Pull every checkable claim out of the copy (numbers, product specs, prices and offer terms, comparisons and "best" or "only" claims, environmental or safety claims, reviews and quotes) and match each one to a source you control, such as the product sheet or the written offer terms. Cut or soften anything you can't source. Allow 10 to 20 minutes per piece.
Checking line by line for "anything that looks wrong" doesn't work well, because AI errors look right. They are fluent, plausible and consistent with the rest of the copy. The reliable method is to treat the copy as a list of claims, then check the list against your own records. A model can help build that list quickly. It cannot tell you which items are true, because it has never seen your stock system, your supplier certificates or your offer terms.
Step one: have the AI list every claim it made
Paste the finished draft into the same assistant, or a fresh chat, with this prompt:
List every factual claim in the marketing copy below as a table with
columns: Claim | Type (number, product fact, price/offer, comparison,
environmental/safety, testimonial/social proof, name/link, other) |
What evidence would prove it | Risk if wrong (low/medium/high).
Include implied claims (e.g. "back in stock" implies stock exists).
Do not judge whether claims are true. Do not rewrite the copy.
[paste copy]An illustrative result for a toy shop's email promoting a wooden kitchen play set:
| Claim | Type | Evidence needed | Risk |
|---|---|---|---|
| "Made from sustainably sourced beech" | Environmental | Supplier certificate or specification naming the scheme | High |
| "Suitable from 18 months" | Safety | Age warning on the packaging | High |
| "Now $89, was $120" | Price/offer | Record of the $120 price being charged, and for how long | High |
| "Our bestselling play kitchen" | Comparison | Sales figures across your play kitchens | Medium |
| "Free delivery this weekend" | Price/offer | Written offer terms, dates and exclusions | High |
| "Loved by over 500 families" | Social proof | Order count or review count | Medium |
Checked against the shop's records: the beech is from a certified source (certificate on file), but the packaging says 3 years and up, not 18 months, because of small parts. The model had borrowed a typical age for play kitchens. That single error could have been a safety complaint.
The claim checklist, by type
Numbers, dates and statistics
- Every figure traces to a record: order counts, review counts, years trading, sizes, weights. How to verify: find the source and write it next to the claim.
- Dates match the calendar: "ends Sunday" is the right Sunday, "since 2014" matches your founding year.
- No borrowed statistics: "80% of parents say..." with no survey behind it is invented. Delete it or cite a real, named source you have read.
Product facts and specifications
- Materials, dimensions, capacity, compatibility and care instructions match the product sheet or packaging, word for word where it matters.
- No features borrowed from similar products. Models often add a common feature (a lid, a removable cover, a warranty) that your version doesn't have.
- Stock claims ("back in stock", "last few") match stock levels on the send date, not the draft date.
Prices, offers and terms
- Prices match the till or shop system at the time of sending.
- "Was" prices were genuinely charged, for a meaningful period. Rules on reference pricing vary; if you use them often, ask an adviser what applies where you sell.
- Offer terms (dates, exclusions, minimum spend, one per customer) are stated or linked, and match the written terms.
- "Free" really is free: no hidden condition such as a minimum order the copy omits.
Comparisons, superlatives and "only"
- "Best", "lightest", "only", "first", "most popular", "lowest price" each need evidence you could show someone. If you can't, change it to a claim you can support.
- Comparisons with named competitors are especially risky; avoid them in AI-drafted copy unless someone has checked current competitor facts.
Environmental, health and safety claims
- "Eco-friendly", "sustainable", "non-toxic", "natural", "recyclable", "plastic-free" need specific, documented backing. Vague green claims attract scrutiny in many markets.
- Age marks, allergen information and safety warnings are copied from packaging or the supplier's documents, never generated.
- No health benefits unless you're allowed to make them and can prove them.
Reviews, testimonials and social proof
- Quoted reviews are real, attributable and quoted accurately. AI sometimes "tidies" a review into something the customer never wrote.
- No invented customers, star ratings or "as seen in" mentions.
- Award claims name the real award and year.
Names, links, trademarks and people
- Brand names are spelled and capitalised correctly, including your own product names.
- Every link goes to the right page and the page is live.
- No real person is named or quoted without their agreement.
Build a fact bank once, so checks take minutes
Most checking time goes on hunting for sources. A fact bank puts them in one document you can also paste into the AI as its only allowed source. An illustrative entry for the toy shop:
PRODUCT: Wooden play kitchen, model PK-2
Price: $89 (from 1 Oct); previous price $120 charged 1 Mar - 30 Sep
Age: 3+ (packaging warning: small parts)
Material: beech from a certified sustainable source
(certificate #, expires {date}, filed in /certificates)
Dimensions: 60 x 30 x 85 cm; assembly required, about 40 minutes
Colours: natural, sage
Not included: pans and play food (sold separately, $24)
Sales rank: 2nd bestselling play kitchen this year (not 1st)
Approved phrases: "certified sustainable beech"; "3 and up"
Banned phrases: "non-toxic" (no test report on file); "bestselling"
Last checked: {date} by {name}Note the last three lines. "Approved" and "banned" phrases stop the same argument happening every campaign, and a "last checked" date tells you when an entry has gone stale. Load the fact bank into a Project in your assistant and add "use only facts from the fact bank" to your drafting prompts, and many errors never get written in the first place.
Words that should trigger a closer look
| Word or phrase | Why it's risky | Safer if you can't prove it |
|---|---|---|
| Best, No. 1, leading | Implies a comparison you'd need data for | "Our most-ordered", backed by your own sales |
| Only, first, unique | Easily disproved by one competitor | Describe the feature itself |
| Eco, green, sustainable, natural | Vague environmental claims draw scrutiny | The specific, documented fact ("certified beech", "recycled card box") |
| Non-toxic, safe, chemical-free | Needs test evidence; "chemical-free" is never literally true | Quote the standard or test the product meets, if you hold it |
| Guaranteed, always, never | Absolute promises create obligations | State the actual guarantee terms |
| Free, save, was | Price claims are closely regulated in many markets | State conditions; check reference prices |
| Clinically, proven, studies show | Implies evidence you probably don't have | Remove |
| Handmade, made in | Origin claims are checkable and often wrong | Say what you actually do ("finished by hand in our workshop") |
A quick way to use this table: search the draft for each word before sending. Anything found needs either a line in the fact bank or a rewrite.
Five claims caught before publishing
An e-commerce homeware brand's bamboo range. The AI draft for a product page read: "Our FSC-certified bamboo utensils are 100% biodegradable." The check found the supplier had no certificate for this range and the utensils have a food-safe lacquer, so "100% biodegradable" wasn't supportable either. After: "Bamboo utensils with a food-safe lacquer finish. Hand wash only." Plainer, and every word true.
A charity shop's appeal post. Draft: "100% of every purchase goes straight to our cause." The shop pays rent, wages and utilities from sales, so this was simply false. After: "Every purchase helps fund our community meals programme. Last year the shop raised enough for about 9,000 meals." The figure came from the charity's annual report, which was added to the fact bank.
A sports equipment shop's ad. Draft: "The lightest carbon racket on the market." Nobody could prove it, and a competitor sold a lighter model. After: "Weighs 265g unstrung: one of the lightest in our range." Specific, checkable, still a selling point.
A subscription box company's email. Draft: "Cancel anytime, no questions asked." The actual terms required cancellation five days before the next billing date to avoid that month's charge. After: "Cancel anytime from your account. Cancel at least 5 days before your billing date and you won't be charged for the next box." The rewrite prevents a string of disputed charges.
A furniture maker's product page. Draft: "Solid oak throughout, built to last generations." The dining table has a solid oak frame and legs but an oak-veneered top on a stable core, chosen so the wide top doesn't split. After: "Solid oak frame and legs, with an oak-veneered top that stays flat through central heating and seasons." The maker had a good reason for the veneer; the AI's version turned it into a misdescription a customer could reasonably complain about.
Write the drafting prompt so fewer claims need checking
The cheapest fact-check is the one you don't need. When the drafting prompt restricts the model to your fact bank and tells it to mark gaps instead of filling them, the claim list shrinks and the risky items show up as placeholders:
Write a 120-word product email for the item below.
Use ONLY facts from the fact bank entry. If a sentence needs a fact
that isn't there (a price, a date, a material, a number), write
[NEED FACT: what is missing] instead of inventing it.
Never use: best, only, eco, non-toxic, guaranteed, clinically.
Fact bank entry: [paste]Illustrative output extract: "Our wooden play kitchen is back, now $89, in natural or sage. Made from certified sustainable beech and built for ages 3 and up... Free delivery this weekend [NEED FACT: offer dates and minimum spend]."
The placeholder is the whole point. The model has flagged the one missing fact rather than inventing "on orders over $50", which is the kind of detail that looks right and costs money when customers hold you to it.
Checking a week of posts in twenty minutes
Batch checking is faster than checking each post as it's drafted. A routine for a small business scheduling a week at a time:
- Two minutes: paste all the week's drafts into one chat and run the claim-list prompt on the lot.
- Eight minutes: tick each claim against the fact bank. Anything not in the bank gets a source found now or the claim removed.
- Five minutes: search the drafts for the red-flag words in the table above.
- Three minutes: check every price and date against the shop system and the calendar on the day of scheduling.
- Two minutes: add a line to the sign-off log.
If a week's batch regularly takes longer than half an hour, the fact bank is missing entries. Add them as you go, and the routine shortens each week.
Who checks, who signs off, and what goes in the log
In a small business the drafter and the checker are often the same person, which is when errors slip through. Split the roles where you can, even informally:
- Drafter: writes with AI, runs the claim-list prompt, fixes what the fact bank contradicts.
- Checker: someone else reads the claim list against the fact bank. Ten minutes, not a full rewrite.
- Approver for high-risk claims: the owner, for anything environmental, safety, health, legal or a new price claim.
A sign-off log can be one line per piece. Illustrative entries:
Date | Piece | Channel | High-risk claims | Checked by | Approved by
2 Oct | Play kitchen launch | Email, 2,100 | Age, price, beech | AB | Owner
3 Oct | Weekend delivery | Instagram | Offer terms | AB | AB
5 Oct | Bamboo range page | Website | Material, care | CD | OwnerThe log earns its keep the first time a customer queries a claim: you can see who checked it, against what, and when.
How the check differs for ads, emails and product pages
- Paid ads: the platform reviews them, but only against its own policies, not your facts. A rejected ad costs time; an approved false claim reaches thousands of people you paid to reach.
- Emails: can't be recalled. Send a test to yourself and check prices, dates and every link on your phone before scheduling.
- Product pages: stay live for months and are read by search engines and AI shopping assistants, which may repeat an error elsewhere. Put a "last checked" date on high-traffic pages in your fact bank and recheck quarterly.
- Social posts: can be edited, but screenshots last. Offer terms and prices get the full check; everything else a quick read.
Why the AI can't check itself, even with web search
It's tempting to finish with "please verify all facts above". The model will reply that it has, and often be wrong. Language models judge plausibility, not truth, and they cannot see your records. With web search on, they can find sources, but the source may not say what the answer claims. An illustrative case: an assistant asked to support "bamboo grows up to a metre a day" cited a page that actually said "some species can grow up to 91 cm in a day under ideal conditions". Close, but the ad would have overstated it. Always open the cited page and find the sentence yourself. There is a general routine for this in a five-minute fact-check routine for AI output, and the reasons models invent details are explained in AI hallucinations explained for business owners.
When a wrong claim has already gone out
Even with a routine, something will eventually slip through. What to do, in order:
- Fix the source first. Correct the product page, pause the ad or edit the post, so nobody new sees the error.
- Decide whether customers relied on it. A wrong price, offer or safety claim matters; a typo in a colour name usually doesn't.
- Correct it plainly where it mattered. A short follow-up email ("Correction: the play kitchen is for ages 3 and up, not 18 months as we said yesterday") is better than silence, and safety corrections should go to everyone who received the original.
- Honour reasonable expectations where customers acted on a price or offer you published, and take advice if the amounts are significant.
- Add the fact to the bank and the word to the red-flag list, so the same error can't be drafted again.
The toy shop in the example above would have needed step 3 for the age claim, which is why that one is worth ten minutes of checking on every piece.
When the check needs a professional
A fact bank and a second reader handle most marketing copy. Get a qualified adviser to look before publishing when copy includes health or medical benefits, financial returns or credit, claims about children's products beyond what the packaging states, environmental claims central to a campaign, comparisons naming competitors, or anything about a competitor's product. Also check the rules on testimonials if AI helped write or edit them; where the legal line sits on AI testimonials covers that. And for social posts specifically, run drafts through the pre-publish checklist for AI-drafted posts, which adds tone and format checks to the claim check here.
Fact-checking questions from marketing teams
Can I ask ChatGPT or Claude to fact-check its own copy?
You can ask it to list the claims it made, which is useful, but not to confirm they are true. The model has no access to your stock system, your offer terms or your supplier certificates, so it can only judge whether a claim sounds plausible. Even with web search switched on, check that the cited page actually says what the answer claims. Verification has to come from your own records.
How long should fact-checking a marketing email take?
For a typical promotional email with a handful of product and offer claims, 10 to 20 minutes once you have a fact bank to check against. The first few take longer while you gather sources. If checks regularly take over half an hour, the drafts contain too many claims; tighten the prompt so the AI only uses facts you supply.
Do I need to fact-check social posts as carefully as ads?
Check the same claim types, but the stakes differ. Paid ads face platform review and wider reach, and emails can't be recalled once sent. A social post can be edited or deleted, though screenshots last. Prices, offer terms, safety and environmental claims deserve the full check in every channel; tone and phrasing can get a lighter touch on social.
Further reads
- How to Catch Made-Up Figures in AI-Drafted Proposals — The same discipline applied to proposals and quotes.
- How to Check AI-Generated Images for Errors Before Posting — Checking the images as carefully as the words.
- Why Your AI Marketing Copy Sounds Generic (and How to Fix It) — Fixing flat copy once the facts are right.
- What to Do When ChatGPT Gets Facts About Your Business Wrong — When AI search gets your business wrong instead.
- Should You Tell Customers When You Use AI in Your Marketing? — Whether to disclose AI use in your marketing.
- How to Write Product Descriptions With AI That Actually Sell — Writing product copy that is accurate and sells.
- How Estate Agents Use AI to Write Property Descriptions — Build the fact sheet at the valuation, prompt in your house style, get portal, brochure and social copy from one input, and check every claim before listing.
- AI Content Workflow for Agencies: From Brief to Approved Draft — Six gates from client brief to approved draft, with the AI's job and the human's job at each, plus a filled-in brief, outline prompt and claims log.
- AI Marketing for Small Law Firms: Content, Reviews and the Rules — How a small law firm can use AI for guides, posts and review replies while staying inside platform, consumer-law and professional conduct rules.
- How Freelance Writers Use AI Without Losing Their Voice — Keep control of the angle and first draft, then use targeted prompts to find weak passages without flattening your writing.
- AI Listing Mistakes That Can Mislead Buyers, and How to Check — Check room counts, measurements, parking rights and altered photographs against evidence before publishing an AI-written property listing.
- How Course Creators Use AI to Build a Course in Weeks, Not Months — A week-by-week plan: AI handles outlines, scripts, edits and quizzes; you supply the expertise, examples and recording. Six weeks for a 6-10 hour course.
- Turning One Online Course Into Emails and Posts With AI — Transcribe the lessons, pull out one idea, mistake and quick win per lesson, and turn each into an email or post. Twenty lessons can fill two months.
- Is It Cheaper to Use AI or a Freelancer for Marketing Content? — A per-piece cost comparison of AI and freelance writers for marketing content, counting your own editing time, with three worked cases.
- What to Prepare Before You Set Up AI Marketing — An eight-part preparation checklist, filled in for a farm shop, with a fact-sheet prompt and a one-page brief you can copy.
- How to Write Website Articles With AI That Still Rank on Google — A step-by-step method for AI-assisted website articles that rank: evidence packs, outline and draft prompts, the edit pass, and what to check in Search Console.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: No vendor prices are quoted. Examples are illustrative; claim categories reflect general advertising-substantiation principles, not any one jurisdiction's rules.