Most pre-publish fact-checks are confirmation passes. The editor reads the draft, nothing looks wrong, and it ships. That process catches typos and obvious errors. It misses the claims that are fluent, plausible, and unsupported, which are exactly the claims that AI-assisted drafting produces in volume.
Google’s guidance on generative content puts it plainly: these models predict likely word sequences, can produce output that sounds right but isn’t, and it’s critical to manually fact-check and review all AI-generated content (Google: using generative AI content). The same discipline pays off on human drafts: confident writers produce confident errors.
What works is an adversarial pass: a review designed to find reasons each claim might be false, not reasons it looks fine. This guide gives you the claim taxonomy, four verdicts, the one rule that does most of the work, and a three-pass ritual you can run in under an hour per article.
The rule: quote or it didn’t happen
Every factual claim in the draft must be backed by a pasted sentence from a source: the exact words, plus the URL and the date accessed. Not “the source says something like this.” Not a link to a 40-page report. The sentence.
This rule does most of the work because it converts a judgment (“does this seem supported?”) into a mechanical check (“is there a matching quote?”). Reviewers get tired. Mechanical checks don’t. And the rule catches the subtlest failure in sourced writing, drift: the source says “many,” the draft says “most”; the source says “can help,” the draft says “will.” Paste the sentence next to the claim and drift becomes visible immediately.
Step 1: Extract and classify every claim
Read the draft once with a single job: list every sentence that asserts something checkable. Classify each by type, because each type fails differently:
| Type | Example | Typical failure |
|---|---|---|
| Numeric | “Cuts review time by 40%” | Invented, or scope-mismatched |
| Causal | “Schema markup boosts rankings” | Correlation presented as cause |
| Comparative | “Faster than manual audits” | No baseline, or an unfair comparison |
| Attributive | “Google recommends X” | Misquoted, outdated, or never said |
| Temporal | “Since the 2025 update…” | Wrong date, wrong sequence |
| Absolute | “Always,” “never,” “guaranteed” | One counterexample breaks it |
Absolute claims deserve special attention: they’re the cheapest to write and the easiest to falsify. Causal claims are the most common unsupported type in SEO writing, because so much of the field runs on observed correlation.
Step 2: Assign one of four verdicts
Each claim gets exactly one verdict:
- OK: a pasted quote matches the claim’s strength and scope.
- SOFTEN: the evidence supports a weaker version. Rewrite the claim to match the quote. “Boosts rankings” becomes “helps Google understand the page,” if that’s what the source says.
- STRIKE: no adequate evidence, and the piece works without the claim. Delete it.
- ASK: only the author or the client can supply the evidence, such as their own data, policy, or process. Send a specific question. The claim stays flagged until it’s answered.
Two rules govern the verdicts. SOFTEN must match the quote, not the writer’s hope: if you can’t rewrite the claim so the quote fully supports it, it’s a STRIKE. And ASK has a deadline. An unanswered ASK at publish time becomes a STRIKE automatically.
Step 3: Run the three passes
The ritual has three passes, in a fixed order:
Pass 1: Extraction (10 minutes). List every claim with its type. Don’t evaluate yet. Mixing extraction and evaluation is how claims get skipped: you get absorbed in checking one and read past the next three.
Pass 2: Adversarial check (20–40 minutes). For each claim, look for the quote. Then ask the adversarial question for its type:
- Numeric: Where was this measured, on whom, and when?
- Causal: What else could explain this?
- Comparative: Compared to what baseline, measured how?
- Attributive: Does the cited page say this today, in these words?
- Temporal: Is this date verifiable in a primary source?
- Absolute: Can I think of one counterexample?
Pass 3: Reconciliation (10 minutes). Apply the verdicts to the draft. Then re-read only the edited sentences for flow. Softened claims often need a small rewrite of the surrounding sentence so they don’t read as hedged.
If you use an AI assistant for passes 1 and 2, give it the role of opposing counsel. “Find every claim in this draft that could be false, and say what evidence would be needed to support each one” produces far more useful output than “check this draft for accuracy,” which invites a friendly confirmation pass. And treat whatever the assistant reports about sources as a lead, not a quote. You still open the page.
Claims that resist checking
Some claims can’t be quoted from a source, and the ritual needs a rule for each kind so they don’t jam the process:
- Opinion and argument. “Most refresh programs fail at the baseline step” is the author’s thesis, not a fact. That’s fine, if it reads as a position. The test is whether a reader could mistake it for a measured finding. If so, add the frame (“in our experience,” “the pattern we see”) or rework it so the reasoning carries the weight.
- First-hand experience. “When we audit sites, the About page is usually the weakest” is supportable only by the author. Treat it as an ASK to the author: can they stand behind it, and would they describe it the same way under questioning?
- Definitions and how-to steps. These are checkable, just not by a quote. Verify them by doing: run the formula, follow the steps, open the setting. A how-to step nobody has executed is an unverified claim.
- Forward-looking statements. “This will matter more next year” can’t be sourced. Keep it as clearly labeled analysis, or tie it to a documented announcement, like a vendor’s published roadmap.
Who runs the pass
The adversarial pass works best when the writer doesn’t run it alone. Writers read their own claims with their own context, and that context fills gaps a reader wouldn’t. The minimum viable setup is a checker who didn’t write the draft, working from the claim list rather than the prose. Small teams can rotate: everyone checks someone else’s piece. A solo writer can get most of the benefit with distance: run Pass 1 the same day, and Passes 2 and 3 the next morning.
Step 4: Keep a verified-facts ledger
Every OK verdict produces a reusable asset: a claim, a quote, a URL, and a date. Keep these in a running ledger. Over a few months, the ledger becomes your team’s citation library. The facts you cite repeatedly (platform guidance, regulatory definitions, your client’s verified numbers) are already checked, quoted, and dated.
The ledger needs one maintenance rule: re-verify anything older than its decay window. Platform documentation might get a six-month window, while a definition from a standard might get two years. When a ledger entry expires, the next article that uses it re-opens the source.
Worked example: a schema article
Illustrative example. A draft article on structured data for local businesses contains these claims:
| Claim (as drafted) | Type | Evidence found | Verdict | Final text |
|---|---|---|---|---|
| “Adding schema boosts your rankings” | Causal | Google says structured data helps it understand page content | SOFTEN | “Structured data helps Google understand what’s on your page” (Google) |
| “Google requires Organization markup for AI Overviews” | Attributive | Google says no special schema is needed for AI features | STRIKE, and replace with the accurate statement | “Google says there’s no special schema required to appear in AI features” (Google) |
| “Most local businesses have no schema” | Numeric | No primary source within the time box | STRIKE | — |
| “Our clients’ pages are indexed within a week of launch” | Temporal | Only the author has the data | ASK | Pending author response |
| “Organization markup never hurts” | Absolute | Markup that doesn’t match visible content conflicts with Google’s guidance | SOFTEN | “Organization markup that matches your visible content is low-risk and widely recommended” |
Five claims, one OK-after-rewrite, two softened, one struck, one pending. The second row is the instructive one. The adversarial pass didn’t just remove an error. It replaced the error with a more useful true statement, which is often what readers actually need.
Make it a ritual, not a heroic effort
The ritual works because it’s boring: the same three passes, the same six types, the same four verdicts, every time. Put it on the publishing checklist as a gate, with one owner, a ledger, and a rule that nothing publishes with an open ASK.
The payoff compounds. Articles that survive an adversarial pass make fewer, stronger claims, and fewer, stronger claims are what both readers and answer engines can safely repeat.
Speed up every pass
Organa Obscura’s SEO & Content Prompt Pack includes a QA & Fact-Check category with fill-in prompts built for exactly this kind of work, alongside nine other categories that cover the rest of the workflow, from keyword and topic research through drafting, on-page SEO, and AEO-ready snippets. All 120 prompts use clear [VARIABLES] you replace with your own context, and the 19-page PDF opens with a primer on prompting that produces usable output rather than generic filler. A Notion-ready CSV and a Markdown version let you drop the whole library into the tool your team already uses, and it works with any AI assistant.
If you want to pair content QA with a structured audit of how answer engines see your site, the SEO Pro Kit bundles the Prompt Pack with the AI-Visibility (AEO) Audit.
Leave a Reply