how to create AI influencers that get 7.3M views for your app (step by step)

stop wasting time trying to get 1k views without understanding how to actually do it.

in this article, i'll cover how to reverse engineer a viral format + how to recreate it by building an AI influencer in a few simple steps.

i've been too lazy lately but i found this method insanely easy to follow.

and honestly?

if you don't know how to go viral yet, this is probably the only article you need because i'll cover the entire process from start to finish.

what do you need?

since we're going to create AI influencers, you just need this pipeline:

gemini -> gpt image 2 -> seedance 2.0

(i'm going to use arcads to make everything easier and faster).

i'll explain exactly why you need these tools and how to use them.

what are we going to create?

i'm scrolling up to 3 hours daily because my whole account is basically a database of app and marketing ideas, so it usually happens that i find viral formats and share them on X.

yesterday i found this account:

annastepsy

1294 followers, 7 videos and over 9M views.

but why am i showing you this account?

it's a fresh new account

we can literally take their format

it's a super realistic AI influencer (and people genuinely can't tell)

this is the most viral video from this account:

7.8M views and 1.3M likes

we're going to analyze every single frame of this video before recreating it, but how do we actually do it?

step 1: analyze the video with gemini

i don't know why but a lot of people still don't realize that you can literally send a tiktok to gemini and ask it to analyze:

frame by frame

the hook

the pacing

the movement

basically everything inside it

this is the exact method i'm using to study viral formats from competitors and adapt them to my apps.

here's the workflow:

download the tiktok (you can do it directly from the platform)

send this prompt to gemini

i want you to analyze this tiktok, from start to finish, without including the watermark or anything about the tiktok UI.

give me these info:

1. pacing

2. physical characteristics of the girl

3. a very detailed frame-by-frame description of the movement she's doing

4. the text hook she's using

5. the visual hook she's using

we're going to use every single information we extract and you'll understand why while reading the full article.

in this case, here's what gemini sent back:

Here is the breakdown of the video clip:1. Pacing

The video follows a very fast-paced, high-energy rhythm. It begins with an immediate, dynamic movement and abruptly cuts as soon as the climax (the fall) occurs. There is no lead-up or slow build, creating an effect of sudden chaos.2. Physical Characteristics of the Girl

The girl has blonde hair pulled back into a ponytail. She is wearing a bright red, form-fitting long-sleeved athletic top. She appears to be in an outdoor setting at night, as her clothes are drenched from rain and glistening under artificial street lighting.3. Frame-by-Frame Movement Description

0:00–0:01 → handheld shaky jogging shot

0:02 → balance starts shifting

0:03 → panic reaction + recovery attempt

0:04 → complete fall forward

0:05 → impact cut4. Text Hook

“When my phone won't unlock tt or insta til I hit 3k steps and the pavement randomly decided it was my time.”5. Visual Hook

Extreme weather + chaotic energy + sudden fall.

as you can see, this is basically free market research.

you're not guessing anymore.

you're reverse engineering attention.

step 2: create the AI influencer

now we're going to use the physical characteristics extracted by gemini as our base prompt.

i'm just adding some extra details like:

handheld tiktok selfie vibe

realistic bad lighting

motion blur

low quality iphone feeling

because perfection isn't viral.

you can literally copy paste this prompt:

A handheld TikTok-style selfie video of a blonde girl running at night in the rain, filmed front-facing on a phone camera. Her wet blonde hair is tied back in a messy ponytail, strands sticking to her face. She’s wearing a tight bright-red long sleeve athletic top soaked from the rain, glistening under blurry streetlights.

The footage feels raw and spontaneous, with slight motion blur, shaky camera movement, uneven framing, autofocus hunting, compression artifacts, low-light grain, and occasional overexposed highlights from passing lights.

She’s breathing heavily while running, laughing slightly, energetic and chaotic.

The vibe is authentic late-night TikTok selfie footage - imperfect, candid, cinematic but low-quality, as if recorded quickly while sprinting through the city in the rain.

you can also ask gemini to improve prompts like this.

just tell it:

“make this feel like a real iphone selfie with bad lighting and shaky movement”

that's usually enough.

i'm going to paste this inside arcads using gpt image 2 in 16:9 format.

look at these insane results:

now that we have our AI influencer, we can move to the most important part.

the actual video.

step 3: create the tiktok

well, technically we already have the entire prompt structure generated in our research.

but i'd still paste everything into gemini and ask it to enhance:

movement realism

pacing

video quality

camera imperfections

because details are what can actually make you viral.

here's the prompt i used on arcads with seedance 2.0:

Vertical TikTok-style selfie video, chaotic and hyper-authentic, filmed on an iPhone front camera by a blonde athletic girl running outside during a storm at night. The entire clip feels accidental, spontaneous, and emotionally impulsive — like a real viral TikTok captured in one take.

The girl has wet blonde hair tied in a messy ponytail, with soaked strands sticking across her face and forehead. She wears a bright red, skin-tight long-sleeve athletic top that becomes glossy and reflective from the rain. Her skin is wet and shiny under harsh streetlights. Heavy rain pours around her while she runs through a dark suburban street at night.

The camera is fully handheld selfie perspective, held at arm’s length while sprinting. The framing constantly shifts imperfectly: sometimes too close to her face, sometimes cutting off parts of her head or shoulders.

0:00–0:01 — She is laughing breathlessly while running fast toward the camera, eyes wide, adrenaline high, speaking or screaming incoherently between laughs.

0:02 — Her footing suddenly slips on the wet pavement. Her expression instantly changes from joyful chaos to genuine panic.

0:03 — She begins falling hard to the right. The phone dips rapidly downward while still pointed at her face.

0:04 — Full loss of balance. Motion blur becomes extreme.

0:05 — Abrupt impact moment. The camera slams near the pavement and the clip instantly cuts.

Visual style:

• authentic TikTok energy

• imperfect iPhone front camera footage

• low-light smartphone noise

• realistic rain physics

• messy framing

• aggressive handheld shake

• cinematic but accidental feeling

• raw viral TikTok vibe

• abrupt ending

and here's the result:

as you can see, this is literally perfect as a visual hook.

people instantly stop scrolling because the brain detects:

movement

chaos

emotion

danger

comedy

all in the first second.

step 4: upload on tiktok

now that we have everything, you can add the extracted hook directly from tiktok.

and yes:

ONLY use hooks from tiktok itself, not external platforms.

here's how i'd warm up the account:

warm it up for ~2 days

search videos in your niche

comment 2x daily

watch entire videos instead of doomscrolling

act like a real human

that's it.

people overcomplicate warm up way too much.

what i'd personally do

i'd probably use this as a low effort format.

i wouldn't even mention the app name directly inside the hook.

i'd only mention it in:

description

comments

pinned replies

why?

because this format works through curiosity.

people will naturally go to the comments after watching because the format triggers an innate sense of hilarity and confusion.

and what are they going to find there?

your app name

that's why this works so well.

another thing i'd do:

i'd probably reduce the quality slightly to make it even more believable.

don't upload ultra hd 4k IMAX cinematic AI videos because people can instantly tell.

imperfections are what make content feel real.

the end

i want to say thank you to everyone that read the whole article.

i've genuinely been enjoying writing these because i can finally share what i've actually been doing lately instead of gatekeeping everything.

and huge huge thank you to arcads for making this possible by sponsoring this content.

Links

• www.arcads.ai/?comet_custom=simonecanciello