Review mining

Store text is written in category language: “comprehensive course content”. Reviews are written in outcome language: “to learn sign language”. The second group is far closer to what gets typed into the search box.

When we ran this on a language-learning app, most of the keywords in its cusp band came from this channel. A title scan would have found none of them.

Keywords are not mined from low-star reviews; the language there is complaint language. But when the same complaint repeats across several rivals, it is no longer a bug - it is a category gap.

That is why 1-2 star reviews go to a separate screen: unmet needs.

You say "personal finance management"; the user searches "where is my money going". You say "language learning platform"; they search "10 minutes english a day". Those phrases are not in your metadata, and they are not in your rivals' metadata either - but they are written down in your rivals' reviews.

The reason is simple: someone writing a review has to describe the product, and they do it in their own words. For a phrase invented in a marketing meeting you cannot even say "at least one real person used this"; for every phrase in a review, you can.

Competition on that source is low because no keyword tool looks there. Everyone's pool feeds from the same three places, and reviews are left outside all of them.

The most productive source is the rival that most resembles you but is larger than you: the review volume is higher, so the extraction is statistically sounder. Bad reviews work too - a missing feature is the word of the person looking for that feature.

Reviews are made of sentences and keywords are made of phrases. There is a conversion between them, and done carelessly it produces rubbish.

The most important rule is pronouns. People write "it really helped me" and "it genuinely surprised me". Those are grammatically sound and they occur often, but nobody searches them on the App Store. Phrases containing first- and second-person pronouns are dropped for that reason.

The second rule is brand names. A rival's own name recurs constantly in its reviews and surfaces as the most frequent phrase in any extraction. A brand name is a keyword, but a rival's brand is no use to you; those repeats are dropped.

The third rule is word boundaries. In writing systems that do not separate words with spaces - Chinese, Japanese - where a word ends is not marked by a space. A separate boundary-finding step runs for those languages; without it the extracted phrases are meaningless.

The fourth is letter folding. A phrase is lower-cased on the way into the pool, and the rule depends on the country: in Turkish and Azerbaijani a capital I becomes a dotless ı, everywhere else a dotted i. Applying one rule everywhere corrupts the phrase, and the corrupted phrase matches nobody searching for it.

Coming out of a review does not privilege a phrase. What survives extraction enters the pool and is treated like everything else: its popularity is measured, its difficulty computed, your current rank for it searched.

Some of them turn out to be genuinely searched, some not at all. What separates them is the measurement, not the source. That is why we do not sell review mining as a "hidden keyword finder" - what it does is add a channel nobody looks at to the pool.

The channel's value is exactly there: among phrases that pass the same elimination, the ones most likely to be lightly contested come from here. Because your rivals never saw those words.

Phrases mined from an app with low review volume stay statistically weak. To get value from this channel you need to look at the few apps in the category with the most reviews.

The phrases mined from the same app's reviews in two different countries do not overlap. Local equivalents, everyday usage and what the product is actually used for in that market all differ.

That is what makes review mining most productive when it is paired with country selection: scanning the reviews of that market's rivals before opening a new country decides, almost on its own, what metadata you should go in with.

The reverse holds too: copying phrases mined in one country into another country's metadata spends that country's 100 characters for nothing.

The scan runs monthly or quarterly. Rank measurement comes free - it is read out of the same search - but reading reviews needs a separate request per app, so this channel runs less often.