Take an app with a 4.2 star average. Is that good? It depends on what sits under it. The same 4.2 can come from a product people like with one bug that makes a small group furious. It can come from a product nobody loves and nobody hates. It can also come from a product that was bad for months and has just been fixed. The average looks the same in all three cases. The work you should do next is completely different in each.
That is the core problem with a star rating. It is a summary, and a summary is built to throw information away. This article explains what gets thrown away, how aspect sentiment gets some of it back, and where aspect sentiment has limits of its own. At the end we connect the idea to the disclosure data this site covers, because the same trap shows up there.
What a star rating actually measures
A star rating is one number per review. The reviewer squeezes their whole experience into it. That experience might include the price, the login flow, a crash on launch, a helpful support reply and a feature they miss. The number does not say which of those drove the score. Two people can both give three stars for opposite reasons. One likes the features and hates the ads. The other has no problem with ads but finds the interface confusing.
Averaging makes this worse. Once you average many reviews, the reasons are gone for good. You cannot get them back from 4.2. You can only guess, and guesses tend to follow whatever the team already believed.
There is a second issue. The platform decides how the headline number is built. Apple, for example, lets developers choose to reset the summary rating when they release a new version, as described on its ratings and reviews page for developers. That is a sensible feature. It also means the headline number is partly a product decision, not only a measure of how users feel. Comparing two apps by their headline rating assumes both numbers were built the same way. Often they were not.
Why the average hides what to fix
A product team needs an answer to a simple question. What should we change first? A star average gives no ranking of problems. It gives one grade for the whole thing.
Consider three patterns that can all produce the same average.
- Most reviews are positive, and a small cluster is very negative about one specific thing, such as billing. The fix is narrow and high value.
- Reviews are lukewarm across the board. No single aspect is broken. The product is simply unremarkable. The fix is strategic, not technical.
- Reviews were poor for a long stretch, and recent reviews are strong after a release. The average is dragged down by history. The right move may be to do nothing and wait.
If you only look at the average, you might ship a redesign in case 1, chase bugs in case 2, and panic in case 3. Each of those is the wrong response.
What aspect sentiment does
Aspect sentiment splits the review stream by topic before it summarises anything. Instead of one grade, you get one grade per aspect. Typical aspects for apps are stability, performance, user interface, price and billing, login, ads, features, support, updates and content.
The mechanics are simple enough to explain in a paragraph. For each review, detect which aspects it mentions. A review that talks about a crash and a price increase counts toward both stability and price. Then, for each aspect, summarise the reviews that mention it. One transparent way to do that is to take the average star rating of those reviews. Our own App Store Review Intelligence tool works this way. The benefit of using the star rating as the sentiment score is that it works in any language. You still need to detect the aspect in the text, but you do not need a separate sentiment model for each language.
The result is a profile instead of a grade. A 4.2 app might score well on features and interface, and poorly on billing. Now the question "what should we fix first" has an answer you can defend.
Three things the profile adds
The first is ranking. Once each aspect has its own score and its own volume, you can sort problems. An aspect with a low score and many mentions matters more than an aspect with a low score and two mentions. A star average never lets you make that trade.
The second is change over time. Aspect scores can move while the headline stays flat. If sentiment about login drops over recent weeks, that is an early warning even if the overall rating has not moved yet. The overall number is slow because it carries all past reviews. An aspect score over a recent window is faster.
The third is comparison. You can read reviews for apps you do not own. Put your profile next to a competitor's profile and you see where each is weak, aspect by aspect. "They are rated higher" is not useful. "They are rated higher on price and lower on stability" is.
Where aspect sentiment can mislead
Aspect sentiment is better than a single grade. It is not the truth. Be honest about its limits before you act on it.
Keyword detection misses things. A review can complain about billing without using any word your detector looks for. Sarcasm and mixed feelings are hard. A review that praises the interface and gives one star because of a crash will pull the interface score down unfairly if you score aspects by the review's star rating. Our method accepts that trade because it is transparent and language independent. More advanced methods score each sentence separately. They are also harder to audit.
Small counts are noisy. An aspect with a handful of mentions can swing hard on one angry review. Always show the count next to the score. If you do not see a count, do not trust the score.
The sample is not the user base. People who write reviews are not a random slice of users. Very happy and very unhappy people are more likely to write. Public review feeds also limit how far back you can read. Apple's public review feed only returns the most recent reviews for an app, so history you did not collect at the time may simply be gone.
Reviews can be fake. The US Federal Trade Commission adopted a rule banning fake reviews and testimonials in 2024. A ban is not the same as a clean dataset. Treat sudden bursts of very similar reviews with suspicion, whatever direction they point.
Free ways to start
You do not need to buy anything to try this. Read the last few pages of your own reviews in App Store Connect or the Google Play Console. Tag each review by hand with the aspects it mentions. Count and average per tag in a spreadsheet. A few dozen reviews are often enough to see which aspect dominates the complaints. Open source text tools can automate the tagging once you know which aspects you care about. A paid tool mainly saves time and keeps a history you would otherwise lose.
The same trap in disclosure data
This site mostly covers official SEC and House disclosures. The single number problem shows up there too.
Take institutional holdings. Many funds report their positions each quarter on Form 13F, and you can read the raw filings on SEC EDGAR. A popular summary is a count: how many tracked funds hold a stock. That count is an average in disguise. It hides whether the funds are adding or trimming, whether the position is large or tiny for each fund, and whether the count rose because of new buyers or because an old holder stayed put. Two stocks with the same count can mean opposite things. The filings also arrive with a delay, which we cover in 13F deadlines and the 45-day lag.
Congressional trading has the same shape. A headline like "members bought more than they sold" is one number built from many filings. The detail lives in who traded, what size range they reported, and how late the report came in. The raw reports are on the House Clerk's financial disclosure site. We explain the reporting flow in how congressional trading disclosures work, and the timing rule in the 45-day rule and why it matters.
The lesson carries over directly. Split the number into its parts before you summarise. Show the count behind every score. Watch the parts that move while the headline sits still. Nothing in this article is investment advice. It is a method for reading data more carefully.
Takeaway
A star rating tells you how a product is doing overall. It cannot tell you why, and it cannot tell you what to fix first. Aspect sentiment gives each topic its own score and its own count, so you can rank problems, spot early changes and compare against competitors. It has limits. Detection misses things, small counts are noisy, and review writers are not all users. Used with those limits in mind, it turns a grade into a to-do list.
If you want the same breakdown approach applied to institutional holdings, the Smart-Money 13F Consensus report splits fund positioning into who holds, who added and who left, so a single consensus number does not have to stand on its own.
Want the signal instead of the raw filings? Get a free report preview. Prefer the tool to the write-up? Browse all data feeds or connect the free MCP server.