AI UGC Avatars Playbook: Create 1,000 Ads Variations using Arcads

Here’s the uncomfortable truth about paid ads: the winners aren’t clever audience hacks or secret attribution switches. The winners are the teams who ship an absurd amount of creative, week after week, and let the data surface outliers.

For years, that level of throughput meant wrangling UGC creators, coordinating reshoots, and paying ~$100 per video—about $100,000 if you wanted 1,000 variants. It was slow, expensive, and a logistical mess.

That constraint is gone. With AI avatars, you can write a script in the morning, generate a talking head that nails the delivery, clip a few edits, localize to Portuguese/Spanish/French, and launch dozens—sometimes hundreds—of variants by afternoon. It feels like stealing time back.

What follows is exactly how to run this play, end to end, distilled from Arcads co‑founder Romain Torres’ workflow and what top performers are doing in the wild.

Gaming’s 8k–20k ads blueprint for non‑gaming brands: Outliers are hidden in volume

Meta and TikTok’s machine learning (ML) now does most of the audience work for you. That shifts the advantage to creative throughput. Volume exposes the outliers you’d never predict in a brainstorm—unexpected hooks, pain points, or a certain face that just gets more people to stop scrolling.

If you’re spending $100/day, a $1.00 cost per click (CPC) gives you ~100 clicks. Nudge that to $0.10 CPC and you’re at ~1,000 clicks. Same budget, 10x traffic. That’s what creative iteration buys you at scale.

Gaming figured this out a decade ago. Studios like King (Candy Crush) and Voodoo routinely run 8k–20k live ads in a single account. Non‑gaming can finally match that velocity with AI‑assisted production.

Why The Old UGC Playbook Broke: The $100k problem

The math killed it. At ~$100 per video, going to 1,000 variants means ~$100,000—before you even talk talent scheduling, reshoots, or localization.

Iteration cycles took weeks. By the time you found a winner, fatigue set in or the window had closed.

Localization in a click: turn one winner into ten markets using AI Avatars

The new workflow is simple: write the script → generate the avatar reading it → do quick clips/overlays → multiply variants. Tools like Arcads remove coordination overhead and make consistency a default.

What you get:

• Cost collapse: a thousand creatives for a few hundred dollars instead of ~$100k.

• Speed: ideate, produce, and launch in hours—not weeks.

• Control: on‑brand delivery with zero talent logistics.

• Instant localization: flip to Portuguese, Spanish, French, etc., without hunting native creators.

Two concrete signals from Romain’s workflow make this real.

First, actor multivariate testing: running the same script with 20 different actors produced click‑through rate (CTR) differences up to ~180%. That doesn’t guarantee 180% more return on ad spend (ROAS), but it often means roughly 2x the traffic for the same spend.

Second, the Learna AI example: an English‑learning app that scaled to ~300 active ads and roughly $800k monthly recurring revenue (MRR) in ~9 months, primarily through paid.

Kill losers fast, pour gas on the top two

Pick a budget you’re willing to burn to learn—say, ~$10,000 over six weeks—and commit. Most teams fire one volley and quit. Don’t. Make it a weekly ritual, even if it’s just you and a Notion board:

• Ideas → In‑Progress → Testing → Learnings.

• Ship 5–10 concepts per week. For each concept, produce ~10 variants (≈100 ads per week).

• In the ad account, one ad set = one concept. Keep 5–10 creatives per set so everything gets a fair shot. Let platform ML distribute within the set; compare performance across sets.

• Next week: double down on the top one or two concepts with +50 variants. Kill the bottom half. Add at least one orthogonal concept to avoid getting trapped in a local maximum.

This cadence is boring, repetitive, and incredibly effective.

Steal like a pro: Ads Library → Whisper → patterns

Start where every good thief does: the Facebook Ads Library. It’s there to make platforms transparent; marketers use it to get smarter.

Set up an automation to watch competitor pages, auto‑download every new video, and run them through OpenAI Whisper for transcripts. Now you have the actual scripts, hooks, claims, and on‑screen pacing in text. Cluster them. What angles repeat? Which CTAs keep coming back? When a team repeats a pattern, it’s a tell.

Then use Perplexity. Paste the entire website of the product you’re promoting and ask: “What pain points does this solve? Who are the personas? What outcomes do they want?” From each pain point, generate 20–50 hooks tied to outcomes. Include a few examples of known winners and instruct the model to “penalize stylistic deviation” so tone stays consistent.

Finally, mine Reddit. Subreddits and comment threads are where users say the quiet part out loud. Pull verbatims into captions and line reads. If you sell to humans, write the way humans actually talk.

MrBeast the process: repeat the winning format with 15s/30s/45s cuts

The structure is predictable because it works:

• Hook that interrupts the pattern and hits a real pain.

• Story arc: problem → failed attempts → product as hero → payoff.

• Proof: quick demo, social cred, or micro‑metric.

• CTA that asks for the smallest next step.

Write in 15s/30s/45s versions. Keep captions and emojis in mind from the start. Use a MrBeast‑style habit: once a format works, repeat it shamelessly like MrBeast with different hooks. Success loves templates.

Swap the face, change the outcome: test 20 actors

Avatars aren’t just cheaper talent; they’re a testing tool. Run the exact same script with 10–20 different actors. Change age, gender, vibe, accent. Watch CTR move. That single lever often decides the week.

For voice, use high‑quality text‑to‑speech (TTS) such as ElevenLabs and match energy to the hook (urgent hook ≠ sleepy voice‑over). In the edit, layer B‑roll/product overlays under the talking head, cut fast, and use kinetic captions, emoji markers, and progress bars. Add a subtle “ding” on success moments—gaming psychology in 0.3 seconds.

Localization isn’t an afterthought. Translate the script, re‑voice it, swap on‑screen text, and adjust currency and cultural references. Launch country‑specific variants side‑by‑side. It’s common for a “solid” US ad to become a “killer” in Brazil or Spain.

Ecom, apps, B2B: formats that print attention

For e‑commerce, the classics work: narrator avatar over product demo, fast problem/solution cuts, and tactile unbox/pack‑and‑go sequences that feel native to Reels/TikTok.

For mobile apps, the duo‑dialogue format is gold: user on the right, app/robot on the left. Show the user make a mistake, then succeed, punctuated by a little reward sound. That’s the Learna AI pattern, and it grabs attention.

For agencies/B2B, try “street interviews” with avatars for scalable social proof, before/after workflow demos with metric overlays, and tight case‑style shorts where you stack three metrics and ask for the click.

Beat fatigue by refreshing intros, hooks, faces

Keep targeting broad and let the platform’s ML work. Structure by concept, not by clever audience slices. In the first 24–72 hours, watch thumbstop rate, 3‑second view, percent video viewed, and click‑through rate (CTR) to prune. When something looks promising, allocate enough spend to assess cost per acquisition (CPA) and conversion rate (CVR) before making a call.

When you find a winner, squeeze it:

• Promote the top creatives within the winning concept.

• Clone into new audiences, placements, and locales.

• Refresh intros, hooks, and actors to stay ahead of fatigue.

One‑click: “turn competitor ad into my ad”

You can build 80% of this with no code:

• Ads Library → Make.com (or Python Ads Library) to poll competitors.

• Auto‑download videos → OpenAI Whisper transcription.

• Perplexity/GPT for clustering, hook generation, and first‑pass scripts with few‑shot guidance.

• Avatars + text‑to‑speech (TTS) built into Arcads; assemble with templates; keep monitoring on in the background.

The advanced loop isn’t much more complicated:

1. Poll competitor pages.

2. Transcribe and cluster themes.

3. Generate on‑brand scripts (“penalize stylistic deviation”).

4. Produce avatars; version actors, hooks, lengths, and locales.

5. Export with clean naming and metadata for import.

Add a growth loop while you’re at it: offer “competitor‑monitor” email alerts with a one‑click “turn into my ad” button. It’s useful on its own and naturally feeds the pipeline.

Track Like A Pro by building your private “Winning Formats Library”

Name assets so you can learn. Use a consistent, machine‑sortable pattern that encodes every testing dimension:

Concept__HookID__ActorID__LengthSec__Locale__v{n}

• Concept: short slug for the idea/theme (e.g., “language‑duo‑dialogue”).

• HookID: numeric/short hash for the opening hook (e.g., H037).

• ActorID: avatar/actor identifier (e.g., A12_female_genZ_US).

• LengthSec: video duration in seconds (e.g., 15, 30, 45).

• Locale: i18n code (e.g., en_US, pt_BR, es_MX, fr_FR).

• v{n}: incremental version for edits (e.g., v1, v2).

Example: language‑duo‑dialogueH037A12_female_genZ_US30pt_BR__v3

Keep a simple database that links hook → script → actor → edit notes → performance. Do a weekly top/bottom quartile review, then codify what you learn into prompts and templates. Over time, you’ll build a private “Winning Formats Library” that new hires (or future you) can use immediately.

Fatigue kills ROAS: rotate motifs like a pro

Stay inside claims guidelines for your category and market.

Use a style guide to avoid uncanny‑valley avatars. Rotate intros, motifs, and actors to manage fatigue.

And remember: “inspired by” structures are smart; copying scripts verbatim is lazy and risky.

6 weeks, 5 concepts, 100 ads: Out‑ship your market and let the math work

Set up the six‑week cadence. Build the Notion board. Pick five concepts. Use Ads Library to steal angles, Perplexity to write hooks, and Avatars + ElevenLabs to produce. Launch fair tests (ad set = concept). Watch thumbstop and CTR early, CPA later. Next week, pour gas on the top two concepts, add +50 variants, and localize to three languages.

AI won’t replace creative strategy. It removes the production bottleneck so teams can execute one. The marketers who embrace volume, structure tests, and iterate weekly will buy the same attention for less—and scale faster without hiring a small army.

Hat Tip to Romain Torres for the workflow and In The Pit podcast for the insights.

PS: This was written with the help of rumoured GPT-5 model

Top Tweets of the day

1/

Automations for local businesses are so cool.

2/

Speed matters. Honestly, the only reason to use paid models is speed because intelligence is converging at a fast pace. It takes 3-6 months to make a model 80-90% as intelligent as a SOTA (state-of-the-art) model.

And lots of problems don't need high intelligence but they need high speed so you can stay in flow state.

Make a lot of money so you can spend $1k-$10k per user per month on high-quality high-speed models. There will be a cheaper $2k per month plan with limits to get majority of users.

Its still cheaper than a decently competent human.

3/

There are 2 types of businesses:

1. One that serves free users (OpenAI)

2. One that serves mostly paid users (Anthropic)

The 2nd one has less overhead and less costs but can make the same amount of money and give quality service to paid users.

Another example would be OpenRouter and Requesty. Both do the same thing but Requesty has no free plan so less overhead, server costs are lower, and fewer support tickets.

Rabbit Holes

1. Leah Belsky on how AI is transforming education — the OpenAI Podcast by OpenAI

2. Agency Is Eating the World by Gian Segato

3. The Transition: Layering sales onto a bottom-up self-serve product by Lenny Rachitsky & Pete Kazanjy

What’d ya think of today’s newsletter? Hit ‘reply’ and let me know.

Do me a favor and share it in your company's Slack #marketing channel.

First time? Subscribe.

Follow me on X.

More Startup Spells 🪄

1. How Oleve Built a $6M ARR AI Empire w/ 4 employees from a Homework App (Quizard AI/Unstuck AI) (LINK)

2. Sabrina Ramanov's Perfect SaaS Retention Strategy (LINK)

3. How Unroll.me used Sharewall instead of Paywall for Viral Growth (LINK)

4. Viral Donut Resume Hack That Landed Job Offers and Closed Enterprise B2B Deals (LINK)