.

Chasing breadth is not the lever. A domain that cites you once and never again moves nothing. Identify the sources that repeat across runs and work on those. Second, analyst coverage should be treated as valuable for buyers and unreliable for engines, which means the same material needs a readable public form. Third, measure the named count separately from the citation count, because a category with a rich source layer can read you often and still name someone else.

Our method is published at github.com/Broadcastwell/audit-method so it can be run without us, and our own score against it, 0 of 40, is at our own visibility page.