Compliance standard 02

Creative Standard: AI Generated UGC Video

Written 2026-08-04. This is the creative sibling of the Store Copy Standard. That document governs what the store may say. This one governs what an ad may say and show, how it is built, how it is checked, and how it is disclosed.

What this is. A review standard for short form video ad creative in the research compound category, written so a reviewer with no production or technical background can apply it. It is grounded in two evidence sets we hold: our 78 ad teardown of the one advertiser running this exact format at volume (August 2026, full assets on file), and the U.S. Food and Drug Administration warning letters to Gram Peptides (reference 721806) and Prime Sciences (reference 721805), both dated March 31, 2026, which are the basis of the store copy standard.

What this is not. Not legal advice. Where this standard and counsel disagree, counsel wins. It also does not approve anything to go live: every creative produced under this standard is delivered as a paused draft and nothing runs without the client's compliance approval and the account owner's explicit written approval.

The one sentence version. Every ad sells the lab, the price, and the paperwork. No ad ever says or shows what a compound does in a body, and no synthetic person ever appears to describe a personal result.

1. The format and the four part script skeleton

The format this standard governs is the single take selfie UGC talking head: one speaker, phone framing, burned in captions, 17 to 34 seconds. This is the format the reference advertiser (Value Research Laboratories) executed across all 78 of its live ads, and their caption structure maps exactly onto the four part skeleton the wider field converged on in 2026.

Every script is written in four parts. The timings are targets, not laws, but a script missing one of the four parts is incomplete and goes back.

Part 1. HOOK, seconds 0 to 3

The speaker is mid sentence from frame one and the caption is on screen from frame one. Cold open. No greeting, no dead air, no logo card.

The hook carries the ad. Write 10 to 20 hooks against one script body and hold everything else constant; the hook is the test variable. Observed working shapes from the teardown, as structure only:

Part 2. PROBLEM, seconds 3 to 8

One problem, stated commercially. The problem is always procurement: paying legacy supplier markups, missing documentation, slow shipping, no way to verify what arrived. The problem is never a body, a condition, or a result.

This is where the reference advertiser placed its quantified stat. We may use a number here ONLY if it comes from the client's own controlled source, per section 2.6. Their "save 35 to 50%" figure is an unverifiable comparative claim and is not available to us.

Part 3. SOLUTION or DEMO, seconds 8 to 20

What the supplier actually offers, usually as a caption checklist overlaying the continuing take: published certificates of analysis per lot, independent testing, transparent documentation, shipping terms, price. Everything in the checklist must be true, current, and sourced from the client's own documented record.

The compliant substance of this section is the same as the store copy standard's: the material and its documentation, never a body.

Part 4. CTA, final 3 to 5 seconds

One spoken and captioned call to action. A discount code is acceptable here and is currently passing review in this category, but note the recon finding: the only ad in the category older than six months makes no offer at all and says only "Learn More". Offer led CTAs buy volume and burn faster. Every batch should include at least one evergreen no offer variant built to last.

Production rules that ride along with the skeleton

2. The claim boundary

This section is written for the reviewer. You do not need to know how the video was made to apply it. Read the script and the captions, watch the video, and apply these tests in order.

2.1 The one question that decides most cases

Does this sentence, image, or implication tell the viewer something about a body, or something about the material and the business?

Body: reject. Material, documentation, price, shipping, testing, the company: allowed. This is the same test that governs the store, and the FDA letters are why: a product becomes a regulated drug when the seller's own words show it is intended to affect the structure or function of a body. The agency applied that test to hedged, research voiced website copy and set the research use disclaimer aside while doing it.

2.2 The hedge removal test

Take any sentence with a softener: "studied for", "associated with", "shown to have potential for", "research suggests". Remove the softener. If a claim remains, the sentence fails. The FDA quoted exactly these constructions back to Gram Peptides; the hedge did not convert the claim into a neutral statement, and it will not for us.

2.3 The hard reject list

Reject the creative outright if it contains any of the following, spoken, captioned, in primary text, or implied by imagery:

  1. Any health, medical, dosing, or outcome claim.
  2. Any description of what a compound does inside a body, however phrased.
  3. Any before and after, in words or pictures.
  4. Any prescription, treatment, or clinical framing, including "patients", condition names, approved medicine brand names, or trial references.
  5. Any first person account of results. See 2.4, this one is doubled.
  6. Administration or regimen content: injection, dosing, schedules, routes, or imagery and icons implying any of them.
  7. Reconstitution supplies in any form: bacteriostatic water, syringes, needles, kits, whether sold, shown, mentioned, or offered free. See 2.5.
  8. Any number describing an effect on people.

2.4 The synthetic testimonial rule

A synthetic or AI generated person who appears to give a personal account of results is TWO violations at once:

There is no compliant wording of a first person results story in this category. Do not soften it, do not hedge it, do not shift it to "a friend of mine". The speaker in our ads is a presenter of commercial facts, never a witness to an outcome. This rule holds even though the realism techniques we use could make such a testimonial look genuine. That capability is precisely why the rule is absolute: a hyper real creator saying a non compliant sentence is a worse problem than an obviously synthetic one saying nothing, because it is more persuasive and it is still a claim.

2.5 The free bacteriostatic water trap, named specifically

The reference advertiser puts "Free BAC water" in its ads and on its landing page. Do not copy it, reference it, or soften it into "reconstitution supplies included". The FDA's Gram Peptides letter ruled that selling bacteriostatic water alongside compounds makes the water itself a drug by demonstrating intended use, and giving it away is not meaningfully different. It is separately a Meta policy violation, since ads may not gift freebies of restricted products or promote paraphernalia. This is the single clearest example of why "it is live and approved right now" is not evidence of safety: the advertiser running it is eleven days old.

2.6 Where claims may come from

Every factual statement in an ad, price, discount, shipping terms, testing regime, certificate availability, guarantee, comes from the client's own documented record or a source the client controls. An unverifiable claim is a hard reject: rewrite the ad without it, do not soften it. This kills, among other things, the reference advertiser's "save 35 to 50%" figure, any "number one" or "most trusted" superlative we cannot substantiate, and any purity statistic not backed by the client's own certificates.

2.7 What the ad MAY say, so reviewers know what good looks like

Price and value framing aimed at labs and research buyers. Competitor price monitoring, if the client actually does it. Published certificates of analysis. Independent laboratory testing. Documentation transparency. Shipping speed and shipping protection. Discount codes. The research use statement. This is the entire permitted persuasion load, and the teardown shows it is enough: the reference advertiser built 78 approved ads from exactly this list and nothing else.

2.8 Everything on the image is copy, including the label on the pack

Added 2026-08-12 from our teardown of 76 live Eon Peptides ads, captured that day from the public Ad Library. This is the ad-side twin of Store Copy Standard 4.11, "the web address is copy", and it is the same defect wearing a different surface.

The rule. Every rule in section 2 applies to text that appears inside the creative exactly as it applies to spoken lines and primary text. That includes burned-in offer text, on-screen captions, packaging, and the printed label on a vial, box, or pouch appearing in the shot.

Why it needs its own rule. Across 76 live ads from a large, established advertiser in this category, the primary text was genuinely clean: zero compound names, zero outcome or dosing language, and the research use line present in 71 of 76. And their vial labels carried the compound names throughout, NAD+, GHK-Cu, BPC-157/TB-500, and the coded blends. The identity their copy carefully avoided was sitting on the pack in the same frame.

That configuration passes a text-only review while the claim is fully visible to a human reviewer, to image OCR, and to anyone who screenshots it. It is not a clever loophole, it is an unreviewed surface.

The mechanical reason it gets missed, which is the same reason 4.11 gets missed: the person writing the copy and the person choosing the product render are usually doing two different jobs, and the compliance pass reads the copy field. Nothing in any ad tool shows you the label text as text.

The check. Before delivery, read every word legible in the frame out loud as if it were the primary text, and apply section 2 to it. If a word would be rejected in the copy, it is rejected on the pack. In practice this means unbranded product in every creative.

Do not solve it with a coded name. Replacing a compound name on a label with an in-house code is obscuring identity rather than removing a claim, and a regulator can read a code as evidence of intent. Show unbranded product.

The count is copy too. The rule reaches one level past the words on the pack: if everything in the frame is copy, then the arithmetic of the composition is copy as well. A countable set of six sealed screens asserts a six screen testing regime exactly as printing the figure would, and it fails for exactly the same reason when we cannot substantiate it. This is easy to miss because nobody experiences arranging objects as writing a number.

The fix is the off frame set. Run the set past both edges of the frame so it cannot be counted. That states breadth without asserting a figure, and it is a better answer than trimming the set down to a number we can defend, because the number is almost never the point. Reach for it whenever a composition shows a quantity of anything the client would have to prove.

2.9 Do not let the compliance line be the first thing a promotion drops

Also from the 2026-08-12 teardown. Of the same 76 ads, the five missing the research use line were all the newest promotional creative, shipped in a two-day window during a scale-up.

The research use statement is one line, it is free, and it is the cheapest evidence of good faith in the entire ad. It goes in the primary text and not only in the image, and it is not optional on offer creative. Treat its absence on a rushed promo as a QC failure, because that is exactly when it disappears.

2.10 Named phrase traps

Specific live competitor lines that are not ours to reuse, kept by name so nobody rediscovers them as good ideas. 2.5 is the founding entry.

3. The QC gate

Run this on every clip before it is delivered, every time, no exceptions for short clips or re renders. This gate exists because we have already been burned twice by generated output that looked fine at a glance: a gloved hand with fused fingers and a chromatogram with garbled axis labels. Look at every frame. A clip that fails any single line is rejected; fix and re run the full gate on the new render, because re renders introduce new faults.

Reject the clip if any of these is true:

Then, after the twelve checks:

The reviewer signs the batch sheet per clip. "It came from the same batch as a passing clip" is not a pass.

4. The disclosure policy

4.1 The position, stated once

We disclose AI generated and AI edited content correctly, everywhere, always. We never strip, launder, or work around provenance metadata or watermarks, and we never buy tools or services promising undetectable AI content. Anyone proposing otherwise stops work and raises it to the account owner.

4.2 Why this is policy and not preference

4.3 What we do in practice

  1. Every ad containing AI generated or AI meaningfully edited video, audio, or imagery is flagged as such in Ads Manager at upload, using the platform's disclosure mechanism, by whoever loads the batch. The batch sheet records the disclosure status per creative so the loader does not have to guess.
  2. Provenance metadata and watermarks pass through our pipeline untouched. Transcoding and editing are fine; steps whose purpose is removing provenance are not.
  3. Where we control disclosure wording and it is truthful, prefer the collaborative framing, on the order of "created with AI assistance", over a bare "AI generated". Research in 2026 found the disclosure penalty runs through perceived effort, and truthful collaborative wording preserved trust better. Our hybrid workflow, human scripted, human reviewed, machine rendered, makes that wording honest.
  4. Do not stage fake humanity around a disclosed AI ad. The measured penalty grows when human involvement cues are undercut by the label. Automated production framing plus the label costs little; faked authenticity plus the label costs more.
  5. The label is not a handicap to engineer around. The evidence says it taxes lazy looking creative, not AI as such. We win on craft: the QC gate above is what makes the label irrelevant.

4.4 What we never do

Strip C2PA manifests. Screenshot relaunder imagery. Re record screens to shed metadata. Use "watermark removal" tools. Misdeclare AI content as camera footage in any platform form. Represent a synthetic speaker as a real customer anywhere, including in comments or organic posts. If a platform labels something we believed was exempt, we do not fight the label; we review our disclosure, not their detector.

5. Delivery format

Creatives are delivered as structured batches, never loose files. Each batch ships as a dedicated folder containing per creative: name, hook text, primary text, headline, CTA, script with the four parts marked, the video file, disclosure status, QC sign off per section 3, and claim sources per section 2.6. Every creative is loaded PAUSED. Nothing goes live without the client compliance team's approval and the account owner's explicit written approval.

6. The record this standard is built on

Enforcement positions, platform detection, and the model field all move monthly. Review this standard against the current record before each new production cycle.