Does pushing back on a refusal actually work? We can only partly answer that

Every FOI response includes a standard notice that you can ask for an internal review if you're unhappy with the answer. Most people don't. We wanted to know, for the ones who do, whether it actually changes anything - and the honest answer is that we can only partly tell.

Getting the measurement right took two attempts

Our first pass at this searched response text for the phrase "internal review" and found it in nearly a third of everything we track - which turned out to be almost entirely false positives. Nearly every FOI response includes a standard appeals-procedure paragraph explaining how to request a review, regardless of whether anyone ever did. Searching for that phrase mostly just measures how universal the boilerplate is, not how often anyone used it.

Fixed properly - requiring language that shows a review actually happened, rather than the generic rights notice - the real rate across the whole dataset is 2.2% (504 releases). Within health & NHS specifically, our sample is smaller (21 releases), too small to report a separate health-specific rate with any confidence - we'd rather say that plainly than manufacture a precise-looking percentage from 21 documents.

What the reviews actually decided

Here's where we have to be honest about a limit rather than paper over it. Real review verdicts turn out to use varied, often legalistic language - one genuine "upheld" decision we read in full, involving a commissioned leadership review at an NHS trust, never uses the word "upheld" at all, reasoning instead through a formal public-interest balancing test. Across the whole dataset we can confidently classify an outcome either way in only 14 of the 504 review releases we found, and just 7 of those are health-related. That's too small a sample to report as a real statistic, and we're not going to dress it up as one.

Reading this responsibly

"Genuine review" here means language suggesting a review actually took place, not just the presence of the standard rights notice - a real methodological fix made partway through this analysis, documented rather than hidden, since the first attempt would have overstated how often people push back by more than tenfold. The dataset-wide mention rate is solid; the outcome split is not, especially at health's smaller sample size, and we're reporting that limitation as the finding rather than forcing a headline number from a handful of documents.