Gmail Search AI Research
A PM research artifact on Gmail's AI features — three product hypotheses (Accuracy, Action Gap, Voice) tested against public Reddit feedback.
The problem
Gmail ships Gemini AI features — summaries, smart replies, search — but a PM on the team has to pick which bets get the next sprint. Is the biggest lever summary accuracy? Is it closing the gap between reading an email and doing something about it? Or is it making AI drafts sound like the person sending them? Each answer points to a different roadmap.
My hypothesis
If I pulled Reddit feedback on Gmail's AI features and categorized each post by three working hypotheses — accuracy (is the summary right?), action-gap (can I do something with it?), and voice (does it sound like me?) — the distribution would point to where the product has headroom.
What I built
A research dashboard at schlacter.me/gmail-search-ai with three hypothesis tracks: Accuracy, Action Gap, and Voice. Each hypothesis has its own tagline, frequency count, competitor signal, and a set of real Reddit quotes backing it. A methodology page documents how posts were sourced and classified.
What broke
Keyword-based classification is a blunt tool — I had to re-read borderline posts to decide which hypothesis they really belonged to. The corpus also skews toward users who had a problem; a real PM signal would need to pair this with Gmail's internal usage data to know which hypothesis maps to the biggest user segment.
What I learned
Picking your hypotheses before you tag the posts is the right sequencing. It makes you commit to a product theory, then test whether the data actually fills those bins. If one hypothesis pulls 5% of signal, that's as useful as one pulling 50% — the distribution is the insight.
If I kept going
Deepen the competitor comparison (Superhuman, Copilot, Notion AI) — each has a distinctive take on one of the three hypotheses. Add a fourth hypothesis if the data suggests one the original three miss. Build a sketch prototype for whichever hypothesis has the biggest headroom.
Prototype
Dashboard hero: three working hypotheses — Accuracy, Action Gap, Voice.
Each hypothesis has a tagline, frequency count, and the real Reddit quotes backing it.
Competitor signal per hypothesis — what Superhuman, Copilot, Notion AI do differently.