1) Look at relative data, not raw ratings. Written reviews or raw numerical scores are what show the nightclub effect the most. What we care about is a ranking of companies (showing the better companies first in search results). This may not be as impacted by the problem. If sour grapes from people who failed interviews overwhelm other signals, we can only look at the scores from people who pass (or normalize the two groups separately).
2) Ask very specific questions (and maybe provide dummy options to attract sour grapes). I suspect that questions like "did this company say they could meet at $250 salary and then offer you less?" will get more accurate answers than "does this company have a toxic culture?". (We do actually want to get culture data, but I want to be really specific about the culture questions we are asking).
3) Use objective facts, not subject opinions. For example, we can tell when a company ghosts a candidate (because they reply through a proxy email that we control). So we're listing the 'ghost rate' for each company.
