Fact-Checking Workflows for Marketing Content

Everyone has a checklist and the errors still get through. The workflow that actually catches them: claim extraction, source of record, the cut rather than hedge rule, and a corrections policy almost nobody has.

Ask any marketing team whether they check facts before publishing and they will say yes. Ask how, and the answer is usually that the editor reads it carefully.

That is not a workflow. It is an intention, and intentions fail predictably under deadline. The errors that get published are almost never the ones somebody considered and got wrong. They are the ones nobody noticed were claims at all.

This is about the mechanics. The checklist of what to look for is a different document, and we set that out in AI content QA.

The first step is extraction, not checking

You cannot check claims you have not identified, and identifying them is a separate task from reading.

Before review, somebody goes through the draft and lists every factual assertion in it. Not the arguments or the opinions. The things that are either true or false: numbers, dates, names, quotations, attributions, descriptions of how something works, statements about what a company or a regulator does.

Doing this mechanically produces a surprise the first few times. A typical two thousand word article contains far more checkable claims than anybody expects, and several of them will be things the writer absorbed rather than sourced.

The sentence that most often slips through is the one that sounds like background. Most businesses now do X. The regulation requires Y. The platform removed Z last year. Nobody reads those as claims, and every one of them is.

Define what counts as a source before you need to

Most arguments in review are about whether something is sufficiently sourced, and they are resolved faster if the standard is written down in advance.

A workable hierarchy.

  • Primary and authoritative. The organisation’s own documentation, a regulator’s own text, a company’s own announcement, the actual study rather than coverage of it. This is the standard.
  • Primary but interested. A vendor’s research about the market the vendor sells into. Usable, and it must be labelled as vendor-published, because it is a primary source for what the vendor claims rather than for the state of the world.
  • Secondary. An article citing something else. Not a source. Follow it to the original, and if the original cannot be found, treat the claim as unsupported.
  • Unsourced. Everyone knows this. It is widely reported. Cut it.

The rule that does the most work is the second one. A great deal of marketing writing repeats a vendor’s figure as though it were an industry fact, and the difference between those two is the difference between honest and not.

Cut, do not hedge

This is the single rule that changes output quality most, and it is the one people resist.

When a claim cannot be verified, it comes out. It does not get softened into reportedly, or studies suggest, or it is widely believed that. Those constructions keep the assertion while removing responsibility for it, which is fabrication with a disclaimer attached.

Writers resist because the hedged sentence felt useful and the paragraph is weaker without it. Almost always the paragraph is fine, and frequently better, because the unsupported claim was doing rhetorical work rather than informational work.

The acceptable alternative is to state the uncertainty as a fact about knowledge. We could not find a reliable figure for this. This is not documented publicly. This is our judgement rather than an industry standard. Those sentences are honest, they are often the most credible lines on a page, and they are quotable in a way a hedge is not.

The writer should not be the checker

Not because writers are careless, but because a writer has already decided each sentence is right. Re-reading your own draft retrieves the reasoning rather than testing it.

Have someone else check, with the extracted claim list rather than the flowing article. Checking a list is a different cognitive task from reading prose, and it catches things reading does not.

Where the team is too small for a second person, impose distance instead. Check the next day, from the list, not from the draft, and with the sources open rather than from memory.

Keep the source attached to the claim

The most common structural failure is that verification happens once and the evidence evaporates.

Somebody checks a figure in March, confirms it, and publishes. In September the figure is questioned and nobody can remember where it came from. The choice is then to recheck from scratch or to trust a decision nobody can reconstruct.

Carry the source link with the claim through the whole pipeline. In the brief, in the draft, in the review, and ideally in the published page. A visible citation is a feature rather than clutter: it lets a reader verify, it makes the page more credible to systems weighing provenance, and it means the next person to touch the page knows what the claim rests on.

Where you cannot cite publicly, keep an internal record attached to the page with the date it was verified.

Treat time-sensitive claims as a separate category

Some facts are true when written and false within a year. Prices, counts, currently and recently, who holds a position, whether a feature exists, what a regulation requires.

These need their own handling, because they fail silently. Nobody discovers a stale claim. A reader simply notices the page is wrong and quietly discounts the rest of it.

Two practices help. Mark decaying claims when you write them, with a note of what would make them wrong. And give any page containing them a review date rather than leaving it to be noticed. The broader decision about what to do with ageing pages is in content refresh.

Prefer durable phrasing where you can. A sentence describing why something is structured the way it is will outlive one quoting this quarter’s figure.

Have a corrections policy, because you will need one

This is the part almost no marketing team has, and it is becoming the part that matters.

You will publish something wrong. The question is what happens next, and the default answer, quietly editing the page and saying nothing, is the wrong one.

Write a short policy covering four things. Who can authorise a correction. How a corrected page is marked, which should be a dated note at the top or bottom saying what was changed and when. Whether anything is ever removed entirely rather than corrected. And who is told, if the error was consequential or was distributed in a newsletter.

Two reasons this is worth the discomfort. A visible correction is a credibility signal rather than an admission of weakness, because it demonstrates that somebody is accountable for the content. And an error that was quoted elsewhere propagates, so correcting it visibly is the only way to interrupt that.

Provenance is weighed more heavily now than it used to be. A site that corrects visibly and dates its reviews is making a claim about its own reliability that an unmarked edit cannot make. The selection logic behind that is in how AI search engines choose what to cite.

Where AI assistance changes the risk

A drafting tool produces plausible specifics with no signal distinguishing the ones it knows from the ones it constructed. A fabricated statistic looks exactly like a real one, and a fabricated attribution looks exactly like a real citation.

Which means the extraction step matters more, not less. Every number, every name, every study, every quotation gets traced to a primary source, regardless of how the draft was produced.

Be especially careful with citations that look right. A plausible author, a plausible journal, a plausible year. Check that the source exists before checking that it says what the draft claims.

The same discipline applies to case studies and client examples, where invention is both easy and damaging, as we set out in writing client case studies that get cited.

What is a waste of effort

Checking the easy claims thoroughly and the background claims not at all. The errors are in the background.

Running the draft through a detection tool and treating a clean result as verification. Those tools do not check facts.

Adding a disclaimer to the site instead of fixing the process. Nobody reads it and it does not make a wrong figure right.

Checking only what a lawyer would worry about. Most reputational damage comes from ordinary errors that make a page look unreliable, not from actionable ones.

Building a thirty-step review process that gets abandoned in a busy month. Three steps that always happen beat ten that sometimes do.

A sensible sequence

  1. Extract every factual claim from the draft into a list, as a separate task from reading.
  2. Write down your source hierarchy once, including the vendor-published distinction.
  3. Trace each claim to a primary source, and cut rather than hedge anything that fails.
  4. Have somebody other than the writer do the checking, from the list.
  5. Keep the source attached to the claim, visibly where possible.
  6. Mark time-sensitive claims and give the page a review date.
  7. Write a one-page corrections policy before you need it.
  8. Review a sample of already-published pages quarterly, because drift happens in what got through.

If you want help building this into an existing content operation, get in touch.