Someone else tested it, and published what he found
Charles Eldering benchmarked six AI patent infringement detection tools across 51 patents, then retested ClaimHit after our V3 release and wrote up both the results and the conversation with our founder. He states that he has no affiliation with any of the vendors. We had no control over what he published and we did not see it before it went out.
Against the same 51-patent portfolio, V3 returned 2,195 company and product reads where the previous version returned 293. On his ranking of how much each tool surfaces, we moved from the lowest to the second highest, and he judged that once overproduction is discounted we produced the largest number of potentially viable leads of the six.
A roughly tenfold expansion in distinct companies. He notes, correctly, that the growth is breadth across companies rather than depth within a product line. That limit is real and we are working on it.
The number of our results independently corroborated by at least one other tool rose more than fivefold. Every other tool in the study saw its share of unconfirmed findings fall as a result, because we corroborated reads none of them had found on their own.
Two products, known in advance. The previous version found neither.
One patent in the set had its claims drafted with two particular products in mind, so the correct answer existed before any tool ran. Our previous version found neither of them. V3 found one of the two, and removed a third lead the evaluator had rated weak. Two red boxes to green, in his words, and a significant upgrade in performance on that patent.
It is one patent, so it is an anecdote rather than a result. It is also the only place in the study where a tool’s output could be checked against a known answer, which is why we lead with it rather than with the volume figures above. Those measure how much a tool surfaces. This measures whether it surfaced the right thing.
He also asked how the new version found what the old one missed, and published the answer. Discovery no longer rests on what a single model happens to recall. Six channels run in parallel: web search, semantic search that matches on meaning rather than words, citation lineage, product libraries that are not indexed as pages, competitors of the companies already found, and several models cross-checking each other. A model never having heard of a product is no longer a reason the product goes unfound.
One thing worth saying plainly. Our results come from a retest on a portfolio we had already seen, against a version released after his first article. The other tools were measured on a single blind pass. That is not a like-for-like comparison and we do not present it as one.
Published on Light Drafts, September 2026. The study compares six tools by name. We link to it rather than reprint it, because ranking other people’s products is his work and not ours.
See who is using your patent.
The pilot is one patent: the full product search and one Hit Chart, with a 30-day money-back guarantee. No subscription, no card on file after that.