A 4.3 average tells you almost nothing. It is compatible with a product everyone mildly likes and with a product half the market loves and half finds unusable — two completely different situations requiring opposite responses.
The information is in the text. Reading it systematically is a skill, and it is mostly about resisting the ways it misleads.
Read the middle
The most useful reviews are three-star, and almost nobody reads them.
Five-star reviews are written in the first flush of enthusiasm, often during onboarding, sometimes in exchange for a gift card. They tell you what impressed someone before they used the product properly.
One-star reviews are written in anger, frequently about a billing dispute or a support interaction rather than the product, and often by someone who was never the right customer.
Three-star reviews come from people who use the product, get value from it, and can articulate exactly where it stops working. That is the most decision-relevant perspective available, and it is sitting in public being ignored.
Separate the product from the relationship
A large fraction of negative reviews are not about the product at all. They are about billing, support responsiveness, a failed migration, or a salesperson who overpromised.
These are real problems and they are different problems, with different owners. Sorting reviews into product complaints and relationship complaints before analysing them prevents a support staffing issue from being read as a product quality issue.
The sorting is usually obvious once you look for it, and it changes the conclusion often enough to be worth the step.
Count, do not impress
One articulate, detailed complaint feels like a finding. It is an anecdote.
The discipline is counting independent mentions of a theme. Three or more independent sources raising the same specific issue is a signal. One vivid account is a story, and stories are dangerously persuasive precisely because they are memorable.
Keep the count visible in whatever you produce. “Onboarding difficulty, 14 mentions across two sites, 9 in the last six months” is a finding a reader can weigh. “Users report onboarding difficulties” is not.
Weight by date, always
Reviews accumulate and never expire, which means a long-fixed problem keeps generating evidence forever.
A product that had a bad year two years ago will carry those reviews indefinitely, and an analysis that does not weight by date will report a resolved problem as current. Conversely, a recent cluster of complaints on a previously clean dimension is the most actionable pattern in review data — something changed.
Practical rule: for any theme, check the date distribution before concluding anything. A theme that is large but entirely historic and a theme that is small but entirely recent point in opposite directions.
Preserve the words
The most common analytical failure is summarising too early.
“Users find the interface confusing” has thrown away the finding. “I still can't find where to change a user's permissions without googling it, and I've been using this for eight months” contains the specific feature, the specific failure, the user's tenure, and the emotional register — all of which are usable.
Keep verbatim quotes with links to their sources. The phrasing is the asset: it feeds positioning copy, objection handling, and the vocabulary problem in search, where buyers describe their problem in words that rarely match your category label.
Reading a competitor's reviews
The same method, with two additions.
Look for structural complaints. A missing feature can be shipped next quarter. A complaint arising from their business model — support ratios implied by their pricing, limits imposed by their architecture — will persist. Structural weaknesses make durable proof points; incidental ones expire.
Look for the segment. Complaints often cluster by customer type. A competitor whose negative reviews come overwhelmingly from small teams is telling you where they stopped investing, and where you can win.
The bias you cannot remove
Worth stating plainly, because it bounds every conclusion drawn this way.
Review writers are not representative. They skew toward the vocal, the technical, the English-speaking, and the extremes. Enterprise buyers mostly do not write reviews. Satisfied users mostly do not write reviews.
This means review analysis can tell you what problems exist and roughly how they cluster. It cannot tell you what proportion of customers hold a view. Any output that implies a percentage is overreaching.
In thin categories the honest conclusion is sometimes “insufficient evidence” — and a report that manufactures four themes from six reviews is inventing rather than analysing.
Doing it repeatedly
Themes rising and falling over time is worth more than any single snapshot. Save the quotes, date them, and compare quarterly. A theme that is growing deserves far more attention than one that is merely large.
Our Social Media Audit works this way — verbatim quotes with source links rather than a sentiment score, themes counted and dated, and an explicit statement when the corpus is too thin to support a conclusion. The wider method is in listening where you don't own the channel.