First, there's a lot of bad science out there. A large part of this is due to the fact that animal models are almost uniformly terrible at modeling real human disease. Second, a therapy may be highly effective but also highly toxic or have other severe risks or side effects, which are hard to tell in preclinical research. The bar here is very high, and trials can easily be pulled over a few "adverse events." Third, it's rare that investigational drugs are used as first-line therapies; so, the patients getting them are usually the ones in worst shape. This can lead to all kinds of problems in demonstrating efficacy. Also, if your drug does get through trials, the FDA looks very hard at who the drug was tested on and can narrowly limit the indications it receives. So, if you develop "promising drug X", and it's trialled on Stage IV glioblastoma patients who express marker Y and are (for whatever reason) over the age of 45 and mostly female, then you're going to be limited to > 45 year old females with Stage IV glioblastoma expressing marker Y, until you can prove safety and efficacy in another population. But the vast majority of drugs never get anywhere near here: they mostly fail in safety or fail to demonstrate effectiveness early on. (An interesting side note here is that many of the signals you look for to estimate your "effectiveness", like size of tumor, don't always correlate to long-term survival. You can have a drug that appears to shrink tumors and patients still dying. It's really common to see good indicators of progress with little or no impact on survival.)

TL;DR: animal models of disease mostly suck.