Why BoldVoice Charged From Day 1 and Ignored the Duolingo Freemium Playbook — Anada Lakra

At launch (during YC, 2021) shipped a plain paywall: ~$10/month with a one-week free trial, no free tier. Rationale: willingness to pay reveals the ICP, so feedback and usage patterns come from people who will actually pay; avoiding paywall experimentation early let the team focus on onboarding and the core product. Paywall/pricing variants came later.

Primary experiment metric is net revenue after refunds per exposed user on a matured cohort, not trial-start rate or initial conversion. Also check failed payments. Maturation window: from trial start, allow the 7-day trial plus an extra week to convert, then at least two more weeks for refunds, since users ask for refunds after they see the charge hit their bank account. Accept slower iteration as the cost.

A/B tested a hard paywall at the end of onboarding: removed the X/dismiss button so people had to start the 7-day trial to proceed. Evaluated on matured net revenue after refunds.

A/B tested a variant where the main paywall kept its dismiss button, but dismissing it showed an offer of X% off (size not disclosed) with no way to close it, so users had to start a trial at the discounted price to proceed.

Paywall with 7-day free trial is shown after onboarding and can be dismissed. Dismissers get the speech assessment with full results breakdown (something people pay coaches for), the rest of the day-one experience (a practice set and, she believes, a coach video) and the first daily plan free. From day two they must start a 7-day free trial to continue. Aimed at skeptics who won't put a card down; she avoids removing the trial paywall entirely so people don't assume the product is free.

Paywall defaults to an annual plan (~$150/year) after a 7-day free trial, with monthly as a fallback and a ~$200/year higher tier. Annual cash arrives right after the trial, so for each paid channel they know subscribers vs spend (first-year LTV/CAC) immediately and reinvest the cash into the next customer. Monthly would leave retention (1 vs 24 months) unknown, and channels differ: TikTok users proved less loyal. Paid marketing run alongside product-led viral features, with strict LTV/CAC discipline.

Early on targeted international students at colleges, borrowing the campus playbook of dating apps/Facebook. Students improve naturally among native speakers, have lower ability/willingness to pay and haven't yet felt the career cost.

Deliberately ignore the beginner learner base Duolingo monetizes via freemium; no teaching basics or vocabulary. Target intermediate-to-advanced speakers, mostly career professionals and US immigrants, for whom English affects income. Price against the alternative (much pricier 1:1 speech/accent coaching), not against hobby language apps. Declined to chase native-speaker speech coaching as a nice-to-have.

Automated email invites go to a varied sample: 2-3 year subscribers, people who started a trial a week ago and converted or didn't. They book a 15-minute feedback call that automatically lands with whichever team members have calendar availability, so everyone (CEO, co-founder, engineers, growth) does at least two interviews a week. Built with automation (team uses Claude) so it stays a habit, not a one-off. Also occasionally bring users into the office to watch them use the app.

Run a super-user group of heavily engaged users with calls roughly weekly. Engineers demo features in progress and collect feedback before release, to check that complex pronunciation features are understandable and not intimidating.

Because happy users book calls and unhappy ones don't, send a three-question mini-survey in an email that people answer by just replying (no clicks), and add a short step in the subscription/trial cancel flow asking what went wrong.

Trained an in-house model that predicts a speaker's origin from ~10 seconds of speech and launched it as a separate web app (accentoracle.com): no download, no gate, no trial; record your voice and get results instantly, then get invited into the BoldVoice app. Shared it in the monthly investor update; a Korean angel investor posted it on Threads. After it spread, invested in accuracy and stability.

Periodically invest in the free viral tool with new unique features, e.g. going from country-level to city-level accent prediction, to re-trigger sharing.

On the free Accent Oracle results page, show a snippet/sample of the fuller results available in the BoldVoice app, so users see what else they'd get instead of taking a leap of faith.

An unprompted TikTok trend: users filmed themselves crying when the app gave them low scores. Not boosted by ads. Treated as a lighter, humorous awareness channel alongside paid marketing.

Launched as an 'accent training app' because accent was the spikiest, most differentiated pain at the last mile of fluency. After reaching brand/revenue scale and raising a Series A, expanding to the same learner's broader spoken fluency: idiomatic phrasing, sentence structure, grammar slips, clarity of ideas, confidence. Also meant to escape the stigma of the word 'accent', which suppresses word of mouth.

Hire founder-type doers who need little direction and coordination. Minimize distractions: one company-wide sync on Mondays so everyone knows what others are doing; otherwise heads-down building and talking to users.

Everyone uses Claude; the non-technical CEO ships PRs herself; the designer turns Figma designs into prototypes that get ~80% of the way before engineers take over. Also used for internal automations (e.g. user-interview invites).

Before any code: cold-reached Hollywood speech/accent coaches (who work with actors) for differentiation. First coach, Ron, recorded lesson videos; ~12 beta users were manually sent the videos by email, emailed back speech recordings, and the coach gave asynchronous feedback. Used the pilot to pitch investors.

A trading card value-scanner app gives the full scanning experience free but limits it to a number of scans, showing a counter in the scanner UI (e.g. '35 scans left', then 34, 33) to prepare users for the paywall.