One uploader, nine funnels, and the anatomy of a utility sold as a subscription.
A full-funnel analysis of ZenDocs from a live crawl of production, which then became the brief: 14 surfaces redesigned on one codified design system.
Counts from the production crawl; funnel readings from internal Mixpanel data. Confidential values are stated as direction and relative magnitude.
- The teardown: three archetypes power 40+ funnels; the shared email gate is the highest-leverage surface on the platform; Forms converts best because the page is the user's document.
- The redesign: 14 surfaces rebuilt on one codified system by me and two UX engineers I direct, ending token drift and systemizing the AI feature funnels.
- The proof plan: the gate experiment reads cleanly against the existing event map, with guardrails on retention and refunds.
A single-file task, sold as a relationship.
ZenDocs converts, edits, signs, and transcribes documents in the browser, monetized as trial-to-subscription. Dozens of intent-matched landing pages feed one upload-to-paywall spine. The tension is visible on the marketing surface itself: the homepage FAQ leads with "How do I cancel my subscription?" because most users arrive with one urgent file and transactional intent, and the business needs them to start a relationship.
Win the task in seconds, then convert the moment of completed value into a trial. The funnel's job is to sell the finished file back at the exact moment it exists.
Three archetypes, one spine, forty funnels.
Not evaluating software. Mid-errand.
Arrives from a query naming the exact task, uploads in seconds, never scrolls. Recognition, not persuasion, carries them to the gate.
Buys a trial to finish one job; early cancellation intent is high before the day 7 to 8 renewal. Hiding the exit produces chargebacks, not retention.
Six findings, ranked by leverage.
One shared modal sits inside every funnel's biggest drop, framed as marketing capture when its legitimate job is same-session file delivery.
The best-converting funnel is the one where the landing page is the user's actual document. The rest of the platform should learn from it, not just celebrate it.
Transcribe and Humanizer are the growth thesis and the least-systemized surfaces: right try-before-gate instincts, no shared archetype.
How to Cancel in every footer, plain refund answers, honest limits. The compliance posture is a retention strategy wearing a legal hat.
An off-token primary blue in production was the highest-severity visual finding. A system only pays back when surfaces converge on it.
The funnels are industrialized, the after is not. Day 7 to 8 is decided by whether the product earned a second task.
An analysis that stops at findings is a report. This one became the brief: 14 surfaces rebuilt on one system, with the style guide as the contract.
Fourteen surfaces, one system, days not weeks.
The home, the funnel archetype, three form landings, both directories, the AI tools, the help center, and the PDF editor, rebuilt as interactive HTML prototypes on a codified style guide: a fixed funnel section order, one glass recipe, and interaction rules that travel with every component. The full system is its own case study.
This is also how the work ships. I lead a team of two UX engineers who build my design work directly in code; I direct their design output and hold sign-off, so review happens against the system rather than against taste. The editor and dashboard are architected to absorb future features without rework, and the AI feature line runs on one shared funnel archetype: Transcriber, AI Humanizer, PDF Translator, Summarizer, Ask AI, and the roadmap behind them. The goal is a product that looks and feels like 2026, and is built so it still will when the next feature lands.
Every finding maps to a shipped pattern.
How we'll know it worked.
The unified email gate ships as a controlled A/B, replaced in place so paywallViewed stays a valid denominator.
Primary: gate-to-paywall progression. Guardrails: purchases, day 7 to 8 retention, refunds, support contacts. A gate that converts better but refunds worse is a loss.
Results roll up by archetype, not by page: one clean read validates the pattern for forty surfaces, which is the economic argument for the system.
What this project proved. The findings that mattered weren't the ones that criticized, they were the ones that specced. Every friction risk in Part I maps to a shipped pattern in Part II. Analysis earns its keep when it compiles.
Method: live crawl of zendocs.com, structured parsing of 13 pages, rendered screenshots at 1440px and 390px; in-flow readings from internal product context. Redesign: 14 interactive HTML prototypes plus a condensed UI style guide.