How-To Guide

How to Fact-Check AI-Written Pages

A claim-by-claim method for checking AI-written content: list the claims, match each to a source, grade the source, cut what fails, and record it. With the three-grade system we use on our own pages.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: November 13, 2026

Why This Step Exists

A language model writes a false figure in the same confident tone as a true one. It does not look facts up unless it is given a tool to do so, and even then it can misread what it finds. Google's guidance on generative AI content says such content should be fact-checked and reviewed before publishing. If AI wrote the draft, the checking has to be done by a process that does not trust the draft.

Step 1: List Every Claim

Go through the page and pull out each statement that could be wrong: numbers, dates, names, prices, quotations, rankings, "first", "largest", "only", and any statement about what a company or person did. Write them as a list. A thousand-word page usually holds twenty to forty.

Step 2: Match Each Claim to a Source

For each claim, find where it comes from and read that source yourself. Not a summary of it. Prefer the origin: the company's own announcement, the filing, the paper, the documentation, the dataset.

  • Be careful with summaries. Search snippets and AI summaries of a page are where details get changed.
  • Check the date. A correct figure from two years ago is a wrong figure today.
  • Check the definition. "Users" might mean monthly, weekly or registered. "Revenue" might be a quarter or a year.

Step 3: Grade the Source

We use three grades and record one for every claim.

GradeMeaningHow it appears in the copy
AA primary source, read directlyStated plainly
BAn established outlet, or two independent sources that agreeStated, with the source named
CA single secondary source or a summaryAttributed and hedged, or left out

Step 4: Cut What Fails

A claim with no source comes out. Do not keep it and add "reportedly". If the page does not work without the claim, the page is not ready.

When two sources disagree, say so and give both, or leave the number out. Do not pick the more impressive one.

Step 5: Check the Parts Outside the Body

Titles, descriptions, tables, FAQ answers and structured data repeat the page's claims, often in rounded or shortened form. Google's guidance specifically mentions titles, meta descriptions, structured data and image alt text, because they can appear in search results. A figure corrected in the body and left wrong in the FAQ is still wrong.

Step 6: Check Your Own Arithmetic

If the page computes anything, such as a percentage, a total or a growth rate, redo it and say on the page that it is your calculation. A total that does not match its rows is the easiest error for a reader to spot.

Step 7: Record It

Keep a short record per batch: each claim, its source and its grade, plus what was left out and why. It makes the next check faster, and it tells you which pages depend on a source that has since changed.

Step 8: Have Someone Else Read It

The final check is a person who did not write the page. They read it as a customer would and ask one question of anything surprising: how do we know this?

Common Ways AI Drafts Go Wrong

FailureExampleHow to catch it
Invented precisionA neat figure that appears in no sourceSearch for the exact number
A real fact, wrong yearLast year's price presented as currentCheck the source date
Merged entitiesTwo similar products described as oneRead the source's own naming
A product that never shippedAn announced or rumoured release written as availableFind the availability page
Vendor claims as factA company's own benchmark stated without attributionSay who measured it

Re-Checking After Publishing

Facts expire. Prices, product names, statistics and "latest" claims need a date and a re-check schedule. We list the items to revisit when a batch ships, and keep the last-checked date on the page.

What This Costs

Checking takes longer than drafting. That is the point of using AI agents for it: one agent drafts, another checks each claim against its source and removes what it cannot support, and a person reviews what is left. See AI marketing agent.

If You Want This Done for You

Presenc AI runs a team of AI agents that does this work, and a person reviews every page before it ships. Request access.

Sources

Google's guidance on generative AI content, read November 13, 2026. The method and grading system are our own practice.

Frequently Asked Questions

List every claim that could be wrong, match each one to a source you read yourself, grade the source, remove anything without support, check titles, tables and FAQ answers as well as the body, and have a second person read it.
A language model generates likely text. It does not look facts up unless given a tool, and it writes an invented figure in the same confident tone as a real one.
Remove it. Do not keep it with a hedge such as "reportedly". If two sources disagree, give both with attribution or leave the figure out.
Google's guidance on generative AI content says to fact-check and review AI-generated content before publishing, including titles, descriptions and structured data.

Hire agents, not agencies.

Presenc is a team of AI agents that builds the pages your buyers search for, and a person checks every one. We take on a few teams at a time.

Request access