A typo costs you a comment. A wrong number costs you a customer, sometimes a refund, and sometimes a letter from someone who bought on the strength of that number and wants to explain what it cost them. That is the shortest reason to fact check ai copy before it publishes.
There is a second reason to fact check ai output properly, and it is less dramatic. A wrong number keeps costing after it is fixed, because people who read it, quoted it, and now see it corrected have to be told. The correction is never as widely read as the claim was.
This page is five wrong numbers, each one a shape we have either published ourselves or watched a client publish, with the reason it survived review and the check that would have caught it. No invented case studies, no client names, no figures I cannot stand behind.
If you want the routine that produces these checks in one pass, it is in verify AI content before publishing. If you want the general method first, start with the seven checks and come back here for the cost argument. Both pages teach how to fact check ai drafts by hand before any tool is involved.
[Cost rising from a typo to a refund, with five stops on the ladder.]

Why one wrong number costs more than ten typos
Three reasons, and each one is structural rather than dramatic.
A wrong number gets repeated. Readers copy figures into their own notes, slides and reply emails. A typo dies on your page. A number travels, and every copy makes the correction harder.
A wrong number gets acted on. A price decides a purchase. A statistic decides a plan. A date decides whether someone believes the page is current. Action is where money moves, and that is why you fact check ai output before publishing rather than after.
A wrong number is defensible in an argument. Typos are obviously errors. A figure that sat on a page for three weeks is a representation, and the fact that it was accidental does not change what the buyer saw.
This is why we built an engine for claims rather than for grammar. Grammar checkers are everywhere, and the expensive failures are not grammatical. They are sentences that read perfectly and state something false. At any scale, this is the category that matters when you fact check ai output.
Case 1: the price from a neighbouring page
Ours. A price on a product page, live for eleven days, copied from a different product in the same catalogue. Same amount, same currency, same one-time wording. The only thing wrong was which product it described.
Why it survived review: every reviewer read for sense, and the sentence made sense. Nobody had the other product’s page open at the same time, so nobody had anything to compare against. This is the failure that made us start to fact check ai written copy at all.
The check that catches it: does the source name the same subject as the sentence? Not a similar subject, the same one. It is the fourth of the seven checks, it takes ten seconds, and skipping it cost us eleven days.
If you fact check ai output on product pages, design the routine around this case rather than around the obvious ones, because it is the failure that looks correct in four different review passes.
[Two sentences with the same price and different product names, side by side.]

Case 2: the percentage that changed quarters
A claim that a tool reduces review time by a percentage, taken from a survey run in one quarter, published in the next, quoted in the third. By publication the figure described a sample that no longer existed.
Why it survived review: percentages look like measurements rather than claims. A reviewer checks the arithmetic and not the dataset, and the arithmetic is usually right.
The check that catches it: the period word, which is the field people skip when they fact check ai generated statistics. A percentage carries a when, and the when lives in the source, not in the sentence. If the source says the survey covered the quarter ending June, the sentence has to be able to say that, or the number is a claim about a moment nobody named.
There is a second trap inside this one. Percentages of small samples swing wildly, so the same question asked twice can produce two very different figures and both are honest. Publishing only the flattering one is not a lie, it is a choice, and the source card is where the choice becomes visible to whoever reads it next. That is how fact check ai work becomes auditable rather than remembered.
That is the argument for a source card rather than a memory when you fact check ai generated statistics. The card holds the date, and the date is what makes the choice inspectable.
Case 3: the sample size that was never a sample
A page states that a certain proportion of writers get fact checking wrong. The number is real, it came from a survey, and the survey asked twenty three people who were in a room together at one event.
Why it survived review: the sentence never mentions the sample, and a reader who does not ask about the sample cannot object to it.
The check that catches it: any statistic carries its basis or it is a claim only the original author could defend. The honest version of most of these sentences is shorter and less exciting. In a survey of twenty three attendees at one event, most said something. That sentence is checkable and useful, and the exciting version is neither. This is the case that convinces most teams to fact check ai drafts on a schedule.
If you fact check ai generated statistics at all, this case justifies the routine. Models produce proportions without populations, because a proportion is a pattern and a population is a specific fact, and patterns travel better in generated text.
[One statistic with its basis attached and removed, showing what changes for the reader.]

Case 4: the date that moved a year
A page describing a regulation, a version or a deadline using a date that was correct last year, one year out, in a sentence that mentions no year at all.
Why it survived review: a date without a year cannot be wrong on its face. Reviewers read the month name and think about the month, not about which year it belongs to.
The check that catches it: write the year into the sentence. It costs four characters and converts an unverifiable statement into a checkable one. Then check it, because a year in the sentence is only useful when the year is right. Dates are the cheapest thing to fact check ai round and the most commonly missed.
This failure is common in anything evergreen. A page written to last gets edited for tone and never for dates, and the numbers age in place while the design is refreshed around them. When you fact check ai drafts in a content refresh, the dates are the first thing to look at and usually the last thing anyone edits.
Case 5: the up to figure with no ceiling
“Up to fifty percent faster.” “From ₹999.” “As many as ten thousand users.”
Why it survived review: hedged claims feel honest. Up to is cautious language, and cautious language reads as accurate even when the ceiling has no source at all.
The check that catches it: a range needs two endpoints and each endpoint needs a source. Up to fifty percent means somebody measured fifty percent, somewhere, on something. If nobody did, the sentence is a claim with a missing measurement wearing a hedge.
There is a legitimate version. A published price list that genuinely starts at a figure, stated as a starting price, on the date of publication, with the list linked. That sentence survives a challenge because the ceiling is visible and the date is attached.
Hedging is not evidence. Endpoints are. That distinction is most of what you learn after the first month of having to fact check ai copy someone else will publish.
[The hedged claim, the source it needs and the honest rewrite.]

The pattern behind all five
Five cases, one shape: the number was true and the sentence was wrong.
None of these failures could be caught by reading more carefully. Every one required comparing the sentence against something else: another page, the survey, the price list, the calendar. Careful reading is a single pass over one document. Fact checking is a comparison between two.
That is the sentence worth putting on the wall. A document cannot check itself, which is why every attempt to fact check ai output by re-reading the same file produces the same false confidence.
The pattern also explains why the people in these five cases were not careless in the ordinary sense. Anyone who has to fact check ai writing for a living meets this once a week. They were careful about the wrong thing, in the same way that reading a contract carefully tells you the terms but not whether they are legal.
The cheapest check that finds them all
One habit in one order finds four of the five in under ten minutes.
- Take the first paragraph that contains a number. Copy the sentence.
- Open the source. Copy the sentence from the source holding the same number.
- Place them beside each other. Underline the subject in both, the currency in both, the period in both.
- If any of the three differs, or is absent, the claim needs work.
That is the whole method, and it is the routine most people skip, because it feels too simple to matter and because it forces you to open the source, which is the step where you discover the source does not say what you thought it said.
Then the mechanical version, for a draft longer than one page:
node bin/saakshi.js scan draft.md
node bin/saakshi.js prove draft.md --ledger ledger.json
node bin/saakshi.js pack draft.mdExit 0 publishes. Exit 3 asks a human to decide and write the decision down. Exit 2 blocks, which in these five cases means the number does not go live. The engine is described at SOVEL SAAKSHI, including the nine things it cannot do.
The policy reason to fact check ai content rather than only the commercial one is dated and short. Google’s guidance on generative AI content asks for a manual review before publishing and states that the review covers metadata as well as body text. The spam policies are where the consequences sit.
[The eleven day copy, corrected, with the two sentences that should have differed.]

How to fact check ai drafts in a weekly routine
A routine beats a rule, because a rule is what you intend and a routine is what happens.
Monday, ten minutes. Scan every title and meta description changed in the last week for numbers, prices, percentages, superlatives and year words. Those fields are generated by templates and read by nobody.
Wednesday, twenty minutes. Check prices against the source of truth, which is the pricing page, not the spreadsheet. Prices are where fact check ai discipline pays for itself fastest. If the two disagree, the page is wrong until one of them is fixed.
Friday, thirty minutes. Take the two pages that earned the most traffic this week and run the full check on both. Traffic concentrates risk, which is why fact check ai routines start with the busiest pages. Highest traffic means highest cost when a figure is wrong.
That is an hour a week and it covers the five cases on this page. It is also the smallest version of a real fact check ai process. The reason to fact check ai content on a schedule rather than at publication is that publication is when you are most optimistic, and optimism is what lets a wrong number through.
If you need a standard to hand to someone else, ai fact checker lists the six claim types and the verdict each one produces. If the problem is a document rather than a page, use verify claims in a document.
[The weekly routine as a small calendar, one task per day.]

What good looks like afterwards
A page that has been through this looks slightly different. Figures carry their subject. Prices carry their currency, their period and their edition. Statistics carry their sample. Dates carry their year. Ranges carry both ends.
None of that makes the writing worse. It makes it slower by about four minutes a page and dramatically harder to get wrong, and the second effect is the one that shows up in your inbox.
There is a commercial version of this too, and it is usually why agencies start. When a client asks how you fact check ai drafts, the folder is the answer. When a client asks how a number was checked, the answer is a folder with a source card and a date, not a promise. That folder is the difference between a supplier and a professional, and it is the reason to fact check ai output before the client sees the draft.
The examples below were rebuilt on 21 November 2026 from drafts that had actually been published.
Learning to fact check ai output is a skill that survives every model change, and the habit is worth more than any single tool you use to fact check ai copy.
Numbered references are the other half of the job, and a citation verification tool handles them.
Questions people ask about how to fact check ai content
Is this about AI writing or about all writing?
Both, and the distinction matters less than it sounds. The failure pattern above is not caused by a model. It is caused by copying a number from a source into a sentence without copying the parts that make it true. Models do that faster, which is the only real difference.
How much does a wrong number actually cost?
It depends on what the reader did with it. A price error costs the difference or a refund. A statistic error costs credibility, which is harder to price and slower to recover. The cost that is always present is the correction, because it arrives at the worst moment and is read by fewer people than the original claim.
Can a spell checker or grammar tool catch these?
No. All five sentences above are spelled correctly and grammatically perfect. Grammar tools check the sentence against the language. These failures need the sentence checked against a document, which is the whole point of learning how to fact check ai drafts rather than how to proofread them.
What if the source is a PDF or an image?
Then you read it yourself. Deterministic extraction over Markdown cannot see inside a screenshot or a PDF table, so copy the sentence by hand into the ledger with the page number. That is a limitation of the tool and it is stated here rather than discovered later.
Do I need software to do this?
Not for one page a week. The ten minute habit finds four of five cases. Software earns its price when volume rises, when a client wants proof, or when two people need to agree on the same standard without a meeting.
Which edition covers client work?
Agency, because the licence is about rights rather than features. Personal covers your own publishing, Pro covers commercial use of your own products, Agency covers deliverables you hand to a client. The engine is identical in all three archives.
Is there an online version?
No, and there will not be one in the v0.x line. Offline is a design constraint, because an engine that uploads client drafts has to be trusted with them, and most agencies cannot make that decision about a draft under NDA.
How do I start this week?
Pick one page with a price on it. Run the four step habit on it tonight, by hand. Most people find one thing, and the one thing is usually the subject rather than the digit, which tells you exactly which habit to build first.
For the full editorial method, how to fact check AI generated content is the pillar. For the policy deadline that makes this urgent, read what Google changed about AI content. For the product itself, with prices and limits published before purchase, see SOVEL SAAKSHI.
And if you only keep one line from this page, keep this one. A number that is true in the source and false in your sentence is still a false claim, and that is the one nobody catches by reading carefully. Every reason to fact check ai output comes back to that sentence in the end.