Because the input is mostly links, the score systematically misreads queries where the barrier is something else: brand trust in finance and health, freshness in news, or a SERP occupied entirely by marketplaces and aggregators. A difficulty of 20 on a query where the top ten are all Amazon listings is not an opportunity, it is a measurement error.
Use it as a first filter, then check the SERP by hand for the shortlist. What you are looking for is a page you could plausibly beat: one with weak coverage of the intent, an outdated date, or a format the query no longer wants. That qualitative check is worth more than any score.