This is likely related to the fact that the app uses both generator=prefixsearch and generator=search.

When the completion suggester is rolled out on Wednesday 16th March, the search results returned by prefixsearch should be significantly improved. At that point, it might be prudent to simply remove the generator=search API call, and just use prefixsearch, because prefixsearch will now come with things like spelling correction and page view powered result rankings. @Dbrant may also want to consider doing the same on Android.

On the Android side, we're actually no longer mixing the two queries. (i.e. this behavior doesn't happen in Android)

Need to decide if we can move away from mixing queries. @Dbrant how did you guys decide to "un-mix" of queries

@JMinor To be more precise, on Android we perform a single query with generator=prefixsearch, and then if that query doesn't return enough results, we follow it up with a full-text query (generator=search). And, if the latter query returns results already contained in the former, we remove them.

In T129916#2149354, @Dbrant wrote:

@JMinor To be more precise, on Android we perform a single query with generator=prefixsearch, and then if that query doesn't return enough results, we follow it up with a full-text query (generator=search). And, if the latter query returns results already contained in the former, we remove them.

The completion suggester (which powers prefixsearch) which was launched on 17th March has cut the zero results rate for non-automata by nearly half (source) and we've also seen a significant increase in clickthrough and overall user satisfaction (source). I'd recommend simply removing the full-text query completely; the data indicates that the completion suggester on its own is probably sufficient for users.

Per @Deskana above, lets prefer the new prefixsearch.

Moving back for estimation, as this is no longer just a bug fix, but a potential change in search logic.