How to create an
AI influencer.

Before you generate a single video, settle three questions: who, what niche, and which face. Skipping this and going straight to a tool is the most common reason AI influencer channels die after 20 posts.

Who is the persona's identity. Are they a real human (you or your spokesperson), a fictional human (a character we help you build), or a stylized non-human (a brand mascot)? All three work. The choice locks downstream: real humans need release-form documentation, fictional humans need a one-time character build, and brand mascots avoid the disclosure rules entirely.

What niche determines the algorithm bucket. Fitness, beauty, finance, business education, wellness all work for the persona format. Stay narrow at the start. Eden Brooks is “women's wellness + supplements” rather than “health” in general. Narrow niches grow faster because the algorithm has a clear signal about who to show your content to.

Which face is the part most builders rush. The face you pick has to look right across many emotional states, many camera angles, many lighting conditions, because the AI will generate dozens of variations of it. Some faces don't generalize well from the reference photos. Spend the day picking five strong references rather than five convenient ones.

With the identity decided, build the face references. Five photos is the working minimum. They have to share style: same hairstyle, same approximate lighting, same makeup or grooming. What varies is angle and expression. Three-quarter left, three-quarter right, dead-on, slight smile, neutral.

If you're using AI-generated faces (a brand spokesperson who doesn't exist), spend extra time here. The five references become the “same person” signal that every future generation will reference. A loose reference set gives you a different face every Tuesday.

Concrete: filenames matter. The pipeline reads photos in alphabetical order and treats the first one as the anchor. Name them 01_anchor.png,02_left.png, etc. So the strongest reference is always the anchor.

The voice profile is the single biggest factor in whether the AI influencer reads “human” or “AI”. The common mistake is using a generic neutral voice and assuming the content will carry it. The content can't carry it; the voice is the first thing the viewer hears.

Write a voice profile of ~40 words that covers: cadence (how fast they talk), vocabulary (specific words they use, specific words they avoid), and anti-patterns (filler words to cut, generic AI tells to avoid). Then tune a TTS voice clone against it. Eden's profile, for reference:

“Confident but not preachy. Says ‘here's the thing’ and ‘the bit nobody talks about’. Never says ‘absolutely’, ‘literally’, or ‘at the end of the day’. Pacing is medium-fast. One stat per video, sourced. Anti-pattern: 90-second monologues. She's in and out in 30.”

With the profile written, clone the voice in your TTS provider of choice (ElevenLabs and fal both work well). 5 minutes of sample audio is enough.

The hook is the first 3 seconds of every video. Algorithm weight on first-3s retention is enormous on both Reels and Shorts. A good hook bank is the difference between 200 views/post and 2,000.

Build 40-60 hooks across five archetypes:

• Question hooks: “Have you ever wondered why...”
• Statement hooks: “Here's the truth about...”
• Contrarian hooks: “Everyone says X. They're wrong.”
• Story hooks: “I almost lost everything because...”
• Problem hooks: “If you're struggling with X, watch this.”

Tag each hook with the content pillar it fits (see step 5). The pipeline can then pick a hook that matches today's pillar rather than picking randomly and ending up with a promotional hook on what was supposed to be a vulnerability day.

Pillars are the categories your videos rotate through across a week. Without them, the feed feels random; algorithms and humans both notice. With them, the feed feels designed even though it's automated.

The four-pillar rotation we use for Eden:

Two struggle days, two education days, two lifestyle days, one convert day. The struggle / education / lifestyle mix balances empathy with utility; convert anchors at the end of the week after the audience trusts the persona.

Daily is the right cadence for AI personas. Less than daily and the algorithm forgets you exist; more than daily and you spread the audience across too many posts to build saved-watch time on any single one.

Pick a fixed time per platform. Instagram's data says evening (6-9 PM local) for most niches; YouTube Shorts is more forgiving and we post Eden in the morning. The exact time matters less than consistency. Algorithms reward predictable upload schedules.

For automation: schedule the generation pipeline 4-6 hours before the post time, so a quality-gate failure has buffer to re-roll a video. Eden generates at 4 AM local, ships at 9 AM.

Both platforms publish AI-generated content under their current AI-content policies, provided you disclose. Instagram uses the standard Meta Business / Creator account flow and their Graph API for publishing. YouTube uses the Data API v3 with OAuth.

Instagram setup: convert your account to Business or Creator (one tap in the IG app), link it to a Meta Business Suite Page, generate a long-lived Page access token with the right scopes (instagram_basic, instagram_content_publish, instagram_manage_insights). Without the insights scope your analytics dashboard will be empty.

YouTube setup:Google Cloud project with the YouTube Data API v3 enabled, an OAuth 2.0 client of type “Desktop app”, run the OAuth flow once per channel to get a refresh token. Keep in mind: apps in Google's Testing mode have refresh tokens that expire every 7 days regardless of activity. Either promote the app to Production (4-6 week review) or re-auth weekly until you have your first paying customer to justify the verification paperwork.

Add a content-disclosure mark at the account level (Settings → Original content & AI disclosures). Per-post disclosure requirements vary by region; default to the strictest.

With the persona built and the platforms connected, run the pipeline daily and wait. For the first 30 days, resist the urge to course-correct on every individual post. Algorithms need a stable signal to figure out who to show your content to; tweaking the niche or voice every week resets that clock.

What to track per week:

• → Views per post, per platform
• → Engagement rate (likes + comments / views)
• → Watch-through on the first 3 seconds
• → Followers gained
• → Which content pillar performed best
• → Which hook archetype performed best

Eden's reference numbers at day 30: 53 reels, 3,070 total views, 4.63% engagement rate, one locked persona across the entire run. That's a healthy small-channel baseline, not a viral hit. Most personas land here for the first three months. Growth comes from one or two breakthrough reels rather than steady accumulation.

By day 60, you should know enough to make persona tweaks: cut a pillar that's not landing, retire hooks that aren't getting first-3s retention, double down on the topics that are working.

We made every mistake on this list with Eden before catching them. Listing them so you don't.

• Face drifts between videos. The reference set isn't locked tightly enough and the generator improvises. Fix: redo the references with stricter lighting and grooming consistency. Five photos that look like the same shoot, not five different days.
• Voice sounds AI. The TTS clone is using too clean a sample. Add a few moments of throat-clear, slight breath, mid-sentence pace shift to the source audio. Synthetic voices that sound human have some imperfection.
• Hooks don't land. You're leading with the topic instead of the tension. “Today I want to talk about iron deficiency” vs “My doctor said my iron was lower than my 65-year-old mom's.” First lands flat; second hits.
• The Instagram container hangs forever.We hit this at three weeks in. Meta's Reels preprocessor needs the moov atom at the front of the mp4 (faststart), not at the end. If you're building this yourself, ffmpeg the file with -movflags +faststart before uploading. Six of Eden's early posts were lost to this bug before we caught it.
• YouTube refresh token dies weekly.Google's OAuth Testing mode has a hard 7-day refresh-token lifetime. If your daily cron silently skips a day every week, this is why. Either re-auth weekly or promote the app to Production.
• Engagement number looks great, but it's on 3,000 views. Small-sample engagement rates are noisy. The number to care about long-term is engagement rate at 50K+ cumulative views per persona. Below that, treat the percentage as directional, not absolute.