AI Fact Checker: 6 Real Claims It Catches Before Publishing

An ai fact checker reads a draft, separates sentences that carry a checkable claim from sentences that carry an opinion, and then checks each claim against a source you supply. That is the whole of it for an ai fact checker, and everything else in this category is presentation. E

An ai fact checker reads a draft, separates sentences that carry a checkable claim from sentences that carry an opinion, and then checks each claim against a source you supply. That is the whole of it for an ai fact checker, and everything else in this category is presentation. Everything else in this category, the dashboards, the scores, the confidence percentages, is presentation wrapped around that one comparison.

The comparison is where products differ. Some ai fact checker tools quietly search the web and compare against whatever they find, which produces a confident answer nobody can defend later. Others compare against a source list the operator writes, which produces a slower answer that survives a client email six months after the draft published.

This page is the six claim types that an ai fact checker catches an ai fact checker actually catches, each with the decision it produces, followed by the three it cannot catch. A page that lists only the strengths was written by marketing, and you can usually tell by the second paragraph.

[Six claim types in a grid, each with the verdict it produces.]

Six claim types drawn with the verdict each one produces
Figure 1: Six shapes cover most of what a draft gets wrong.

What an ai fact checker actually does

Four operations run in order, and none of them is clever in the way marketing likes.

Extract. An ai fact checker starts here. Find every fragment that holds a number, a date, a price, a percentage, a quote, a named person or a ranking word. Extraction is deterministic, which means the same draft produces the same list every time, and two runs can be compared.

Bind. Attach each fragment to an entry in a ledger you wrote: the URL, the exact quote, the date you looked, the subject and the tier. The ledger is the input that makes everything else possible, and building one for a single page takes about twenty minutes.

Judge. Compare claim and quote on four dimensions. The amount, the currency, the period and the named subject. All four have to line up before anything is called supported.

Prove. An ai fact checker ends here. Write a verdict and a hash so the decision can be reopened. A verdict you cannot reopen is a mood with a timestamp.

An ai fact checker that skips binding is a search box with an opinion. That is the most common shortcut in the category and the reason so many tools feel impressive for a week and vanish from a routine by the third.

Claim type 1: a price with no subject

The most common claim in commercial writing is also the most fragile. A number, a currency, and nothing saying what either of them buys.

A draft says six thousand nine hundred and ninety nine, one time. A source says a different product costs the same. Both statements are true about the world. The draft is still wrong, because the claim names a subject the source never mentions.

An ai fact checker catches this by requiring a subject in the same sentence or the one immediately before. When the subject is missing, the claim cannot be bound. When it cannot be bound, it comes back unverified, and unverified blocks publication.

The repair is usually a clause rather than a rewrite. Name what the price buys and the claim becomes checkable, and it often becomes correct in the same edit, because naming the subject is what forces you to look at the source again.

[One price claim before and after the subject is added.]

A price claim with and without a named subject, showing which one can be bound
Figure 2: The same amount twice. Only the second sentence can be checked.

Claim type 2: a figure with no currency

Several countries use the same symbol, several style guides drop the symbol, and several spreadsheets keep the currency in a column header that nobody copies.

An ai fact checker treats a bare number as unverifiable rather than guessing. A tool that guesses the currency will be right most of the time, and that is precisely the problem. The remaining times it approves a claim that is wrong by a factor of eighty, with a green tick attached.

Two habits follow. Write the currency in the sentence rather than in a footnote. When the source is a table, copy the header row with the number, because the header is where the currency lives and the header is the first thing a copy and paste drops.

Half the currency errors we have reviewed came from copying a cell without its column title. A tool cannot repair that, which is why the habit matters more than the software.

Claim type 3: a period that changed shape

Period claims fail when the period moves without anyone noticing. Monthly becomes annual, per user becomes per seat, one quarter becomes one year, gross becomes net.

An ai fact checker compares period words as part of binding. When the claim says monthly and the quote says per year, the claim is unsupported even when the amount matches exactly. The engine does not scale it, because scaling by twelve is an inference, and inferences belong to a person who writes down their reasoning.

This is where a tool shows whether it respects the operator. A tool that quietly multiplies is producing a figure that appears in no source. That figure may well be correct and it is still not evidence, and a page built on it has no defence when someone asks where the number came from.

The useful output for a period mismatch is review. You either find a source with the matching period or you rewrite the claim, and either way a human decided, and the decision has a date.

[The same figure as a monthly and an annual claim, marked as different claims.]

Monthly and annual versions of one price, marked as different claims
Figure 3: Twelve times the number is not the same claim.

Claim type 4: a quote about a different product

Quotes carry more weight than the same words without quotation marks, which is why a misattached quote travels further than a misattached sentence.

Four failures appear repeatedly in review. The quote is from the right company and the wrong product. The quote is about the same product and a different edition. The quote is a paraphrase wearing quotation marks. The quote was true before a rebrand, so the name no longer exists on the page.

An ai fact checker catches the first three mechanically, by comparing subject words in the quote with subject words in the claim. It catches the fourth only when the ledger records the old name as an alias, which is why a good ledger is maintained rather than typed once and abandoned.

The paraphrase failure deserves its own sentence, because it is the one that feels harmless. Quotation marks around a summary are fabrication with extra steps, and a reader who trusts them is being told something false about what a source said.

Claim type 5: a tier that was never named

Tier claims are where careful writers still lose money. A page compares three editions, quotes the cheapest, then says something about the middle one.

The rule that closes it is strict and worth memorising. The source must display the tier the claim names. A price list showing all three editions can support a claim about any of them, because the tier appears in it. A page showing one edition supports that one only.

An ai fact checker enforces the direction that matters. A general price list can support a specific tier claim when the tier is visible. A single tier page can never support a claim about a different tier, no matter how reasonable the arithmetic looks.

Getting this wrong costs a tick you did not earn, and an earned tick is worth more than a generous one.

Claim type 6: a ranking with no source

Best, fastest, cheapest, most accurate, number one. Ranking words are claims, and they are the claims most likely to live in a title tag where no human reviews them before they reach a search result.

An ai fact checker flags the ranking word and asks for two things: a definition and a source. A definition is not optional. Cheapest by what measure, across what set, on what date. Without those three the claim cannot be checked by anyone, including the person who wrote it.

For an ai fact checker, the honest version of most ranking claims is a stated criterion rather than a superlative. Not the cheapest tool, but cheaper than the three tools we compared on 12 November. That sentence carries a scope, a set and a date, and a buyer can act on it without being misled.

The three things an ai fact checker cannot catch

Every honest product page has this section and ours lists nine limits in the archive. Here are the three that matter in this article.

It cannot decide what a name means in the world. If two sentences both use the same product name and one of them refers to a different product that shares the name, the engine treats them as the same subject. The operator has to notice, and the operator usually does, because they know their own catalogue.

It cannot read images, PDFs or slides. Extraction is deterministic over Markdown text. A figure inside a screenshot is invisible, and screenshots are one of the most common places a stale number hides.

It cannot rate the quality of a source. A forum post stating a price is treated the same as the vendor pricing page. Ranking sources is a judgement, and judgements belong beside the ledger entry, written down.

An ai fact checker that claims to do those three is claiming to decide meaning, read pictures and judge credibility. None of that is true of any tool here, and the products that claim it are the ones worth testing before you pay.

[The three limits, each with the claim it cannot decide.]

Three limits drawn as three claims the engine declines to judge
Figure 4: The misses matter as much as the catches.

How to read the four exit codes

The output of a run is a number, and the number is the point of the run.

0  PUBLISH   every claim has a witness
3  REVIEW    inference or estimate, a human decides
2  BLOCK     no source, or a source about something else
1  ERROR     the engine could not run

Two habits make this useful rather than decorative. Treat 2 as a hard stop, because every attempt to soften a block under deadline pressure is how a wrong figure reaches a customer. Treat 3 as work rather than permission, because review means somebody writes down what they decided and why.

An ai fact checker that returns a confidence percentage instead of a code has handed you a feeling. Percentages invite negotiation at eleven at night. Codes end conversations, which is what you want before a launch.

[One run, one code, and the action each code asks for.]

A blocked run showing exit code 2 and the claim that caused it
Figure 5: Blocked, with the reason attached. This is the output that stops a mistake.

What one run leaves behind

A folder named for the date, the time and the draft. Inside it, a report in HTML, PDF, DOC and Markdown, a proof file with a hash, and a source card for every bound claim.

That folder is the difference between an ai fact checker and a spell checker. A spell checker tells you about a typo and forgets by lunchtime. This tells you about a claim, keeps the receipt, and six months later the receipt is still there when someone asks where the figure came from.

The report also survives a staff change. The person who wrote the draft will not always be the person answering the question, and a folder is a better handover than a memory.

[The folder after a run: four report formats, the proof file and the source cards.]

Results folder with four report formats, a proof file and source cards
Figure 6: Everything a reader would need to recheck the decision, in one folder.

Building the ledger: where the sources actually come from

An ai fact checker is only as good as the ledger behind it, and the ledger is the part people abandon. Four sources cover almost everything you will need.

The vendor’s own pricing page. For a price claim this is the only source that counts, because it is the page the company itself will point at. Copy the row rather than the cell, and copy the date from the page footer when there is one. If the page shows three editions, the quote is stronger than a page showing one, because the tier appears in it.

The vendor’s documentation for anything technical. Seat counts, limits, supported platforms. Documentation pages are usually dated, which makes them useful for the date field on the card, and they are usually stable, which makes them less likely to break between writing and publishing.

The primary document for anything regulatory. A guidance page, a standard, a policy. Google’s own material is a good example of a primary source that anyone can read and quote: the generative AI content guidance is dated and specific, and quoting it is safer than quoting an article about it. For anything about how pages are read, the Search documentation is the second place to look.

Your own records for anything you state about yourself. A date you launched, a number of tests you ran, a count of gates that passed. These are the claims a reader can hold you to most easily, and the ones most often written from memory. An ai fact checker will hold you to them, which is uncomfortable and useful in equal measure.

Two rules about how much source is enough. One entry per claim, and no entry that supports nothing. The second rule is the one that keeps a ledger usable after a year, because ledgers rot when they collect links that once explained something to somebody.

An ai fact checker run against a well kept ledger takes seconds. The same run against a ledger that is really a bookmark folder takes longer than reading the draft, which is how people conclude the tool is slow when the input is the problem. The ledger is not the paperwork around the work. For this kind of tool, the ledger is the work, and the run is just the part that tells you whether the work was done.

Where the limits and the licence are published

The nine limits are on the product page before purchase, not in a support article afterwards. The three editions differ on rights rather than features: Personal for your own publishing, Pro for commercial use of your own products, Agency for client deliverables.

If you want to see how the engine behaves before buying it, the honest test is to run it against a page you already published. Most people find one thing, and the one thing is usually a subject mismatch rather than a wrong digit.

Every count here comes from a release built on 20 November 2026 and checked the following morning.

SOVEL SAAKSHI is the product this page describes, and its nine limits are published before purchase rather than after.

An ai fact checker that publishes its own limits is easier to trust than one that does not. Ask any ai fact checker vendor what it cannot decide, and treat a vague answer as the answer. A good ai fact checker is a small, boring tool.

A claim that needs three sources belongs with a claim verification tool rather than a single pass.

Questions people ask before buying an ai fact checker

Is an ai fact checker the same as an AI detector?

No, and the difference decides which product you need. A detector reads style and guesses authorship. An ai fact checker reads claims and checks sources. One answers whether a machine wrote the text, the other answers whether the text is true. If you can only buy one, buy the second.

Does it need an API key or an account?

No. The engine runs offline with no key, no login and no network call, and a release gate enforces that rather than a promise in a paragraph. The trade is that sources come from your ledger rather than from a live search, so the engine cannot invent evidence it was never given.

What happens if I have no ledger at all?

Every claim returns unverified and the run exits 2. That is correct behaviour rather than a bug. The ledger is the input that makes comparison possible, and one page’s ledger usually serves every future draft about the same product.

Can it handle a price list with three editions?

Yes, and this is the case an ai fact checker handles better than a digit comparer. A list naming all three editions supports a claim about any one of them. A page showing one edition supports that edition only.

What if the claim is a range or an estimate?

Ranges and estimates are judged as estimates, which produce review rather than publish. You decide whether the range is acceptable and record the decision with a date. The engine will not quietly convert an estimate into a fact.

How much does it cost, and is it a subscription?

Personal is ₹2,499, Pro is ₹6,999 and Agency is ₹14,999, paid once, plus GST where it applies. The licence is perpetual for the v0.x line, with no renewal, no seat meter and no activation step.

Where do the nine limits come from?

From nine rounds of our own audit, not from a competitor comparison. The PDF font limitation, URL redaction by parameter name, deterministic Markdown extraction, all of it is published before purchase, because a buyer who discovers a limit after paying is a refund conversation.

If you want the editorial method behind the tool, how to fact check AI generated content covers seven checks in order. If you want the command level workflow with the ledger fields spelled out, start with verify AI content before publishing. For the difference between a checker and a detector, read ai hallucination checker. The engine, its prices and its limits are at SOVEL SAAKSHI.

The last thing worth saying is the one that took us longest to learn. An ai fact checker does not make you careful. It makes carelessness visible, which is a different and more useful thing, and it is the reason our own eleven day price mistake has not repeated.

Connected Systems & Architecture

This publication connects directly to the formal SOVEL software registry, production engines, and architectural documentation: