How should eco DTC use ChatGPT for ads without inventing green claims?
Quick answer
Treat ChatGPT, Claude, and Gemini as rearrangers of an allowed-claims kit, not as scientists. Freeze kit lines and a banned-prompt list before anyone opens a chat. The model may shorten, combine, or restyle approved lines for a named ad format. It may not invent percentages, unnamed studies, plastic-free, carbon-neutral, or competitor bash. A human runs the QA gate before Ads Manager. Platform remix lives in Advantage+ brand safety. Weekly creator batches live in the UGC system. After Meta already rejects, use the rejection playbook.
Vs other AI and claims guides
| Guide | Job |
|---|---|
| Advantage+ brand safety | Meta remixes assets you already uploaded. Platform AI, not a ChatGPT paste |
| UGC creative system | Weekly creator briefs, rights, and fatigue. Not a chatbot workflow |
| Meta eco rejections | After Ads Manager already flags the ad |
| Support claims macros | CS paste answers. Same kit, different desk |
| Substantiation folder | Where proof lives. This guide is how you feed that proof into an LLM |
| AI marketing tools | Stack overview. Not a paste-and-publish gate |
| This guide | Human plus LLM ad drafts before Ads Manager |
Freeze the kit before you open the chat
| Paste in | Never ask |
|---|---|
| Allowed-claims kit: every paid line plus the evidence filename | “Make it more eco / greener / premium sustainable” |
| Banned list: plastic-free, toxin-free, carbon neutral, net zero, eco-friendly as a trait, studies show | “Add a statistic” or “what would a study say” |
| Folder index: filenames only, from the substantiation folder | “Cite a university or lab we work with” when the file is not listed |
| Format: Meta primary, headline, Google RSA, Demand Gen, or UGC hook | “Compare us to fast fashion / conventional detergent / plastic bottles” unless that comparison is a kit line |
| The draft, marked untrusted | An empty chat with only the product URL and “write ads” |
If the kit cannot support the format, stop. Do not let the model pad with generic planet copy. Score a shaky number in the green claims calculator before it ever reaches a prompt.
How to prompt (steal this sequence)
- Fresh chat. Paste kit, banned list, and folder index first. Do not rely on memory from last week’s thread.
- Name the format and the job (prospecting vs replenishment). Retention ads do not get new environmental claims.
- Ask for rearrangement only: “Use only kit lines A, C, and F. Shorten for Meta primary. Do not add claims.”
- Run the QA gate prompt on the draft. Require PASS/FAIL, a claim inventory, and a kit-only rewrite.
- Human gate before Ads Manager. If FAIL, rewrite from the kit. Do not “tone it down” in a third chat.
Failure modes (what the model invents)
| Invention | What it looks like | Fix |
|---|---|---|
| Invented % | Rounds a lab metric, annualises a sample, or turns “contains X” into “X% less impact” | Numbers must match kit unit and denominator |
| Unnamed studies | “Studies show,” “clinically proven,” “research suggests” | Name the study only if the kit names it |
| Banned superlatives | Plastic-free, toxin-free, carbon neutral, 100% sustainable | Delete. Do not soften into “practically plastic-free” |
| Competitor bash | Fast-fashion smear or “unlike conventional bottles” | Only if that comparison is an approved kit line with evidence |
| Fake proof files | Cites lca-2024.pdf that is not in the folder index | Treat as FAIL. Update the index or drop the cite |
| Fake offer | Invented waitlist, drop, or charity partnership | Offers are a human input, not a model flourish |
EU shoppers still sit under EmpCo. Unscoped “eco-friendly” in a ChatGPT headline is the same risk as typing it. See EU green claims.
Human QA before Ads Manager
- Every environmental, health, or superiority phrase maps to a kit line in the inventory.
- Landing page already says the same thing. Ad-only claims are how rejections start. See Meta rejections.
- No banned words in primary text, headline, description, or image overlay.
- Tone, offer, and audience are human decisions. The model does not pick a 20% code.
- If you will feed the line into Advantage+, freeze it in the platform brief too. That is the brand-safety job, after this gate.
Go / no-go and how to measure
| Call | When |
|---|---|
| No-go | No kit, no folder index, or people are prompting from the homepage hero |
| Go | Kit is frozen, QA prompt ran, a human signed the inventory, LP matches |
| Partial go | US-only until EU lines are scoped; or static ads only, no Advantage+ expansion yet |
Count invented-claim incidents, time-to-publish after FAIL, and live ads that drifted from the kit. Speed is not a KPI. CS still needs the same phrases in macros or they will re-invent what you just killed.
The model is a rearranger
ChatGPT will happily write a better green story than you can prove. That is the bug. Your job is to starve it of invention.
Frequently asked questions
How should eco DTC use ChatGPT for ads without inventing green claims?
Freeze an allowed-claims kit and banned prompts. Ask the model only to rearrange kit lines. Run the QA gate, then a human check, before Ads Manager. Do not ask it to invent percentages, studies, or plastic-free.
Can ChatGPT add a statistic if it sounds conservative?
No. If the number is not in the kit with the same unit, it is invented. Conservative rounding is still a new claim.
Is this the same as Advantage+ brand safety?
No. This is a human pasting a chatbot draft. Advantage+ remixes assets already in Ads Manager. Run this gate first, then platform guardrails.
What if the draft already went live and Meta rejected it?
Stop generating more variants. Use the rejection playbook, then bring the rewritten line back through this gate.
How is this different from the substantiation folder guide?
That guide is how you store proof. This guide is the ChatGPT workflow that is allowed to read that proof and is forbidden to invent the rest.