Portfolio pages fail in two opposite ways: they are so vague they persuade no one, or they are so punchy they overclaim. AI makes the second failure cheaper. Ask for “a case study that sounds impressive” and you will get percentage lifts, glowing quotes, and category leadership language with no source trail.
Freelancers sell trust. Invented proof burns it twice - once with the prospect who checks, and again with the past client who recognizes their project dressed in fiction.
Treat portfolio claims like advertising claims
In the U.S., advertising claims must be truthful and substantiated (FTC advertising and marketing guidance). Endorsements and testimonials have their own guides (FTC Endorsement Guides: what people are asking; regulatory text at 16 C.F.R. Part 255). There is no special exemption because a model wrote the sentence (FTC crackdown on deceptive AI claims and schemes). In the EU, misleading actions and omissions toward consumers are restricted under the Unfair Commercial Practices Directive.
You may sell only to businesses. The practical standard is still useful: if you would not defend the claim on a call with the former client on the line, cut it.
Do not publish AI-generated testimonials, metrics, or “results” that you cannot document. Fabricated social proof is a claims problem, not a writing problem.
Evidence pack before prose
Gather, offline:
- Client permission to describe the work (email is enough for many cases; some NDAs forbid any public mention - obey the stricter rule)
- One-sentence problem statement in your words
- Constraints (timeline, stack, brand rules)
- Deliverables actually shipped
- Outcomes you can prove (screenshots, analytics exports you are allowed to share, before/after artifacts)
- What you will anonymize (company name, revenue, personal data)
If permission is missing, stop. Write a hypothetical process sample labeled as illustrative instead of a fake case.
Strip personal data about the client’s customers and staff. Anonymize competitively sensitive figures unless permission explicitly covers them. Do not paste full contracts into a consumer model to “extract a case study.”
Workflow
Step 1: Fill an evidence sheet yourself
Use the fields in the portfolio evidence checklist. No AI yet.
Step 2: Structure-only generation
Using only the evidence sheet below, draft a 200-300 word case
writeup with sections: Context, Constraints, What I delivered,
Evidence of outcome, Limits / what this does not prove.
Do not invent metrics, quotes, logos, awards, or client names.
If evidence is missing for a section, write "not evidenced" instead
of filling it.
Evidence sheet:
[paste]
Step 3: Invention audit
Highlight every number, comparative (“3x”, “industry-leading”), and quotation. Each highlight needs a source in the evidence sheet or it goes.
Step 4: Permission audit
Confirm the public version matches what the client allowed: named vs anonymized, metrics yes/no, screenshots yes/no.
Step 5: Authorship accountability
You remain the accountable author of the portfolio claim (keep your name on AI-assisted work). If AI helped polish, that does not reduce your duty to substantiate.
Keep a private “evidence” folder per case (permission email + screenshots). The public page should be regenerable from that folder without opening a chat history.
Illustrative scenario (labeled)
Illustrative scenario, not a measured case: A freelance SEO writer asks a model to “punch up” a case study. The draft adds “helped increase organic traffic 140% in three months.” The writer never measured traffic. A prospect asks for the analytics screenshot on a sales call. The writer has nothing. The lost deal is the mild outcome; a public complaint is the worse one.
Anonymized cases still need evidence
Anonymizing the client name does not license invented metrics. “A Series B SaaS company” with a fabricated conversion lift is still a false claim. Anonymization protects identity; it does not create proof. If you lack measurable outcomes, write about process quality, constraints handled, and artifacts you can show.
When NDAs block public cases
Some NDAs forbid public mention entirely. Options that stay honest:
- Ask for a narrow written exception (anonymized, no metrics)
- Use a labeled illustrative walkthrough of your method with fictional data clearly marked
- Request a private reference call instead of a public page
Do not “approximate” a forbidden case until it becomes unrecognizable fiction presented as fact.
Validation and fallback
Validation: every claim maps to a file in the evidence folder.
Fallback: publish a process breakdown (“how I approach X”) with no client story until permission and proof exist.
Tie this to proposal drafts so sales decks reuse only the same permissioned proof, not a second, looser narrative.
