How ElevenLabs Turns Feature Launches Into a Growth Engine – Luke Harries

Growth team works hand in hand with product/research. Each new model or feature (e.g. Scribe v2 speech-to-text, image/video in the app) gets: a launch video and X/tweet thread showing capability and benchmarks; a new landing page; and refreshed Meta, LinkedIn, Google (and TikTok) ads using the launch's hooks. Core messaging for the X post is iterated alongside ad copy. A separate in-house motion design team makes longer, fewer-variant brand/launch videos.

Analyzed all top-performing and bottom-performing Meta and Google ad copy and built a custom GPT on it. For each launch, take the brand X post and pass it through the GPT to iterate it into high-performing Meta and Google ad copy.

Started with agencies to test many formats and find what fits, then brought creative in-house to raise quality and quantity. Two 'AI creative producers' (motion designers by background) study data and competitors daily using Motion, then make After Effects motion-design ads layered with ElevenLabs text-to-speech voiceovers, music and AI-generated images. Swapping hooks, voiceovers and music gives quick variants while keeping quality high; UI-based flows show the abstract AI product well.

An in-house creator (Alec) with a YouTube channel makes person-on-camera studio videos showing how to get the best out of ElevenLabs tools. AI avatars considered for the future, but kept human for quality.

Instead of one general ad ('how to make images and videos with AI'), use AI to produce segment-specific versions (how podcasters, or people making apps, can create with image and video).

Budget is mainly bottom-up: each channel is its own growth engine. For Google search (biggest channel): strong ad copy, localized to all geos, high-converting landing pages, then scale 10-20% week over week. At each step ask what the next bottleneck is: first funnel optimization, then localization, and so on. Meta handled similarly with quantity plus quality of creative.

Previously paid an expensive localization SaaS plus an agency. Team kept finding ChatGPT phrasing better, so they ripped out that infrastructure and built their own GitHub Action plus a thin library that extracts/wraps all strings. Every PR sends new strings to an LLM with a per-language localization-guide prompt. Localized apps then enable localized ads.

Once apps are localized, localize all ads: dub videos with ElevenLabs' dubbing model and translate ad copy by passing it to an LLM with good prompts.

For self-serve web (bulk of spend), judge channels on direct ROI. Run in-platform lift studies (A/B tests with holdout groups) since a known brand gets many conversions that would happen anyway, especially from retargeting. Pull conversion data from each platform, apply incrementality adjustments that are updated every quarter, and allocate budgets from the adjusted numbers. Mobile, self-serve web and enterprise treat brand vs performance differently.