Why unsourced claims are worth an audit
Most content libraries contain claims that were true, or believed true, when they were written. Some figures have since moved. Some were copied from an article that copied them from somewhere else. A few were never true.
Three things make that worth fixing now.
- Google weights trust above everything else in E-E-A-T. Its guidance on helpful content says of experience, expertise, authoritativeness and trustworthiness: "Of these aspects, trust is most important."
- Answer engines favour sourced content. In the Princeton GEO study, citing sources was among the strongest interventions, and it raised the visibility of a site ranked fifth in search results by 115%.
- A wrong figure under your name is a liability. It gets quoted back to you by prospects, by competitors and, increasingly, by AI assistants that found it on your site.
We ran this audit on our own two flagship guides in September 2026. We found we had attributed two figures from the Princeton study to the wrong methods: the paper reports 41% and 28% as the gains of its best methods on two different measures, not as the separate effects of statistics and quotations. We had also carried forward click-through figures from a secondary source that did not match the original report. Both guides now say what the sources say.
What counts as a claim
Audit these, and only these:
- Statistics: any number describing the world. "AI Overviews appear on 15.69% of keywords."
- Named research: "a Princeton study found."
- Superlatives and firsts: "the most accurate," "the first platform to."
- Product capabilities: what a product, yours or a competitor's, does or does not do.
- Prices and limits: yours and anyone else's.
- Dates and events: launches, announcements, policy changes.
- Quotations: anything inside quotation marks.
Opinions, recommendations and reasoning are not claims in this sense. "We think most teams should start with their top twenty pages" needs no citation. "Most teams start with their top twenty pages" does.
Step 1: pick the pages
Do not start with the whole library. Start with the pages where a wrong claim costs the most:
- the twenty pages with the most search traffic,
- the pages AI engines already cite, if you track that,
- pricing, comparison and product pages, and
- anything quoted by sales.
Step 2: extract every checkable claim
Read each page once, for claims only. Paste each into a sheet with the page, the exact wording and the link it currently carries, if any. Reading for claims is different from reading for sense, and trying to do both at once is how claims get missed.
Step 3: classify each one
| Class | Meaning | Default action |
|---|---|---|
| Sourced | Links a source that says exactly this | Check the source still says it |
| Sourceable | True, but no link | Find the primary source and link it |
| Internal | About your own product or data | Link the page that states it |
| Unsupported | No source found | Cut it, or rewrite without the claim |
Step 4: find the primary source
This is where most of the time goes, and where most of the errors hide.
- Go to the original. If an article cites a study, open the study. Figures drift as they are summarised: rounded, attributed to the wrong method, or stripped of the condition that made them true.
- Check the date and the population. A figure from one country, one industry or one month is not a general fact.
- Check the number says what you say. "Clickthrough fell 58%" and "the presence of an AI Overview correlates with a 58% lower clickthrough rate" are different sentences, and only one of them is what the study said.
- Prefer the operator's own documentation for product facts: a vendor's pricing page for its prices, a crawler operator's documentation for what its bot does.
- Note when you checked. Prices and plan limits change. A date beside the claim tells the reader how fresh it is.
Step 5: decide
For each claim, one of four outcomes:
- Link it. The source says exactly what the sentence says.
- Rewrite and link. The source supports a narrower claim. Change the sentence to match it.
- Attribute it. The claim is contested or comes from one party. Say whose claim it is: "Google called the study inaccurate."
- Cut it. No source supports it. If the paragraph depended on it, rewrite the paragraph.
Resist the fifth option, keeping the claim and hoping. A reader who checks and finds nothing assumes the rest of the page is the same.
Step 6: record, then re-date honestly
Change the page's updated date only when the content changed, and keep a short note of what changed: which claims were corrected, cut or newly sourced. The date signals freshness to readers and to crawlers; changing it without changing the content teaches both to ignore it.
The audit sheet
Copy this structure:
| Page | Claim as written | Class | Current link | Primary source | Checked on | Action | Rewritten claim |
|---|
One row per claim. The sheet becomes the list of pages to re-check next quarter, sorted by how fast their facts change.
Making it the default
An audit fixes the library you have. The lasting fix is writing claims with their sources attached from the start, which is the idea behind citation-first content. Veritas drafts that way, with a citation on every knowledge-based claim, and its content scoring checks drafts as they are written. For why unsourced claims appear in AI drafts in the first place, see What the Research Actually Says About AI Hallucinations in Marketing Content.
Want to see which AI answers mention your brand?Start a 7-day trialSee pricing