A research repository that researchers actually open

The problem we were handed

Fieldnote sells a research repository to B2B product teams. The product had strong first-month engagement, new teams would pile studies in, tag things, share clips, and then usage would fall off a cliff. Accounts stayed subscribed because the research inside was valuable, but the tool itself was increasingly ignored. The churn risk was three quarters out, not immediate, but the revenue retention math did not work unless active usage recovered.

The ask was broad. “Why are researchers not coming back, and what should we build to fix it?” I had nine weeks and a team of four.

What research showed

The repository was full. Researchers were storing studies in it. But nobody was searching it.

Instead, researchers were asking each other in Slack. “Did anyone run something with enterprise admins recently?” would get answered by the one teammate who remembered. If that teammate was on leave, the research was, functionally, gone.

The repository was indexed by study. Studies were tagged, searchable, and browsable. But a researcher does not want a study. A researcher wants the thirty seconds in the study where the participant said something specific about enterprise admins. That thirty seconds lived inside a two-hour video, inside a session, inside a study, inside the repository. The tool made you dig through four layers to find it.

The core decision

We changed what the repository returns. Instead of returning studies and letting the researcher drill in, it returns clips, thirty-second annotated moments, with the study, session, and participant attached as metadata. The clip is the first thing a researcher sees, and it is the thing they can paste into a Slack thread or a deck.

This sounds small. It was not. It required re-indexing every existing study, rewriting the search backend, rebuilding the browse UI, and changing the mental model every existing user already had. We tested it in a private beta with eighteen accounts before rolling it out.

Design detail that mattered

Three details moved usability more than I expected.

The hover preview. Twelve seconds of video, muted by default, auto-play on hover. That single affordance cut the time researchers spent opening dead-end clips by more than half.

The participant demographic chips. Every clip shows three small chips beneath the transcript snippet, industry, role, seniority. Researchers could filter on those without ever opening a clip. We had not planned to add the chips until post-launch; researchers asked for them unprompted during the private beta and we added them in week seven.

The transcript snippet. Showing the exact sentence, not a generic preview, turned out to be the single most requested feature during testing. Researchers treat transcripts as evidence. They want to see the exact words.

What did not work

An early prototype tried to generate automatic summaries of each clip using an LLM. Researchers hated it. The summaries were usually accurate but lost the specificity that made the clip useful. We cut the feature entirely. It turns out that a summary is a different thing than a quote, and researchers were searching for quotes.

We also tried a “related clips” panel on the clip detail view. It was ignored. Researchers who found a good clip were done, they did not want adjacent clips, they wanted to get out of the tool and into their deck. We shipped without it.

Outcomes

Monthly active researchers tripled within six weeks. Time-to-first-relevant-clip dropped from twelve seconds to four on the standard task-based benchmark. Expansion revenue, seats added to existing paying accounts, grew 24% in the quarter following launch, mostly because teams who had mothballed the tool reactivated it and added their new hires.

Reflection

The retrieval model was doing a different thing than the research workflow. Fixing the mismatch was worth more than any new feature we could have added. Research tools are the clearest case I know of where the model in the product must match the model in the researcher’s head. Get it wrong and no amount of polish will bring them back.