AI UGC ads look fake for one reason: the phone is in the shot.

Scroll past a few AI-made "customer" videos and you can spot them in half a second, usually before anyone speaks. The giveaway is rarely the face. It is a phone visible in the hand of a person who is supposedly filming themselves with it, a shutter button drawn over the picture, a mug with a logo that does not exist, a man drinking coffee and talking at the same time.

So what are the rules that stop a generated person from looking generated, how fast can they talk, where do the captions go, and what does one of these reels cost? This is the pipeline Ultim built for its UGC style, with the defects that produced each rule.

What a UGC reel is, in this pipeline

A real-looking person in the business's world talks to the front camera of their phone while doing something in between: a welder holding a machined block on the shop floor, a founder with a cold brew at a kitchen-table desk, a cook with a ladle mid-service. One still image is generated as the start frame, then an image-to-video model animates it into 5 second clips of the person speaking a scripted line, in their own generated voice. Four clips, influencer-style captions, a quiet music bed and a 2.5 second end card make a 20 second reel. The brief, the script and the publishing around it are the same as in the rest of the pipeline.

Two modes. Basic: a believable person for the industry, imperfect on purpose (tired eyes, a bandana, an oil smudge). Boost: an unexpected presenter, funny and good-natured, never mocking, such as an 88-year-old with a neon-green mohawk. Boost reels may add one or two silent b-roll clips between lines.

Rule 1: the phone is never in the frame

The image has to be exactly what the front camera sees. The phone itself is outside the frame, there is no screen overlay, no record button, no camera interface. The first round of start frames broke this three times: a visible phone in the hand on two frames, and a shutter button with a flash icon drawn onto a third.

The fix is not more negatives. When the prompt was given an extra line, "no phone anywhere in the scene, no device with a screen", the model drew the entire iPhone camera app, status bar and PHOTO/VIDEO tabs included, over the picture. Every extra mention of the word "phone" or "screen" is a cue to draw one. The sentence lives once, in a fixed style suffix, and nowhere else.

Rule 2: no readable text, so describe everything as blank

A bare "no logos" is ignored. What works is describing every object that would normally carry print as blank: "a plain hoodie with no print", "blank white price cards", "a soft-blurred unreadable dashboard", "a crushed plain silver unprinted aluminium can". One boost frame had to be regenerated because the first attempt drew a recognisable soda-can design and printed mail in the gutter; the second, with "plain blank white envelopes", came out clean.

Rule 3: one prop, both hands accounted for

One arm reaches out of the bottom of the frame towards the camera; that hand holds the phone, out of view. The other hand holds exactly one prop. Leave a hand undescribed and the model invents a second phone, a second cup or an extra hand.

If the business sells a physical product, one or two real product photos from the site go in as reference images. Without that, a 3D-printed puzzle came out as a cardboard jigsaw and a desk lamp as a generic round-shade lamp. With two references, the model copied both faithfully.

Rule 4: nothing covers the mouth while the person speaks

In our six-model test every founder clip had him drinking and talking at the same time, because the prompt asked for "a sip". The owner's note on every one was the same. The rule now: the prop stays still and away from the face, no sip, no bite, no hand near the mouth, written explicitly before the dialogue and naming the prop. Mid-actions belong in clips with no dialogue. In the two pilot reels the model obeyed this in 10 of 10 clips.

Rule 5: 2.2 words per second

The model speaks at about 2.2 words per second with natural breaths. It does not speed up to fit a long line; it drops the end. Measured in the pilots, 10 to 11 word lines took 4.5 to 4.7 seconds of a 5 second clip. So: at most 11 spoken words per 5 second clip, at most 8 for the hook, counting "sixty-seven" as two and "dot com" as two.

A line that is too short is also a defect. A 6 word hook leaves 3 seconds of silence, and the model fills it with whatever text follows the quote in the prompt. In the first take of both pilot hooks it read the pronunciation note aloud: "Ultim is pronounced ULTIM. The domain is set as ULTIM dot ESSEY." The fix: the pronunciation direction goes before the quote, the hook gets an explicit "after the line he stays quiet and gives a small nod", and the clip is trimmed to the end of the line plus 0.9 seconds.

Captions like the influencers use

Word-level captions from speech-to-text timestamps, aligned to the script so the spelling is the script's. Groups of 2 to 3 words, at most 18 characters, uppercase, one line, a bold condensed sans at 92 px, white with a black outline, the word being spoken in the brand's accent colour.

Position matters more than the font. The captions sit at 66% of the frame height, below the chin, because on these selfies the mouth is at 40 to 48%. A caption over the mouth defeats the point of a person talking to you. Each clip is normalised to -16 LUFS, the mix to -14 LUFS, with the music bed ducked about 20 dB under speech.

What it costs

Item	Model	Cost
Start frame or b-roll frame	Nano Banana Pro, 9:16, 2K	$0.09
Talking clip, 5 s, audio on	Seedance 1.5 Pro, 720p	$0.175
B-roll clip, 5 s, audio off	Seedance 1.5 Pro, 720p	$0.0875
Short reel (hook plus 2 lines, about 15 s)	1 frame, 3 clips	$0.62
Standard reel (hook plus 3 lines, about 21 s)	1 frame, 4 clips	$0.79
Boost reel with 2 b-roll	3 frames, 4 talking, 2 b-roll	$1.15

The real pilots matched the estimates. Two complete 20.6 second reels (a founder for Ultim, a welder for a job-shop directory) took ten Seedance generations including two retakes: 350 credits, $1.75. A boost demo reel for a fictional junk-hauling brand, with three talking clips, two b-roll clips and four start frames including one rejected, came to 212 credits, $1.06. Word timestamps for the captions add about $0.002 per reel.

Seedance 1.5 Pro is the default because of the six-model test: correct words, lips on the words, props that keep their shape. Veo 3.1 Fast can take one hero clip at $0.30 when the budget allows.

Basic or boost

The owner rated the welder reel (basic) and the junk-hauling reel (boost) 5 out of 5, and the note on a weaker frame was "too weak, does not catch the eye". So even a basic presenter needs one memorable, specific trait. Boost adds one or two surprises in the look; the surprise is never in the behaviour.

Which to use is an empirical question, and the pipeline treats it as one: each reel tests one variable, the 24 and 48 hour reads decide, and a format that fails three times is benched.

What this means for your reels

If you are making AI UGC yourself, write the rules into the prompt once and check every frame against them: phone outside the frame, no camera UI, every printable surface described as blank, one prop and two hands, the product from a reference photo, the mouth free during every line, 11 words per 5 seconds. Each of those came from a defect, and each one costs a regeneration when you skip it.

If you would rather paste a link: Ultim's UGC style runs these rules with captions on by default, basic or boost, from the facts on your site. The first thing to check is still the words. Paste your site into /tools/reel-hooks and see which of the five hooks you would want a person to say to camera.

FAQ

What does an AI UGC reel cost to make?

In Ultim's pipeline a standard reel (one start frame plus four talking clips, about 21 seconds) is $0.79 in generation; a short one $0.62; a boost reel with two b-roll clips $1.15. Two pilot reels cost $1.75 including retakes; a boost demo cost $1.06.

Why do AI UGC videos look fake?

Most often because the generated frame shows the phone, a camera interface or readable text, because the person eats or drinks while talking, or because the line is too long for the clip. Each has a fixed rule that prevents it.

What is the difference between basic and boost?

Basic is a believable person from the business's world, with one memorable trait. Boost is an unexpected, funny presenter, with one or two surprises in the look, plus one or two silent b-roll clips between lines.