Fact-Checking AI Marketing Content: A Source-Verification Workflow

• Extract every checkable claim first, give it an ID, and only then edit the prose.
• A source counts only after a person opens it and finds the passage. A citation supplied by a model is a lead.
• Re-derive every number from its inputs and write the arithmetic down.
• Titles, meta descriptions, structured data and alt text are claims too.
• Keep a provenance log with the published version. The template is below.

What does it mean to fact-check AI content?

Articulate describes AI fact-checking as verifying AI output so it holds no factual, grammatical or other errors, which protects credibility and avoids preventable legal and ethical risk. Its explanation of why the errors happen is that AI predicts the most probable response to a prompt from patterns in large datasets and is not trained for accuracy.

The errors that result are commonly called AI hallucination: content that sounds true and is not. This page treats the problem as an editing task with a defined output, a verdict on every claim and a log that proves the work was done.

Why does AI content need a fact-check before it publishes?

It needs one because generative models do not look facts up, and Google now says so in its guidance to sites. Search Engine Journal reports that Google revised its generative AI guidance on 1 October 2026 to add three sentences: models predict a likely sequence of words from training data, their output may contain inaccuracies, and it is critical to manually fact-check and review all AI-generated content before publishing.

The same report says the December 10, 2025 version asked for accuracy but had no manual fact-check instruction. It also says the review now extends to metadata, listing title elements, meta descriptions, structured data and the alternative text of images. The year of the October change is inferred from that comparison, so confirm it on Google's guidance page and its documentation updates page.

The advice is not new. According to Search Engine Journal, Gary Illyes said in an April 2024 episode of Search Off the Record that generative AI output always has to be fact-checked, and in an August 2025 interview that AI content should be human curated.

Articulate lists what is at stake when the check is skipped: reputational damage, an ethical duty to give audiences accurate information, search engine penalties for content judged low quality or inaccurate, and legal or regulatory exposure. Its own example is a job aid drafted with AI after a safety notice, which read well and still held several inaccuracies and a diagram of machines the team did not own.

What are the five steps of a source-verification workflow?

The five steps are extract, match, re-derive, decide and log, run in that order on every piece. Each step has one question, one output and a stop rule, so a reviewer can tell when the step is done.

extract claims → match to opened sources → re-derive numbers → decide → log

Steps 1 to 3 are where errors are found. Steps 4 and 5 are what make the work repeatable, and they are the steps most guides leave out.

Step 1: How do you extract the claims from a draft?

Read the draft once with no editing, and write down every statement that could be true or false as its own numbered line. A sentence with three facts becomes three claims, because each may fail on its own.

Then label each claim with a type. The type sets how hard the match has to be, and the table gives the standard of proof used in this workflow.

Opinions sit at the bottom of the table because they cannot be verified. Originality.ai makes the same point about its own tool, saying a statement such as which ice cream flavour is best cannot be checked.

Do you check every claim or only numbers, dates and quotes? Check every verifiable claim, and give numbers, dates, quotes, attributions and product claims the full match and re-derivation. Background statements get one quick match, and opinions are labelled or cut.

Step 2: How do you match each claim to a source?

Open a source, find the passage that supports the claim and record where it is. A source that you have not opened does not count, including one the model named.

Articulate suggests prompting the AI tool to include sources, which makes validation faster, and searching the source with Ctrl+F for the specific statistic or quote. If the draft has no source, it suggests entering the statistic in a search engine, then checking credible sources such as government or non-partisan research bodies, a deeper academic search through Google Scholar, or fact-checking sites such as Snopes, FactCheck.org and PolitiFact.

Articulate also gives the stop rule: if a claim still cannot be validated, reconsider using it. Kanso's version is stricter about sources. Prefer the primary document, such as a regulator, a platform's own documentation or the study itself, and when the only support is a news report, either trace it to the primary or attribute the claim to that report by name.

The filled log below shows why. Every row about Google rests on one trade-press report, so each is marked secondary and carries a note to confirm it on Google's own page.

Step 3: How do you re-derive the numbers?

Rebuild each figure from the inputs the source gives, write the operation down and compare the result with what the draft says. A figure that cannot be rebuilt either gets its inputs added to the draft or is removed.

Here is the exercise on figures from Originality.ai's page. It reports an accuracy study of its fact checker with a result of 86.69%, against 86.67% for GPT-5, and describes that as a close tie with a slight lead.

• The gap: 86.69 minus 86.67 equals 0.02 percentage points. A draft that says the tool "beat" GPT-5 by a wide margin fails this check.
• The error share: 100 minus 86.69 equals 13.31%. Dividing 100 by 13.31 gives 7.5, so the honest phrasing is about 1 in 7.5, not 1 in 8.
• The caveat: the extract does not define accuracy, so the 1 in 7.5 reading assumes it means the share of verdicts that were correct. Say so, or leave the derived figure out.

Two habits catch most arithmetic slips: separate a percent from a percentage point, and note the base the figure is a share of. These are Kanso's habits, not a published standard.

Step 4: How do you check dates, consistency and metadata?

Check that every fact is current, that the draft does not contradict itself, and that the same review covers the metadata. Articulate lists the first two as separate checks: AI can reference outdated information on fast-moving topics, and a draft can state a claim in one section that a later section contradicts.

For timeliness, Articulate suggests searching a newer date range, so that a claim from a 2024 report prompts a search for a 2025 or 2026 report. In the log, write the source's date next to the claim and put "as of" in the sentence whenever the fact can change.

When does a claim need an expert?

A claim needs an expert when it sits in a specialist field and the tool is a general-purpose one. Articulate says general AI tools are less effective in fields such as medicine, law and engineering, and advises consulting a professional when content on technical or niche subjects comes from an off-the-shelf tool.

Does the check cover titles, meta descriptions and alt text?

Yes, because Google's guidance, as Search Engine Journal reports it, applies the manual review to metadata that can appear in search results. That means title elements, meta descriptions, structured data and image alt text each get rows in the log, with their own location.

Where do AI fact-checking tools fit in the workflow?

They fit in step 2, as a faster way to find candidate sources, and nowhere else. Originality.ai's own page describes a checker that works through content sentence by sentence, fetches relevant URLs and uses their content to verify each fact, returning a potentially true or false flag with an explanation and source links.

The vendor also states limits. The fact has to be verifiable, it is steering clear of political topics for now, and its own study led it to say the tool should be an aid and not the final source of truth.

A tool's verdict is a row for the log, not a replacement for it. Record it as the source the tool suggested, then open that source yourself.

What goes in a provenance log?

A provenance log holds twelve fields for every claim: where it sits, what it says, where the support is, what the numbers rebuild to, and who decided what. None of the three top-ranking pages read for this article includes one, which is why Kanso publishes a template.

Use four verdicts, so that no row can stay vague.

• Supported: a source an editor opened says exactly this. Keep the claim.
• Supported with change: the source says something close. Rewrite, attribute or date the claim to match it.
• Not supported: the source says something different. Rewrite from the source or cut.
• Could not verify: no source found in one pass. Cut it or mark it clearly as an opinion.

Start the log before drafting if you can. The content brief template includes source fields, so a brief can name the primary sources the draft is allowed to rely on.

What does a filled provenance log look like?

It looks like the table below, which logs eight claims from a hypothetical short draft about the topics on this page. The draft is invented for the demonstration; the sources are the three pages read for this article, checked on 28 October 2026, and the derived numbers are the ones worked in step 3.

Of the eight rows, four are supported as written, three are supported only after a change and one is not supported. The four supported rows all rest on a secondary report, which is why the log carries the primary or secondary field: it shows how many claims still need the primary document opened.

Rows C05 to C07 show the value of re-derivation. The figure matched the source each time, and the draft's wording around it was the error.

How do you run the workflow on a small team?

Run it as a gate that sits between the draft and the editor's style pass, with a cut-early rule so that verification never grows into research. This is Kanso's suggested routine; the sources above do not prescribe one.

1. Extract the claims before any style edit, because rewriting sentences changes what they claim.
2. Sort the claims by type. Numbers, dates, quotes, attributions and product claims get the full match and re-derivation.
3. Give each claim one pass to find support. If it fails, soften or cut it instead of hunting.
4. Add a second reader only for pieces with wide reach or high stakes. The risk-by-task matrix for approval gates helps decide which.
5. Save the log with the published version and re-check dated facts when you refresh the page.

The review time is a real cost, and it belongs in the decision to use AI for a task at all. The break-even formula for AI-assisted tasks counts review time and redo time against the time saved.

Kanso's Writer Agent turns Search Console opportunities into long-form drafts in your brand voice, for you to review and publish yourself. Its description does not include a claim-verification step, so treat each draft as unchecked until its log is complete.

What is documented and what is our analysis?

The documented points come from the three sources named in each item, and the analysis is Kanso's own reasoning built on them.

Documented

• Per Search Engine Journal, Google revised its generative AI guidance on 1 October 2026 to say outputs may contain inaccuracies and that manual fact-checking and review are critical before publishing.
• Per Search Engine Journal, the December 10, 2025 version had no manual fact-check instruction, and the review now covers metadata.
• Per Search Engine Journal, Gary Illyes said in April 2024 and August 2025 that AI output needs fact-checking and human curation.
• Per Originality.ai, its own study found 86.69% accuracy against 86.67% for GPT-5, and the vendor says the tool is an aid.
• Per Articulate, the checks are sources, cross-checking, consistency, timeliness and expert review.

Analysis

• The five-step order, the stop rules, the eight claim types and the standard of proof for each.
• The twelve-field provenance log and the four verdicts.
• The reading of the Originality.ai figures as a 0.02-point gap and about 1 in 7.5, which depends on how that study defines accuracy.
• The advice to limit second readers to high-stakes pieces. No test or survey of editorial teams was run for this page.

Fact-checking AI content: frequently asked questions

Sources

• Google Tells Sites To Fact-Check AI Content Before Publishing, Search Engine Journal
• Google Search's guidance on using generative AI content on your website, Google Search Central
• Fact & AI Hallucination Checker, Originality.ai
• How to Fact-Check AI Content Like a Pro, Articulate