The question shouldn't be about accuracy - they all read what they read - it should instead be about consistency. A dyno cell that can give repeatable results is a useful tool. One that gives wildly different results each time is completely useless.