Zovda
Ratings are the most-read number in commerce and the least informative, because an average flattens time, volume, and the reason people were unhappy.
The same 4.2 can describe a product that is improving, one that is failing, and one nobody has bought in a year.
Read the series, not the score
Plot the rating and the review count by week. A product with a steady score and collapsing volume has a different problem from one with falling scores and stable volume.
The shape of the series is the diagnosis. The average is a summary written by someone who never looked.
Separate the complaint types
Delivery, packaging, fit, durability, and support are different failures with different fixes. Most datasets merge them into one sentiment field.
Where the review text exists, tag the complaint type. Where it does not, at least keep the ones that mention a replacement or a return.
Watch for the incentive
Some reviews are written because someone was asked. Some are written because someone was compensated. The distribution of ratings changes shape when an incentive is introduced, and the change is visible in the timing.
Plot review count by day and look for spikes that line up with a campaign.
Keep the text and the date
A dataset with ratings but no dates cannot answer the question anyone actually has, which is whether the problem is still happening.
Zovda keeps reviews as a series through Datuva, where each row carries its date and its source listing. The data research collection shows the series structure.
Tell us what you are deciding about a product. We reply within one business day.