Results
In March 2026, ElevenLabs was buying Google, Meta, and TikTok traffic for three business lines and landing it on pages nobody owned. Over nine months I turned paid landing pages into an owned program, then into an engine where Claude designs, builds, tests, and maintains pages in production.
168 pages, 88 tests, 52 wins
168 landing pages published from April to November 2026 across the Creator, Agents, and Developer lines
88 concluded tests: 52 variant wins (59%), 25 control wins, and 11 draws, with a 32% median lift among winners
A variant only wins when it beats control on both subscriptions and est. MRR per visitor
32 of 38 head-to-heads won against the marketing site
38 randomized paid-traffic tests put a dedicated landing page against the page campaigns used before; the landing page won 32 (84%), with a 40% median lift
Studio: +18.5% sign-up CVR and 170 vs 115 subscriptions, the largest single Q3 contributor
Enterprise ICPs: +323% contact sales CVR and two closed-won deals worth $39,600 ARR, against none on control
Developer API (Dark): +35.5% API key creation; the dark, code-first pattern became the API template
About $618K gross incremental ARR run-rate
$274K in Q2 and $344K in Q3 (tracker estimates), with the Q2 target met
Sign-up to subscription reached 6.9% at the Q3 close against a 6.5% target, with the Core Experience team
Weeks to a day
A page used to take weeks; a handful now ship in a day from one prompt loop
Figma is no longer in the production path
The Problem
Paid campaigns landed on generic marketing-site pages, agency-built Webflow pages, or docs. join.elevenlabs.io took its first paid click on 19 March 2026; before that, every click landed somewhere that wasn’t built for it.
No message match
A “speech to text” ad and an “audio to text” ad arrived at the same hero
Campaigns pointed at the same elevenlabs.io URLs whatever the query, the audience, or the model being advertised
Weeks per page
Every new page meant Figma design, a developer or agency build, manual QA, and a readout written by hand
Paid media launches weekly, so campaigns kept launching without a page made for them
No revenue read
Measurement was split across VWO, GTM, and PostHog with no revenue tie-in
Wins were argued on sign-ups rather than money, and nobody could say what a page was worth
The opening signal
A hand-run Expressive Mode hero test on 9 March, three headlines against each other before any testing tool was wired up, lifted sign-ups 13.6%. It was the first evidence that the paid page itself was a lever worth owning.
The Idea
Own paid conversion end to end, prove it with randomized tests judged on revenue, and only automate steps that had already been done by hand many times.
One owner, four jobs
Build: an LP design system from the logged-out site and the logged-in product, then a pipeline that ships pages without a hand-off to a developer or an agency
Test: every new page runs head-to-head against the page it replaces, on live paid traffic
Measure: subscriptions and est. MRR per visitor, never sign-ups alone
Feed back: winners get 100% of their campaign traffic, and learnings go to the marketing site, the product, and a playbook
Rules set before the first test
A split between sign-ups and revenue is a draw, whatever the sign-up lift says
Sample ratio, capture gaps, and UTM breaks are checked before a p-value is read
Approved figures, banned words, and claims wording are checked before copy is written, not after a page is live
No new vendors
Figma, Framer, VWO, PostHog, Hex, BigQuery, and Claude were already paid for: no procurement, no agency retainer
The Action
Four phases, each one useful on its own. Automation only replaced steps that had already been done by hand many times.
What changed for one page
Before, a page had hands on every step and took weeks. Now it’s prompt, review, ship: briefs arrive through the Asana intake, strategy comes from what already wins, and the readout, trackers, and Slack post close out in one workflow.
Cadences now running
Hourly: LP Autopilot, from Asana ticket to Framer branch
Daily: a Hex CVR exception report and tracker sync
Weekly: at least one new test launched
Monthly: a Framer content sweep of claims, models, stats, pricing, tables, and imagery
Before & After
Six of the head-to-heads. Controls are the marketing site or the legacy page; variants are join. pages built with the design system.
Studio: +18.5% sign-up CVR
Three-word promise, gradient headline, single CTA, and proof above the fold. Bounce fell from 41% to 21%, and subscriptions rose from 115 to 170.
ControlVariant · WonDeveloper API: +35.5% API key creation
A dark, code-first page with a runnable snippet and one CTA, “Get your free API key”. Dark beat light with theme as the only difference, then beat five more controls.
Control
Variant · WonSound Effects: 17.2% vs 12.3% sign-up CVR
The marketing page asked visitors to try the tool before they knew why; the landing page showed the result first, with a review and three free-plan reassurances. Bounce 32% vs 60%.
Control
Variant · WonEnterprise: +323% contact sales CVR
An inline demo form instead of a redirect, with a customer quote and the award badges. Two deals closed from the landing page, $39,600 ARR, against none from the control.
Control
Variant · WonAgents: +25% sign-ups
A “choose your path” fork sends developers to self-serve and enterprise teams to a demo, with the decision in the hero. 514 vs 409 sign-ups and 22 vs 17 contact sales, the first test in the program with real contact-sales volume.
Control
Variant · WonSeedance 2.5: +143% on the primary goal
Script-to-finished-video promise, model name in the hero, prompt and example in one frame. Est. MRR $1,863 vs $650; a retest in October confirmed it, and it became the base for pages the engine built next.
Control
Variant · WonHow a result is read
Every readout carries the results table, an executive summary, and both screenshots. The verdict is called on subscriptions and est. MRR, not sign-ups: in the Agents test the control won on subscriptions, but the landing page won on est. MRR and contact sales.
| Metric | Controlelevenlabs.io/studio | Variantjoin.elevenlabs.io/studio |
|---|---|---|
| Signed up | 1,367 | 1,645 |
| Subscription created | 115 | 170 |
| Starter subscription | 51 | 68 |
| Creator subscription | 62 | 102 |
| Pro subscription | 2 | 0 |
| Contact sales | 1 | 2 |
| Signed up → subscription | 8.18% | 10.39% |
| Bounce rate | 41.43% | 21.42% |
| Metric | Controlelevenlabs.io/agents | Testjoin.elevenlabs.io/agents/v6 |
|---|---|---|
| Signed up | 409 (0.28%) | 514 (0.35%) |
| Subscription created | 67 (0.046%) | 61 (0.041%) |
| Starter subscription | 26 (0.018%) | 32 (0.022%) |
| Creator subscription | 39 (0.027%) | 27 (0.018%) |
| Pro subscription | 2 (0.001%) | 2 (0.001%) |
| Growing Business subscription | 0 | 1 (0.001%) |
| Contact sales | 17 (0.012%) | 22 (0.0015%) |
| Signed up → subscription | 16.38% | 11.87% |
| Est. MRR | $1,204/mo | $1,869/mo |
Beyond the Landing Pages
The dark, code-first developer page pattern was recommended for the marketing site’s API pages
SFX deep-link sign-up routed visitors straight into the tool; they bought larger plans, about 9% more monthly revenue, so the deep link rolled out to the main page
A contact sales form redesign cut raw submissions 23% while qualification rose from 8.0% to 11.1% and SQLs from 55 to 64
A Growth Engineer was hired to absorb front-end builds, and Landing Page Engine became a Q4 input goal
What I Brought
Design
Built the LP design system and Framer toolkit, and designed the pages and test variants
Turned the winners into a template library of five patterns
Strategy
Owned paid landing page conversion end to end: build, test, measure, and report
Set the scoring rules: subscriptions and est. MRR per visitor, with sample integrity checked before significance
Systems
Built the Claude-driven engine: strategy project, Framer MCP and CLI pipeline, asset automation, QA, close-out, and scheduled automations
Wrote the operating manual and claims registry that every person and agent touching join. follows
Collaboration
Worked with Paid Media, the Core Experience team, and the Growth Engineer hired in Q4
Reflection
Vocabulary match beats page quality. The same speech to text page won on “speech to text” queries (+24.5% sign-up to subscription) and lost on “audio to text” queries (-11%).
Interactive-first loses on paid traffic. Three consecutive hero tests that led with an interactive component lost to the simpler control.
Removing navigation alone does nothing. Six no-nav tests lost across Google, Meta, and TikTok; content strategy and UX carry the wins.
Write it down once. Every surprise in a readout becomes a rule in the operating manual, so the next page starts from the rule instead of rediscovering it.
More screens
AI Voice Replicator: subscriptions 101 vs 62 against Voice Cloning v5
Dubbing Studio: variant win with a 73% lift over the marketing site
Music: variant win with a 51.13% lift over /music
AI Audiobook Generator (Google): variant win with a 40% lift over the marketing site
Avatars: variant win with a 38% lift over the marketing site
AI Image Generator: variant win with a 15% lift over the marketing site