← Back

Tree

A content-discovery tool where people compose their own cross-format feeds from contextual concepts.

Project overview

Type: Product · 8-week 0→1 · web (responsive)
Project type: 0→1 · Content Distribution · Search & Personalization · Consumer
Role: Lead Product Designer · Opportunity Framing · UX Research · Experiment Design
Methods: Continuous discovery (opportunity solution tree) · JTBD interviews · A/B
Tools: dscout · Otter.ai · Figma + Dev Mode · Amplitude · Statsig
Case thesis: Designing a discovery tool where users compose their own cross-format feeds from contextual "concepts," filled by indexing and linking to the open web, giving people direct control over what they consume.

The context

Algorithmic feeds optimize for engagement, so a user scrolls through a lot of low-interest content to reach the occasional piece worth their time, with little say over the mix. Keyword search returns generic results that miss the specific slice of a topic a person cares about. Tree's bet was to let people build their own recurring feeds, called presets, from concepts (contextual themes that carry more than a keyword), across every format, including text, audio, video, and images.

The problem

A topic is not an interest. In a feed-relevance test, users could name a topic in seconds, but a feed built from that topic label tested as only 38% "worth my time," sitting at the level of the algorithmic feeds they said they wanted to leave (behavioral, feed-relevance test). When asked what they wanted within the topic, users listed a median of 5 contextual qualifiers (era, region, format, length, treatment, specific figures) that the topic label threw away. The value of the product lived in capturing that context, and a topic-level feed discarded it.

The goal

Let a user turn a fuzzy interest into a feed precise enough to return to, and keep it improving, measured by feed relevance, presets a user keeps and reuses, and retention, rather than by how many content formats the platform indexed.

Empathize — Users named a topic in seconds but wanted a specific slice of it, and a topic-level feed tested as generic as the algorithm

In this section: Research foundation · Key insights

Research foundation (method)

Phase 1 — Interviews (n=20, ~40 min, recruited via dscout, transcribed in Otter.ai): what people follow, where current feeds fail them, and how they describe an interest.
Phase 2 — Survey: ~1,093 panelists invited; 200 analyzable responses (18.3% response rate); ~138 completed every item (12.6% completion rate). Attitudinal percentages are computed on the 200 analyzable responses; question types (select-all vs single-select) are labeled per question in the appendix.
Phase 3 — Feed-relevance test (n=28): each user rated three feeds on the same interest — an algorithmic feed, a feed built from a topic label, and a feed built from the same topic plus their stated context — on a "worth my time" scale, presented blind and counterbalanced.
Phase 4 — Prototype pilot (Amplitude-instrumented, web, 75 users): behavior on the concept-builder and presets, including the setup A/B.

Key insights

1. The contextual slice is the interest, and the topic label loses it. Asked "what do you like?" users gave a topic; asked "what about it?" they produced several qualifiers that changed everything. Triangulation:

Behavioral: topic-only feeds rated 38% worth-my-time — level with the algorithmic feeds users were leaving — while the same topic plus the user's stated context rated 81% in the concept test.
Verbatim (P6, documentary viewer, history hobbyist) — coded: Topic underspecifies: "I don't want 'history,' I want the fall of specific empires told as long documentaries, not hot takes."

2. Users recognize a good concept faster than they can author one. People struggled to write down what they wanted from a blank prompt, then immediately confirmed or rejected a feed once they saw it. Recognition was fast; authoring from nothing was slow.

Observational (setup sessions): the median user wrote a vague first concept and only sharpened it after seeing example results.
Verbatim (P14, design-news reader) — coded: Recognition over authoring: "I couldn't have told you what I wanted, but the second I saw the feed I knew which half to throw out."

3. The supply has to be full from the first session. Users abandoned a new feed that returned thin results, regardless of how well it was configured, so the platform could not wait for user-generated content to accumulate.

Attitudinal (survey): the top reason respondents gave for dropping a new app was an empty or repetitive feed.

Dashboard — A topic is not an interest

A topic is not an interest
Scope: feed-relevance test (n=28) + survey (n=200)
Guiding question: How much does context change feed relevance?

  Feed rated "worth my time"
    Algorithmic feed ......................... ~37%
    Topic label only ......................... 38%
    Topic + user's stated context (concept) .. 81%

  Contextual qualifiers users named per topic .. 5 (median)

Key Insight: A topic label barely beats the algorithm; the relevance jump is
the context the topic discards, so the product's job is to capture that
context rather than the topic alone.

Define — The feed had to be built from the contextual slice of an interest, captured as a concept

In this section: POV · How Might We · Principles · Insight→decision map

POV statement. A person leaving algorithmic feeds needs to compose a feed from the specific contextual slice of a topic they care about, because a topic-level feed returns content as generic as the feed they left.

How Might We

How might we help a user turn a fuzzy interest into a concept precise enough to fill a good feed?
How might we capture contextual qualifiers (era, region, format, length, treatment) without making setup feel like work?
How might we keep a feed converging on the user's taste as they use it?

Design principles (each traceable to an insight)

The concept carries the context. A concept holds facets (subtopic, format, length, region, era, treatment), so the feed targets the slice the user means. (Insight 1)
Elicit, then tune by example. Setup proposes facets and lets the user sharpen by reacting to real results. (Insight 2)
The user owns the feed. Presets are transparent, editable, toggleable, and offer a chronological option. (the goal)
Supply from the open web, by reference. Tree indexes existing content and links to it on its source, so feeds are full on day one without re-hosting. (Insight 3 + rights model below)
Each format consumed at its best. A distraction-free reader, a native audio player, and full-bleed video present each format optimally.

The vocabulary, made precise (so the model is unambiguous)

A facet is one contextual qualifier (e.g., format = documentary, length = long, era = antiquity).
A concept is a bundle of facets on a subtopic (e.g., "the fall of specific empires, as long documentaries").
A preset is one or more concepts saved together as a reusable feed (e.g., a "long-form history" preset).
The active feed is the sum of whichever presets the user has toggled on.

Insight → decision map

Insight (from Empathize)	Concrete design decision
Topic-only feeds rated 38%; context lifts to 81%	A concept is built from facets, so the feed targets the contextual slice, and several concepts compose one preset
Users recognize a good concept faster than they author one	Setup seeds from a sentence or an example piece, proposes concept facets, and tunes via "more / less like this" on real results
New feeds die when results are thin	The web-indexing layer fills any new concept immediately, linking to each source, so a first preset is never empty

Ideate & Craft — A concept builder elicited the facets of an interest and tuned the feed by example

In this section: Design execution · Supply model & rights · Before → after · Other deliverables

Design execution

Concept builder — a user seeds a concept with a sentence or an example piece they like; Tree proposes facets (subtopic, format, length, region, era, treatment) drawn from what it indexed, and the user accepts, edits, or adds. The feed updates live as facets change.
Presets — a preset holds one or more concepts; toggling presets reshapes the active feed, so a person can keep separate feeds for basketball, editorial design, or long-form history and switch between them.
Example-based tuning — "more like this" and "less like this" on real items converge the feed on the user's taste over a few sessions.
Filters and order — feeds filter by location, format, and duration, and offer a chronological view alongside the recommended one.
Optimal per-format consumption — text opens in a distraction-free reader, audio in a native player, video full-bleed, and places and events render with map and time context.

Supply model & rights (the load-bearing constraint, made explicit)

Tree is a discovery and link layer, not a content host — the same posture as a search engine. It indexes content that lives on its original servers, shows a title, source, and snippet, and sends the user to the origin to consume it, crediting the creator. It respects robots.txt and publisher opt-outs, and for audio and video it relies on official APIs and embeds (which carry their own playback and attribution terms) rather than scraping or re-hosting media. This is what makes "index the open web" defensible: attribution plus link-out plus official embeds is a referral relationship, not republication. The design treats the source link and credit as a first-class, non-removable part of every item — a rights requirement that doubles as a trust signal.

Before → after

	Before (algorithmic feed / keyword search)	After (Tree)
What shapes the feed	An engagement algorithm, or a single keyword	Concepts the user composes from facets
Specificity	Topic level	The contextual slice the user means
Supply	One platform's user-generated content	The open web, indexed and linked, with attribution
Consumption	One format, one surface	Each format presented at its best

Other deliverables

Built in Figma with Dev Mode handoff: the concept-builder facet system, the example-based tuning pattern, the preset switcher, the per-format consumption views, and the source-attribution treatment for indexed content.

Dashboard — What the concept builder produced (Variant B cohort, in-product)

What the concept builder produced
Scope: Last 30 days · web · pilot builder cohort (Variant B)
Guiding question: What did the seed-and-refine builder deliver in use?

  Feed relevance ("worth my time") ...... 82%
  Saves or opens per 10 items shown ..... 4.3
  Presets kept and reused at 7 days ..... 81%

Note: these are the builder cohort's absolute results. The A/B section below
compares them head-to-head against the one-shot freeform variant.

Prototype / Test — A one-shot "describe what you want" box was fast and produced feeds users abandoned; a seed-and-refine loop produced feeds they kept

In this section: The experiment · What it taught

The most appealing first idea was a single freeform box: the user describes what they want in a sentence and Tree builds the feed. It felt flexible and modern — and it is the naive-AI approach many products ship by default — so it was A/B tested against the seed-and-refine concept builder in Statsig across the pilot.

The failed variant. The freeform box was the fastest path to a finished feed, but a single sentence underspecified the interest (users wrote vague descriptions and did not know how much context the system needed), so feed relevance landed at 44% — only a few points above a bare topic label — and 63% of those feeds were abandoned or edited into the ground within a week. Capturing taste needed more than one sentence could hold.

One sentence is fast and underspecifies the feed
Scope: Statsig A/B · web · pilot (75 users; ~37 / arm)
Guiding question: Which setup produces a feed the user keeps?

  Variant A — One-shot freeform description
    Time to a finished feed ......... fastest
    Feed relevance ("worth my time") 44%
    Feeds kept at 7 days ............ 37%

  Variant B — Seed + facet refinement + example tuning
    Time to a finished feed ......... +90 sec
    Feed relevance ("worth my time") 82%
    Feeds kept at 7 days ............ 81%

Read with care: ~37 users per arm. The relevance and 7-day "kept" gaps are
large and consistent; the 30-day retention figure (below) is directional.

Reconciliation: the freeform variant (44%) edges a bare topic label (38%)
because a sentence carries some context; the builder (82%) lands on the
concept test's prediction (81%) — two methods agreeing.

Key Insight: The extra ninety seconds of refinement is what turned a feed
the user dropped into one they returned to. Variant B shipped.

What it taught. A blank natural-language box looks flexible and captures little, because users cannot state intent they have not yet seen results for; taste is captured through a seed plus an example-driven loop that converges. The seed-and-refine builder shipped.

Outcomes & reflections

In this section: Causal chain · Limitations · Competitive context · Reflections · References

Causal chain (pilot, 75 users; the comparison is the shipped builder vs the rejected freeform variant)

The numbers below compare the two pilot cohorts: Variant A (the one-shot freeform box — the naive-AI baseline) against Variant B (the concept builder that shipped). They are not a product-vs-world comparison; the algorithmic/topic-only baseline was measured separately in the relevance test (~38%).

Within that A/B, the concept builder elicited contextual facets and tuned by example, so feed relevance rose from 44% (freeform) to 82% (builder) "worth my time," which got users saving and opening more of what they saw (saves or opens per 10 items 1.6 → 4.3) and returning to their feeds, which lifted weekly active presets per user from 1.4 → 3.2 and day-30 retention from 41% → 60% (directional, given ~37 per arm). Across both cohorts, the web-indexing supply meant a new user's first preset returned a full feed immediately, removing the empty-platform start that ends most new feeds.

Metric	Freeform (Variant A)	Builder (Variant B)	Δ
Feed relevance ("worth my time")	44%	82%	+38 pts
Saves or opens per 10 items shown	1.6	4.3	~2.7×
Presets kept and reused at 7 days	37%	81%	+44 pts
Weekly active presets per user	1.4	3.2	~2.3×
Day-30 retention (directional)	41%	60%	+19 pts

Reading note: the win here is that a considered design beat the naive one-sentence box, not that Tree beat the open algorithmic feed. The thesis that concepts beat topics is carried separately by the relevance test (38% → 81%).

Scale note: in discovery, relevance compounds into habit — a feed worth returning to becomes a daily surface, where a generic one is abandoned after a session.

Limitations (stated, because a portfolio claim is only as strong as what it concedes)

Small pilot. 75 users, ~37 per A/B arm. The relevance and 7-day "kept" gaps are large and consistent; the 30-day retention lift is directional, not statistically powered.
Two baselines, two methods. The thesis (38% topic vs 81% concept) is self-reported in the relevance test; the A/B (44% vs 82%) is in-product behavior. They agree, which strengthens the claim, but neither is a long-horizon measurement.
Supply depends on index coverage and rights. Day-one fullness assumes the index is broad and that source APIs/embeds remain available; a publisher opt-out or an API term change narrows supply for the affected formats.
Curator layer is out of core scope. Letting people subscribe to feeds others compose adds a social/moderation surface — the exact axis that complicated comparable products — and is deliberately deferred, not designed here.
Interest data is sensitive. A user's concepts and presets are a precise map of their interests; the product needs an explicit stance on what is stored, shared, and retained.

Competitive context

Discovery is a crowded space with a notable graveyard. Artifact — the AI-driven news-discovery app from Instagram's co-founders — wound down in 2024 after concluding the market opportunity wasn't large enough and after drifting from a reader into a social network with moderation overhead. Feedly, Google News, Flipboard, Ground News, and RSS readers occupy adjacent ground. Tree's wedge is user-composed control: the feed is built and owned by the user from explicit, editable concepts, across formats, sourced by reference from the open web — not an opaque algorithm and not a social network. The lesson taken directly from Artifact's arc is to keep the social/curator layer out of the core and let the value sit in control and relevance, not in a feed of other people to moderate.

Reflections (transferable principles)

In discovery, the unit of value is the contextual slice of an interest, and the design's job is to elicit that context, because users recognize a precise concept faster than they can author one from a blank prompt.
Capturing taste takes a seed plus an example-driven refinement loop; a single freeform description looks flexible and underspecifies what the system needs to return a feed worth keeping. (The A/B is the evidence: the naive box lost to the considered one.)
Indexing existing web content by reference gives a new platform full supply from its first session, solving the cold-start problem that empties feeds before a user base exists — and treating the source link and credit as non-removable is both the rights model and a trust signal.

References (for the external claims)

Artifact, the AI news-discovery app by Instagram's co-founders, announced wind-down in 2024, citing insufficient market opportunity; coverage notes its drift toward a social network and moderation overhead (TechCrunch, Engadget, The Verge, 2024).
Search-engine-style indexing (index, snippet, link to source) versus re-hosting, and the role of robots.txt and official media APIs/embeds, frames the rights posture of the supply model.