Finding keywords in your reviews without collecting chatter

Reviews use the words people search with — and many more that nobody searches. Most of the work is filtering.

Why reviews are worth reading

Users describe an app in their own words. Those words often differ from how the listing was written, and some never show up in autocomplete because too few people type them for Play to suggest them. A phrase that several different reviewers use to name what the app does is a strong keyword candidate.

Most of what you extract is not a search term

Reviews are sentences, not queries. A naive extraction counts every frequent word and phrase and returns praise, complaints, verbs and filler. The first time we ran it, 154 of 203 extracted terms were single words, and almost none of them were things anyone would type into a store search box.

Filters that remove most of it

• Count reviews, not mentions: a term has to appear in at least two different reviews, not twice in one.
• Let a single word through only if it already appears in your own listing, with root matching. This rule needs no word list, so it works in any language.
• Drop phrases that start or end with an evaluative word; keep the same word when it sits inside a real product phrase.
• Leave out reviews below three stars. There people describe problems rather than the product, and you end up with a vocabulary of complaints.
• At equal frequency, prefer the longer phrase: a two-word product phrase says more than a single word.

Then measure

A review term is a hypothesis. Add the ones that survive to your pool and let them be measured like any other keyword — rank, popularity, difficulty. Keep those that are searched and describe your app. None of this requires storing the review text anywhere.