Most restaurants on most platforms sit within a narrow band near the top of the scale, which makes the ratings almost useless for comparison. The compression is produced by who leaves reviews and by how platforms collect them.
Reviewers are not a random sample
Leaving a review takes effort, and the people who make that effort are disproportionately those who had an unusually good or unusually bad experience.
Ordinary, acceptable meals, which are the large majority, generate very few reviews of any kind.
The result is a distribution with a peak at the top, a small cluster at the bottom and very little in the middle where most meals actually sit.
Prompting shifts the balance
Businesses actively ask satisfied customers to review them, often at the moment of payment when the experience is freshest and most positive.
Dissatisfied customers are rarely asked and frequently just leave, so the solicited portion of the reviews skews high by design.
Delivery and booking platforms that prompt every user after every order collect more representative data, which is why their averages sit lower and spread wider.
The scale is not used as a scale
People treat a five-point scale as a pass or fail judgement rather than a graded measurement, so the middle values are used far less than the ends.
A merely fine meal often receives the top rating because the reviewer does not want to harm a business over something unremarkable.
Once that norm is established, a middling score reads as a complaint, which pushes future reviewers further toward the top.
Platforms filter what appears
Review systems suppress entries they judge to be fraudulent, incentivised or from suspicious accounts, and the criteria are not published.
Those filters remove a substantial number of reviews in both directions, and the displayed average is calculated only from what remains.
Two platforms showing different averages for the same restaurant are usually not disagreeing about the food but working from different surviving sets.
Volume carries more information than the score
A restaurant with a great many reviews has been sampled across seasons, staff changes and both quiet and busy nights.
A high average built on a handful of reviews carries almost no information and is the easiest kind of rating to influence.
Reading the substance of recent reviews for recurring specifics is more informative than any comparison of averages, because specifics survive the compression that scores do not.