When ClaimHit returns a list of potential infringers, the useful question is not "what is the score" — it is "how was each candidate confirmed." ClaimHit answers that directly: every candidate is tagged either Good Match or Possible Match, and the difference between those two labels is the difference between a candidate confirmed on a manufacturer's own page and one that was proposed but could not be independently verified.
This matters practically. A Possible Match on a product with limited public documentation is often more worth investigating than a Good Match on a product with extensive public specs — the former may simply lack a public datasheet, not lack the feature. The labels tell you where the evidence stands today, not where the infringement case ends up.
Why a single AI output is not enough
Every AI model has blind spots. Training data coverage varies, different models weight evidence sources differently, and reasoning patterns applied to the same claim language can diverge significantly between providers.
A single model naming a product is a starting point, nothing more — models can and do hallucinate plausible-sounding products that do not exist, or attribute a real product to the wrong manufacturer. So ClaimHit never treats a model's output as an answer. It treats it as a lead to be checked.
That is the core idea: models propose, evidence decides. Several independent frontier models, from different providers, each propose candidates in parallel. Convergence between them is a useful early signal — but it is not the final word. Every candidate, however many models named it, then has to survive verification against live web evidence.
Stage 1 — Parallel candidate discovery
ClaimHit runs several frontier models from different providers in parallel against your patent's inventive contribution. Each independently proposes products, companies, and technical standards that may implement the claimed invention.
Using models from different providers is deliberate. They have different training data and different failure modes, so a candidate that several of them surface independently is less likely to be a single model's hallucination. But this stage only produces leads — nothing here has been confirmed yet.
Stage 2 — Live web grounding
Every proposed candidate is then searched for on the live web — semantic and keyword search across manufacturer sites, datasheets, regulatory filings, and technical documentation. ClaimHit is looking for a real, reachable page that actually describes the product the models named.
This is where most hallucinated or mis-attributed candidates fall away: if a model invented a product, or attached a real product to the wrong company, there is usually no manufacturer page to back it up.
The pipeline is built around one principle: a model naming a product is a hypothesis, not a finding. Nothing reaches your results as a confirmed lead until it has been checked against evidence that exists outside the model.
Stage 3 — Noise filtering and category-fit
Before verification, ClaimHit removes structural noise — patent-database pages, academic aggregators, articles and guides that mention the technology but are not products, and entries that mimic a manufacturer page without being one.
It then checks category fit: is the candidate actually in the product class the patent covers? A patent on a bimodal HDPE pipe resin should surface resin and pipe manufacturers, not a testing laboratory or a recycled-plastics trader that happens to mention polyethylene. Candidates in the wrong product class are dropped here.
Stage 4 — Manufacturer-page verification
For surviving candidates, ClaimHit attempts to confirm the product on the manufacturer's own domain — fetching the page, checking it is reachable, and matching the product name and claim-relevant terms against the page content.
The outcome of this stage determines the label you see. A candidate confirmed on the manufacturer's own page, with claim-element coverage and verifiable evidence, is surfaced as a Good Match — the strongest leads. A candidate that the models proposed but that could not be independently confirmed on a manufacturer page is surfaced as a Possible Match, shown for further investigation.
Good Match means the candidate was confirmed on the manufacturer's own page with claim-element coverage and verifiable evidence. Possible Match means multiple models proposed it, but it could not be independently confirmed. Both are leads worth investigating; neither is a determination of infringement.
What the labels do not tell you
Good Match and Possible Match are research signals, not legal determinations. They tell you which candidates have public evidence behind them right now — nothing more.
Infringement is a legal conclusion that requires claim construction, prosecution-history analysis, and a qualified patent attorney. ClaimHit's job is to make the first phase — finding and verifying the candidates worth a closer look — fast and evidence-backed, so attorney time goes to the targets that actually warrant it.
Treat any result, Good Match or Possible Match, as a starting point for qualified patent counsel, not a conclusion.