How to Verify AI-Written Facts Before You Publish: A Source Ledger Method, Tested on a Real Conflict
To verify AI-written facts, split the draft into single claims, find the strongest primary source for each, record the source and date in a ledger, and label every claim Verified, Hedge, Conflict open, Stale or Unsupported. Publish only verified claims as fact, hedge the rest, and hold anything unresolved.
In this guide
- Check the claims that carry weight first: numbers, dates, rules, quotes, and anything about money, health or law
- Even a company’s own pages can disagree. In my example, a pricing page, a help article and its snapshots told different stories
- A conflict you can’t resolve is a result. Publish a hedged sentence or hold the claim
- Record the date you checked. Facts that were right last year can be wrong today
Why do AI drafts need a verification step?
AI tools write fluent text whether or not the underlying fact is right. They can state outdated figures, blend conflicting sources into one confident sentence, or name a source that doesn’t say what they claim. Fiverr’s own AI guidelines list “AI-generated hallucinations” among the quality problems it reviews in disputes, which shows how seriously platforms treat the issue.
For a publisher the risk is practical. A wrong fee, rule or date misleads readers, damages trust and weakens your case for search and AdSense quality. Verification isn’t an extra. It’s the part of the work a chatbot can’t do for you.
What is the five-step source ledger method?
A source ledger is a simple table with one row per claim. It forces you to write down where each fact comes from and how sure you are.
- Extract claims. Split the draft into single-fact statements. “Gumroad charges 10% + $0.50 and includes processing” is two claims. Separate them.
- Classify each claim. Number, date, rule, quote, comparison or opinion. Numbers, dates and rules need primary sources. Opinions need none.
- Find the strongest source. Start with the company, regulator or original document. Use the tiers below.
- Record evidence. Note the link, the date you checked, and whether the source supports, contradicts or only repeats the claim.
- Assign a status and wording. Choose one of the five labels and write the sentence you can safely publish.
The five status labels
| Status | Meaning | What to publish |
|---|---|---|
| Verified | A primary source states it, and you checked it on a stated date | State it as fact, with the date |
| Hedge | Only secondary sources, or a figure that varies by situation | Say “reported” or “about”, and name the variable |
| Conflict open | Credible sources disagree and you can’t settle it | Hold, or tell readers to check the current page |
| Stale | The source is old or was superseded | Don’t use it |
| Unsupported | No source found, or the AI cited one that doesn’t say it | Cut it |
Which sources should you trust most?
| Tier | Source type | Watch for |
|---|---|---|
| 1 | The primary document: a company’s own help or policy page, a regulator, the original study | Snapshots and cached copies may differ from the live page |
| 2 | Independent reporting or analysis that links the primary source | Check that it quotes the current version |
| 3 | Competitor or vendor blogs | They benefit from a particular reading |
| 4 | Aggregators, forums and AI answers | Often repeat each other, which looks like agreement but isn’t |
Ten pages repeating one claim aren’t ten confirmations. Count independent sources, not copies.
Worked example: do Gumroad’s fees include card processing?
This is a real claim set from an article I was writing. It looks simple: “Gumroad takes 10% + $0.50 per direct sale.” Here is what happened when I put each claim in the ledger.
| # | Claim | What I found | Status | Safe wording |
|---|---|---|---|---|
| 1 | Gumroad charges 10% + $0.50 on direct sales | Gumroad’s own help article states it. Third-party guides agree | Verified | Gumroad lists 10% + $0.50 per direct sale (help article, checked Oct 2026) |
| 2 | That fee includes card processing | The help article says it does not include credit card processing or PayPal fees. Some sources describe the pricing page’s headline as if it were all-in | Verified: it does not | Card processing is charged on top of the 10% + $0.50 |
| 3 | Processing adds 2.9% + $0.30 | One version of the help article lists that amount. The live page text I fetched showed only the headline. Other sources use standard Stripe rates | Hedge | Card processing is about 2.9% + $0.30 for typical US cards and varies |
| 4 | Discover sales cost 30% with processing included | Help article snapshots say a flat 30%, including all processing | Verified | Sales through Gumroad’s marketplace carry a flat 30%, processing included |
| 5 | Gumroad refunds none of its fees on refunds | Two versions of the same help article disagree (details below) | Conflict open | Check Gumroad’s current refund wording before relying on it |
| 6 | Gumroad charges 5% + $0.25 | An article the search index lists as about three years old says so. Current sources say 10% + $0.50 | Stale | Don’t use |
| 7 | Gumroad has been the merchant of record since January 2025 | Third-party guides say so. I did not open a primary Gumroad source for it | Hedge | Reported to have been merchant of record since January 2025 |
The refund conflict, in detail
Claim 5 is the most instructive. One version of Gumroad’s fee article says it doesn’t refund any fees when processing refunds. Another version of the same article says Gumroad returns its own fee on the refunded portion and keeps only the payment-processing portion, because processors don’t return that. A competitor review and a blog both repeat the first version as if it were current. The live page text didn’t load for me, so I couldn’t tell which version is live.
That is an unresolved conflict, so the ledger says: don’t publish a flat claim. The honest sentence is the one in the table.
How can AI help without making things up?
Use AI for the tedious parts of verification and keep the judgment yourself. It is good at extracting claims and comparing text you supply. It is unreliable at remembering facts or producing links, so open every link it gives you.
If you use Claude, the downloadable claim ledger builder skill packages these steps so you can run them with one request. Read the file before you install it.
What mistakes break verification?
- Treating repetition as confirmation. Ten blogs copying one number are one source.
- Skipping the date. Without a “checked on” date, you can’t tell stale from current.
- Verifying only the easy claims. Check the numbers and rules first, because those are what readers act on.
- Hiding conflict. A hedged sentence with a pointer to the source is more credible than false certainty.
- Trusting AI-provided links. Open each one and confirm it says what the AI says it does.
FAQ
How do I verify AI-generated statistics?
Can I ask AI to fact-check itself?
What counts as a primary source?
What if two official pages disagree?
How long should verification take?
How often should I re-check published facts?
Should I tell readers when I couldn’t verify something?
Where to start
Take your next draft, pull out the ten claims readers are most likely to act on, and put them in the ledger. Fix those first. Even that small step catches the errors that do the most damage.
Gumroad Help Center: Gumroad’s fees · Cartmango: Gumroad pricing · Fungies: Gumroad alternatives (competitor) · Checkout Page: Gumroad fees · WalletMinded (older article) · Fiverr Help Center: Using AI on Fiverr. Fees and policies change; this example reflects what was visible on October 11, 2026. General information, not legal or financial advice.
