Search & Discovery That Speaks Music

The search problem every music licensing platform has — and most haven't solved

Soundstripe serves video creators, brand teams, and agencies who need licensed music fast. Their library is large and high-quality. But there's a fundamental mismatch at the heart of music search: users think in feelings and references ("something cinematic, like a Nintendo ad"), and search engines understand tags and metadata ("orchestral, 92 BPM, C major").

The gap between those two languages was killing conversions. Users would search, get results that felt slightly off, search again with different words, get frustrated, and leave. Downloads — the primary proxy for subscription intent — were low. Dead-end sessions were high.

Two compounding gaps between how users search and how music gets found

Gap one: the language problem

Music licensing users are often video editors, brand creatives, and agency producers — people who think in visual and emotional language, not musical terminology. When they search "upbeat background music," they could mean a dozen different things. When they search "sounds like a Nike ad," the old system had nothing to work with.

• Search queries were vague by nature — users were describing a feeling or a reference, not a track
• Each dead-end result set required starting over from scratch — no way to refine in context
• Sessions that didn't find the right track quickly tended to end without a download

Standard search experience: keyword-based, no refinement, no context. One shot per query. Search often returns thin or no results due to challenges matching natural-language queries to song metadata tags.

Gap two: the invisible metadata problem

Soundstripe had invested heavily in rich metadata for every track — tempo, key, mood tags, instrumentation, use-case labels. This metadata was already powering the filter system and search ranking. But it was almost entirely invisible to users while browsing.

The only way to see a track's full metadata was to navigate to its individual song page — a separate destination that almost no one visited. Users in the browsing flow, scrolling through results, had no way to understand why a particular track matched their search, or what made it different from the track next to it.

• Rich metadata existed — users just couldn't see it without leaving the browsing context
• Users couldn't learn the vocabulary of what they liked — no feedback loop between results and search terms
• Two tracks that sounded similar on the surface had no visible differentiation in the list view

Gap two: rich metadata lived behind each track but was invisible in the browsing flow. No way to see why a track matched or how it differed from the one next to it.

Sole designer on both features, end to end

I designed both the AI search experience and the metadata slide-out feature independently, working with a PM, three engineers, and data/analytics throughout. A second designer provided peer review. These weren't exploratory concepts — both shipped to production.

Two features. One coherent system.

Feature 1 — AI search: intelligence through conversation, not just queries

The first design decision was where the AI should live. One path was to keep it buried in a search bar — a slightly smarter autocomplete. I pushed in the opposite direction: give the AI a presence and let it guide the search through dialogue.

The result was Supe — a named AI music supervisor embedded directly in the product. Rather than asking users to find the right search terms, Supe asks clarifying questions, offers structured choices, and refines results conversationally. The intelligence is visible in the exchange, not just the output.

• Natural language prompts ("sounds like a Nintendo ad, but more cinematic") kicked off a real dialogue — Supe asked follow-up questions to narrow the vibe before surfacing results
• Suggested response chips reduced friction for users who weren't sure how to describe what they wanted
• Each clarifying turn refined the result set — so the list of tracks updated as the conversation got more specific
• The interface had personality (a name, an avatar, an intro message) without feeling like a gimmick — it was still in service of finding the right music

Natural language in, curated results out. The intelligence is in the results — the interface stays familiar and out of the way.

The compound refinement pattern

After each result set, the interface surfaced three AI-generated follow-up prompts — short suggestions for how to refine the current search ("fast, urgent, high energy," "mid tempo, heroic, soaring," "slow build, big ending"). These weren't new searches — they were extensions of the current one, building specificity progressively.

This solved the dead-end problem directly. Instead of a failed search meaning start over, it meant a suggested next step. The refinement prompts gave users a path forward without requiring them to already know what to say next.

Feature 2 — Metadata slide-out: making the invisible visible

The metadata slide-out was designed around a simple insight: users needed to see what made a track tick while they were listening to it, not after clicking away to a separate page. The solution was an inline expansion — a lightweight panel that appeared beside the track row in the results list, showing mood tags, use-case labels, tempo, key, and instrumentation without disrupting the browsing flow.

• The panel opened as a drawer from the right edge of the screen — the track kept playing, the list stayed intact, nothing navigated away
• Technical metadata (duration, BPM, key) appeared at the top alongside catalog info — useful at a glance for editors and music supervisors who need it fast
• Richer contextual data followed below — genres, instruments, moods, characteristics, and keywords — giving users multiple vocabularies to understand a track
• Tags were clickable, turning any label into a filter — so a mood or instrument that fit could immediately seed the next search

The metadata slide-out: inline, lightweight, and anchored to the track — everything visible without leaving the results list.

Why these two features work together

The AI search and metadata slide-out solve different parts of the same problem. AI search meets users at their language and gets them to good results faster. The metadata slide-out helps them understand why a track is the right one — and teaches them the vocabulary to search better next time. Together, they close the loop between how users think about music and how the system organizes it.

Both features moved the metric that matters.

Downloads are Soundstripe's primary indicator of subscription intent — a user who downloads a track is significantly more likely to convert to a paid plan. Both features directly moved this number.

What I'd do next

The natural next step would be personalization — using download history, saved tracks, and search patterns to proactively surface music that matches a user's taste profile before they even search. The metadata system was already rich enough to support this. The AI search behavior and the slide-out engagement data would give the personalization model strong signal. The infrastructure was there; it just needed the next design layer on top of it.