List every factual claim in the AI draft, attach evidence you hold to each one (a dated report, a certificate, survey data, client figures with permission), and rewrite or cut any claim the evidence doesn't support as a customer would read it. File each claim with its evidence, so you can produce it if challenged.
Timing is where most businesses go wrong. Advertising rules in the major markets expect you to hold the evidence before a claim is published, not to go looking for it after a complaint arrives, and they judge the claim by what a reader would take it to mean, including what it implies. "Trusted by leading brands" implies you can name some, and that they'd agree.
Why AI drafts arrive full of claims you can't prove
A language model writes what marketing copy usually looks like, and marketing copy is full of social proof, superlatives and round numbers. Ask for an About page with no facts attached and it supplies them. Here is a prompt an illustrative eight-person PR consultancy might use for its own website, and a realistic output:
Write a punchy 60-word About section for our PR consultancy. We work
with technology and consumer brands. Make us sound credible.Founded in 2014, we're an award-winning PR consultancy trusted by over 200 leading brands. Our data-driven approach delivers an average 300% uplift in media coverage, and our team of former journalists has secured front-page stories in every major newspaper. We're the fastest-growing independent agency in the sector, and we guarantee results.
That's nine claims in about sixty words, and the consultancy can support almost none of them as written:
- "Founded in 2014". The company records say 2015. Wrong.
- "Award-winning". Shortlisted once, won nothing. Unsupported.
- "Over 200 leading brands". 64 clients in eleven years, and "leading" is undefined. Wrong and vague.
- "Data-driven approach". Implies measurement; true only if you can show the reports.
- "Average 300% uplift". Invented. No such calculation exists.
- "Team of former journalists". Three of eight staff. Misleading by implication.
- "Front-page stories in every major newspaper". An absolute claim, and false.
- "Fastest-growing independent agency". A ranking with no source.
- "We guarantee results". No PR firm controls what journalists publish. Cut.
Finding claims like these is the first half of the job, and the routine for pulling them out of any draft is set out in the routine for fact-checking AI marketing copy. This tutorial is about the second half: what evidence each claim needs, how to hold it, and how to write claims that stay inside it.
What counts as evidence, claim type by claim type
Evidence has to match the claim as a reader understands it. A single great campaign is evidence of a single great campaign, not of an average. A shortlisting is evidence of a shortlisting. The table below is the reference the consultancy's checker works from.
| Claim type | What the AI wrote | Evidence that holds up | Evidence that doesn't |
|---|---|---|---|
| Facts about your business | "Trusted by over 200 brands" | Client list with start dates, company records | A rough memory of "loads of clients" |
| Results for clients | "Average 300% uplift" | Per-client coverage logs with a defined before and after period, plus permission | One exceptional campaign presented as typical |
| Superlatives and rankings | "Fastest-growing agency" | A named, dated third-party ranking you appear in | Your own estimate, or a ranking from three years ago |
| Comparisons | "Twice as responsive as other agencies" | A like-for-like, dated comparison with its method written down | Competitors' own claims you can't check |
| Awards and accreditations | "Award-winning" | The winner's letter: award, category, year | A shortlisting, or an award a former employee won elsewhere |
| Testimonials | "They transformed our profile" | The client's original words, written approval, contact details | A quote paraphrased or polished by AI after approval |
| "Up to" figures | "Up to 50 pieces of coverage per launch" | The spread across launches, showing a meaningful share come close | A single best case |
| Research findings | "73% of journalists prefer email pitches" | The survey itself: who was asked, how many, when, the exact question | A figure the AI produced, or an article quoting an article |
| Environmental | "Our eco-friendly packaging" | Specific, certified figures for the specific thing claimed | A general sense of being green, or offsets behind a "neutral" claim |
| Guarantees | "We guarantee results" | A written guarantee with terms you will honour | Nothing; cut the claim |
Three rows need a closer look. Testimonials sit under their own rules: advertising codes expect you to be able to show a testimonial is genuine, and several major markets have banned fake reviews and testimonials altogether, with at least one regulator's rule singling out AI-generated fakes. Where the legal line sits on AI-written testimonials covers that ground in detail. Research findings are where AI is most dangerous, because a fabricated statistic looks exactly like a real one; the "73%" above is the kind of figure a model will produce on request. And environmental claims changed recently. If you sell to customers in the EU, rules that apply from 27 September 2026 ban generic environmental claims such as "green", "eco-friendly" or "climate friendly" unless you can show recognised excellent environmental performance, ban claims that a product is neutral or reduced-impact because of carbon offsetting, and ban claims about a whole product or business when only one part qualifies.
Implied claims: what the reader hears, not what you wrote
The claims that catch businesses out are often the ones nobody typed. "Our team of former journalists" doesn't say "everyone here is a former journalist", but that's what a reader hears. Logos under "Trusted by" imply current clients who are happy to be named. "Senior-led" implies that seniors do the work, not just the pitch.
AI is useful here, because it can read copy the way a sceptical stranger would. A prompt that works:
Read the copy below as a sceptical potential client. List every claim
it makes, including claims it only implies. For each one, write:
- what a reader would assume is true
- what evidence would prove it
Don't judge whether the claims are true; just list them.
[paste copy]An illustrative excerpt of what comes back for the consultancy's homepage:
1. "Our team of former journalists"
Reader assumes: most or all consultants are ex-journalists.
Evidence needed: staff list showing their backgrounds.
2. Logo strip under "Trusted by"
Reader assumes: these are current clients who endorse the agency.
Evidence needed: client status and permission to show each logo.
3. "Senior-led campaigns"
Reader assumes: senior staff do the day-to-day work.
Evidence needed: how accounts are actually staffed.
4. "Coverage in the titles that matter"
Reader assumes: regular top-tier coverage for typical clients.
Evidence needed: coverage breakdown by outlet tier.The consultancy checked the logo strip and found two of the nine logos belonged to clients who had left, one of them on bad terms. Nobody had written anything false; the page just kept implying something that had stopped being true. Implied claims deserve a row in the register like any other.
A substantiation file with one row per claim
A substantiation file (or claims register) is a table of every claim you're willing to make, the exact wording, and where the proof lives. It turns "can we say this?" from a debate into a lookup. The consultancy's register, filled in after its website clean-up:
ID Claim, exact wording Evidence Evidence date Review by
C-01 "Founded in 2015" Incorporation certificate 2015 never
C-02 "412 pieces of client coverage in the Monitoring-tool export, filtered 31 Aug 2026 30 Nov 2026
12 months to August 2026" by date, saved as PDF
C-03 "Shortlisted, [award], Tech Campaign Organiser's shortlist email Mar 2025 Mar 2027
of the Year 2025" (drop after 2 yrs)
C-04 "Three of our eight consultants are Staff list with prior roles Sep 2026 each hire or leaver
former journalists"
C-05 Testimonial, head of marketing at a Approval email with exact words, Jun 2026 Jun 2027, or if they
software client contact details on file stop being a client
C-06 "Monthly mentions went from 4 to 17 a Client's coverage report and Jul 2026 Jul 2027
quarter in year one" (software client) written permission
C-07 "We reply to journalist requests Inbox timestamps, 40 requests: Aug 2026 quarterly
within two working hours" median 47 minutes, 38 of 40 under
two hours
C-08 Logo strip: 7 current clients Permission emails, client status Sep 2026 monthlyEach row names an owner in the full version, and each evidence file sits in one shared folder named by ID, so C-07's timestamps are in a folder called C-07. Note what C-07 shows about evidence: "within two working hours" is supported because 38 of 40 sampled requests met it, and the sample is saved. If only 25 of 40 had, the claim would need rewording, perhaps to "usually within the same morning".
The register also becomes the drafting source. Instead of asking the AI to "make us sound credible", you give it the register and forbid anything outside it:
Write the About section for our website, about 80 words.
Use only claims from the register below, word for word or shortened
without changing their meaning. If the section needs a claim that
isn't in the register, write [NEEDS EVIDENCE: the claim] and carry on.
No superlatives (best, leading, fastest, first, only) unless they
appear in the register.
[paste register rows C-01 to C-08]An illustrative result: "Since 2015 we've run PR for technology and consumer brands, securing 412 pieces of client coverage in the 12 months to August 2026. Three of our eight consultants are former journalists, and we reply to journalist requests within two working hours. [NEEDS EVIDENCE: 'most of our clients stay for more than two years'] We were shortlisted for Tech Campaign of the Year 2025." The flagged sentence is the AI reaching for a retention claim; the consultancy either pulls the client data to support it or deletes the sentence.
Rewriting claims down to the evidence you hold
When evidence falls short, you have three choices: find better evidence, narrow the claim until the evidence covers it, or cut it. Narrowing is usually the best of the three, because specific claims persuade better than vague ones anyway. Six rewrites from the consultancy and its clients:
- Before: "Award-winning PR consultancy." After: "Shortlisted for Tech Campaign of the Year 2025."
- Before: "Trusted by over 200 leading brands." After: "64 clients since 2015, from two-person start-ups to listed software companies."
- Before: "Our data-driven approach delivers an average 300% uplift." After: "For one business software client, mentions went from 4 to 17 a quarter in our first year together (their figures, shared with permission). Results depend on the news and the budget."
- Before: "Our team of former journalists." After: "Three of our eight consultants are former journalists."
- Before: "We guarantee results." After: "We don't guarantee coverage, and we'd be wary of anyone who does. We do commit to a target list agreed with you, follow-ups within 48 hours and a coverage report every month."
- Before, in a consumer client's release: "The most sustainable packaging in the category." After: "Packaging made from 80% recycled cardboard, certified by [scheme]." The client supplied the certificate; the superlative went because nobody had compared the whole category.
The guarantee rewrite shows a useful move: when a claim can't be supported, replace it with a commitment you control. Readers can't verify "we guarantee results", but they can hold you to "a coverage report every month".
Client press releases: getting the client to prove it
A PR consultancy drafts claims on behalf of clients, which splits the responsibility. The claims are the client's, but the consultancy's name is on the relationship with every journalist who receives the release, and journalists check "first", "only" and "largest" in minutes. The practical answer is a claims confirmation email for every release, sent as a separate item rather than buried in the full draft. For drafting the release itself, see how to write a press release with AI that journalists will read.
Subject: Claims in the launch release: evidence needed by Wednesday
Hi [first name],
Before the release goes out, we need you to confirm the evidence for
each claim below. Please reply against each number.
1. "The first scheduling app built for dental practices"
Evidence: how you know no earlier product exists (search notes,
market research). If unsure, we suggest "built for dental practices".
2. "Used by 1,200 practices"
Evidence: active account count and the date it was taken.
3. "Cuts admin time by up to 6 hours a week"
Evidence: the data behind it. How many customers saved close to
6 hours? If only a few, we suggest quoting the typical figure.
4. Quote from [founder's name]
Please confirm the exact wording, attached separately.
Anything without evidence by Wednesday comes out of the release.
[first name]The first item is there for a reason. Picture a release announcing "the first AI-powered scheduling assistant for dental practices", drafted by AI and waved through. A trade journalist searches for two minutes, finds a competitor that launched something similar the previous year, and the story becomes a paragraph about the claim rather than the product. The client is annoyed with the journalist, then with the agency. "Built for dental practices" was true, and would have been enough.
The same discipline applies to results figures in case studies. If you're turning client interviews into published stories, writing case studies from customer interviews with AI covers getting the figures and the permissions in one pass, and catching made-up figures in AI-drafted proposals covers the same problem when the audience is one prospect rather than the public.
Evidence goes stale: dates, renewals and review
A claim that was true in March can be false by September without anyone touching the copy. Every row in the register needs a review date, set by how fast its evidence decays:
- Counts and totals (clients, coverage, customers): review quarterly and restate with an "as of" date or a fixed period, as C-02 does.
- Awards: state the year, and drop them after two years unless they're genuinely significant.
- Certifications and accreditations: review by the expiry date on the certificate, not the date you added it.
- Testimonials and logos: review when the client relationship changes, and at least yearly. A glowing quote from a client who has since left is a liability.
- Comparisons and competitor prices: date every comparison, and recheck before each campaign that uses it. Competitors change prices without telling you.
- Third-party statistics: check for a newer edition of the report before reusing a figure.
Set a recurring calendar entry and give the review to whoever owns the register. At the consultancy's size, a quarterly pass through 20 to 30 rows takes about 45 minutes, and any row past its date is pulled from live copy until it's refreshed.
When someone challenges a claim
Sooner or later a competitor, a customer or a journalist asks you to justify something. The register is what makes that a calm afternoon rather than a scramble. Work through it in this order:
- Find the row. If the challenged wording isn't in the register, that's your first finding: the claim went out without evidence, and the honest move is to pause it now.
- Read the evidence against the claim as the challenger reads it. "Reply within two working hours" backed by 38 of 40 requests is defensible. The same wording backed by a one-week sample from a quiet month is not.
- Pause first if in doubt. Taking a claim down while you check costs little. Leaving a weak claim up while you argue costs credibility, and possibly more.
- Change it everywhere at once. Claims spread into proposals, directory listings, social bios and pitch decks. Add a "where used" note to each row so a change reaches every copy, not just the homepage.
- Reply with the evidence, briefly. "The figure comes from our coverage log for the 12 months to August 2026; here's a summary" ends most challenges. A formal complaint to a regulator or a letter from a competitor's lawyer is different: that's the point to bring in a lawyer before you reply.
Afterwards, add a line to the row noting the challenge and the outcome. If one claim type keeps drawing questions, that's usually a sign the wording promises more than the evidence, and the fix belongs in the drafting prompt, not just the page.
The consultancy's own website, claim by claim
For the full picture in numbers, here's the illustrative clean-up of the consultancy's five AI-drafted pages (Home, About, Services, two case studies):
- Claims found: 31, including 7 implied claims surfaced by the sceptical-reader prompt. Time: 40 minutes.
- Supported as written: 12. Evidence located and filed: 1 hour.
- Narrowed to fit the evidence: 11, like the rewrites above. Evidence gathered and rewrites done: 2 hours 30 minutes, most of it pulling coverage exports and chasing two clients for permission.
- Cut: 8, including the 300% average, the ranking, the guarantee and two departed clients' logos.
- Register built: 23 rows. Time to format and file: 30 minutes.
That's about 4 hours 40 minutes once. After it, a new proposal or press kit drafted from the register needs 10 to 15 minutes of claim checking instead of an hour, because most of its claims are already rows with evidence behind them. At an internal cost of $60 an hour, the clean-up cost about $280, which is less than a morning of senior time spent answering a single complaint about a claim you can't support. The less obvious gain is in the copy itself: the rewritten pages carry specific numbers and dates where the AI draft had adjectives, and that tends to be more persuasive to the clients the consultancy actually wants.
Further reads
- How to Write an AI Content Policy for Your Marketing — Make 'no evidence, no claim' part of a written policy.
- AI Content Approval Workflow: Draft, Check, Sign Off — Fit the evidence check into a draft, check and sign-off flow.
- How to Write Facebook and Instagram Ad Copy With AI — Write ad copy with AI that stays inside what you can prove.
- How to Catch Outdated Information in AI Answers — Spot claims that were true once and have since gone stale.
- How to Check Sources and Citations in AI Research — Verify third-party statistics before they enter the register.
- How to Keep Your Brand Safe When AI Places Your Ads — Keep claims and placements safe when AI runs your ads.
- 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.
- 12 AI Marketing Mistakes Small Businesses Make (and the Fixes) — Twelve ways AI marketing goes wrong in small businesses, how each one shows up, and the specific fix and check that stops it happening again.
- Which Marketing Tasks Should a Small Business Never Hand to AI? — Fourteen marketing tasks where AI may draft but must never decide or publish, each with a real-looking failure, plus a sign-off policy you can copy.
- How to Audit Your Website for AI Content That Needs Fixing — A content audit for sites with AI-drafted pages: inventory, leftover searches, a five-check score, keep-fix-merge decisions and a translation agency's numbers.
- Social Media Pre-Publish Checklist for AI-Drafted Posts — A grouped pre-publish checklist for AI-drafted social posts, with why and how to check each item, platform limits, AI labels and a worked example.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: published advertising substantiation guidance from advertising regulators (evidence held before publication; claims judged as consumers read them; testimonials must be genuine); consumer-protection rules on fake reviews; Directive (EU) 2024/825 on consumers and the green transition, applying from 27 September 2026, and law-firm summaries of its banned practices.