ChatGPT Fact Checks Itself
Have ChatGPT check and refine its own sources for improved accuracy.
Use This When
Campaign planning, content calendars, ad creative, copy tests, hooks, CTAs.
Inputs Needed
Offer, audience, pain points, proof, tone, CTA, objections, channel, length limits.
Expected Output
Copy variants organized by hook, body, proof, objection handling, CTA, and recommended test priority.
The Workflow Prompt
You are a direct-response copywriter and conversion strategist. Objective: ChatGPT Fact Checks Itself Context: Have ChatGPT check and refine its own sources for improved accuracy. Original task: Task: - [Write a 500 word article on the benefits of magnesium for sleep].Phase 1 - Draft: Produce the summary and list sources beside each claim.Phase 2 - Critique: Re-examine your own draft and show me your actions below. Flag any sentence whose source is weak, older than 2018, or not open-access. If ≥ 1 sentence is flagged, rewrite ONLY those sentences with stronger evidence or delete them.Return both phases in Markdown with headings DRAFT and FINAL. If no reliable data found, answer “INSUFFICIENT EVIDENCE”. Inputs I may provide: Offer, audience, pain points, proof, tone, CTA, objections, channel, length limits. Operating instructions: - First, restate the objective in one clear sentence. - If critical information is missing, ask up to 5 focused questions. If there is enough information to proceed, make practical assumptions and label them. - Use a Standard response style. - Be specific to the business, audience, channel, and constraints provided. - Avoid generic AI advice. Give concrete recommendations, examples, templates, copy, or steps I can use. - When current facts, competitors, laws, prices, policies, or market claims matter, use current research and cite sources. - Do not expose hidden chain-of-thought. Provide a concise rationale or decision summary instead. - End with a short QA checklist that helps me verify the output. Required output: Copy variants organized by hook, body, proof, objection handling, CTA, and recommended test priority. Caution: Avoid generic output; require concrete examples, assumptions, and next steps.
QA Follow-Up Checklist
After the AI returns its output, verify against:
- Output is specific to the provided business/context.
- Assumptions are clearly labeled.
- No unsupported claims without source checks.
- Next actions are clear and usable.
- Hook, offer, audience, proof, objection, and CTA are addressed.
Follow-Up Prompt
Now turn the result for 'ChatGPT Fact Checks Itself' into a client-ready version: tighten wording, remove fluff, add missing assumptions, and provide the next 3 actions.
Avoid / Cautions
Avoid generic output; require concrete examples, assumptions, and next steps.
How Different Verticals Use This Workflow
Restaurant & Hospitality
A boutique winery in the Finger Lakes drafts a 500-word post on the sleep benefits of moderate red wine consumption. The fact-check phase flags 4 of 9 claims as based on pre-2018 studies and one as referencing a since-retracted paper. The rewrite removes 3 sentences entirely and softens 1 into 'some research suggests'. Final post avoids the alcohol-health-claim regulatory issue that would have triggered if they'd published the first draft. Saves them from a potential cease-and-desist.
Retail & E-commerce
A supplement brand writes a product page for an ashwagandha extract. The fact-check phase flags 60% of stress-reduction claims as either single-study or industry-funded research. The rewrite produces FTC-compliant language ('preliminary research suggests' rather than 'clinically proven'). Brand avoids a Consumer Affairs complaint that ate their lookalike competitor's quarter. The 'INSUFFICIENT EVIDENCE' fallback fires on adrenal-fatigue claims, which they remove entirely.
Professional Services & B2B
A B2B marketing agency drafts a thought-leadership post claiming '67% of buyers make decisions before talking to sales.' The fact-check phase reveals the original source is a 2011 Forrester study widely misquoted. The rewrite replaces the stat with a 2024 Gartner study with the actual number (44%). The post publishes with a correctly attributed citation and the agency's credibility holds when a prospect's CMO pushes back on the data in a sales call.
Beauty & Personal Care
An esthetician drafts a blog post on retinol's collagen-boosting effects. The fact-check phase flags 5 sentences citing 'studies show' without specifying. The rewrite forces named studies (Kligman 1986 and a 2019 follow-up) and removes one paragraph entirely about microbiome impact (insufficient evidence). The post passes a client's compliance review and runs on the medspa's site without legal review delay.
Local & Trade Services
A general contractor writes a homeowner guide on whether spray foam insulation off-gasses. The fact-check phase catches a 2008 source treated as current and an industry-association source treated as neutral research. The rewrite cites a 2022 EPA review and notes the off-gassing window honestly. Avoids a small but real liability — the contractor's previous draft would have minimized a real customer concern.
Frequently Asked
Why use this instead of just running ChatGPT with web search on?
Web search retrieves sources. This prompt forces a self-critique step that catches weak citations even when the source exists. The model will happily cite a 2014 nutritional study and treat it as current. The fact-check phase forces it to flag sources older than 2018 or behind paywalls and rewrite. Use web-search-only for fact gathering. Use this prompt when the output will be published with your name on it and a stale source could damage credibility. Magnesium-for-sleep is a real example where 80% of online sources are weak.
Is this safe to use on medical or financial content?
It's a guardrail, not a license. The prompt reduces obvious citation problems but doesn't make AI content compliant for regulated topics. If you're publishing health, finance, or legal claims, treat the output as a first draft to hand to a credentialed reviewer. The 'INSUFFICIENT EVIDENCE' fallback in the prompt is the safety valve — if you ignore it and publish anyway, the prompt didn't protect you. For pharma or medical device content, don't use AI at all; the FDA scrutiny isn't worth the time saved.
What's the most common failure mode here?
The model self-critiques softly. It flags 2 sentences when there should be 7. Force it to flag a minimum percentage ('flag at least 30% of sentences for review') or it'll rubber-stamp its own draft. Second failure: the rewrite phase quietly replaces a weak source with a weaker one (a Reddit thread instead of a 2016 paper). Demand that rewrites cite open-access peer-reviewed sources with URLs, and read every replacement. AI fact-checking only works if you verify the verifier.
How is this different from running the draft through a separate prompt for review?
A two-prompt workflow gets you the same draft you started with because you forgot the original spec by the time you wrote the review prompt. This prompt forces the model to hold draft and critique in the same context window so the rewrite addresses the exact claim, not a remembered approximation. The single-prompt structure also makes it impossible to skip the critique step — which is what happens 70% of the time when you split it into two prompts and run out of time.