Checking a Claim Before You Publish: The Free Tools and the Order to Use Them In

Most errors in published articles are not lies. They are claims that were true somewhere, once, and were copied without anyone going back to check. A price from last year. A statistic from a report that measured something slightly different. A screenshot that was real but of a different product. None of these takes long to catch if you check the right thing, and all of them survive indefinitely if you check the wrong one.

What follows is the sequence we run on every ToolNest article before it goes out. It uses nothing you need to pay for. It does not catch everything, and the last section is about what it misses.

First, name the type of claim

The mistake that makes fact-checking feel slow is treating every claim the same way. They are not the same. A sentence like "Otter's free tier caps conversations at 30 minutes" is a vendor fact; "40% of meetings could have been an email" is a statistic; "the CEO said the product was finished" is a quote; and a screenshot of a pricing page is an image. Each of those has one tool that answers it and several that waste your time.

Claim type	What settles it	Tool
Price, limit, feature	The vendor's own page, today	Their site, then Wayback Machine to record it
Statistic	The report or dataset that produced it	Quoted search, then the earliest link that cites a source
Quote	The original transcript, post or interview	Quoted search; the publication's own archive
Image or screenshot	Where it first appeared	TinEye, Google Lens
Viral or political claim	Whether a fact-checker has already ruled	Fact Check Explorer
Historical "the page used to say"	An archived copy	Wayback Machine

Vendor facts: the page, then the archive

Go to the pricing or documentation page and read it. Not a review of it, not a comparison table, the page. This sounds too obvious to write down and it is the step most often skipped, because a comparison site already has the number in a tidy table. That table was accurate on the day someone built it, and pricing pages change under a stable URL without any notice.

Then record what you saw. Open the Wayback Machine, paste the URL into Save Page Now, and you have a dated copy you can link to. Every source on this site carries a "checked on" date for the same reason: when Fathom or Granola change a plan next quarter, the article is correct as of a date rather than silently wrong. If the archive already has a capture from the last few days you can use that; the calendar view shows every capture with a blue dot for a clean save.

The limitation is real. Many modern pages load their content with JavaScript after the page arrives, and the archived copy is a blank frame with a spinner. When that happens, take a screenshot with the date visible in your system clock, and keep it. It is a worse record than an archive link, and it is far better than nothing.

Statistics: follow the chain to the end

A number in an article was almost always copied from another article. Your job is to find the end of the chain. Search the exact figure and its subject in quotation marks. Sort mentally by date and open the earliest one that names a source. Follow that. Repeat until you are reading the report, the survey methodology or the dataset — or until the trail ends at a post with no citation.

When you reach the origin, check three things: what was actually measured, when, and of whom. "Users spend 40% of their time in meetings" turns out to be a survey of 300 managers at large American firms in 2019, which is not nothing but is not "users" either. Write the qualified version. It is less quotable and it is true.

When the trail ends without a source, the number has no known origin. That is a finding. Either drop it or publish it with that stated, and we would drop it. This is also the check that AI research tools fail most reliably; they will hand you the number with a citation to an article that quotes it, which looks like a source and is not.

Images: find the first copy

A screenshot proves that a page looked like that on someone's screen at some point. It does not prove when, whose, or whether it was edited. Reverse image search answers the first two questions often enough to be worth the thirty seconds.

TinEye is the tool for provenance because it can sort results by oldest, which tells you where the image first appeared on the indexed web. Google Lens has a much larger index and is better at identifying what is in the picture, but ranks by relevance rather than age, so it is the second tool rather than the first. Run both. If the oldest copy is on a satire account or attached to a different product, you have your answer.

Both are weak on the same things: heavy crops, filters, and images that were generated last week and have never been posted before. A negative result means "not found", not "original". For a screenshot that matters — a pricing page, a message from a vendor — the stronger check is to reproduce it yourself, which is faster than any search.

Viral claims: see if the work is already done

Google's Fact Check Explorer searches the published fact-checks of organisations that mark their work up with ClaimReview structured data — the same markup that puts a "Fact check" label under a search result. Paste the claim or a person's name and you get a list of verdicts with links to the full pieces.

Its coverage is what you would expect from who does fact-checking: strong on politics, health and viral news, thin on software, pricing and anything technical. For our subject matter it rarely has a direct hit. It is still worth the query, because when it does, someone has done in a day what would take you an afternoon, and their working is shown.

Quotes: context is the check

Search the quote in quotation marks and find the original — the interview, the post, the transcript, the earnings call. Then read the paragraph either side. The quote is usually real; what is missing is the qualifier before it or the reversal after it. If the original is behind a paywall, the Wayback Machine sometimes has it, and a library card often does. If you cannot find the original at all, attribute it to where you found it — "quoted in" — rather than to the person, because you do not actually know they said it.

The five-minute order

1. Read your draft for claims, not for prose. Mark every number, price, quote and "always" or "never". Most paragraphs contain one or two.
2. Sort them by type using the table above. This takes a minute and saves ten.
3. Vendor facts first. They are quickest, they are the ones readers act on, and they are where this site's own corrections have come from.
4. Statistics next. Budget the most time here. A chain of five articles takes five minutes to walk.
5. Images and quotes last, because they are usually fewer.
6. Write down what you could not confirm and either cut it or say so in the text. Never smooth it over.

What this does not catch

A primary source can be wrong. A vendor's pricing page can be out of date on the vendor's own site, a report's methodology can be bad, a transcript can be mis-transcribed. Everything above gets you to the origin of a claim; none of it evaluates the origin. That takes domain knowledge, and it is the part that cannot be reduced to a tool.

It also does not catch what you did not think to check. The claims that survive to publication are the ones that sounded obviously true, and those are exactly the ones nobody marks in step one. The only defence we have found is a second reader who has not seen the draft, which is why our publishing week has a review day that is not the writing day. It is slower. It is also the reason this article has a source list rather than a corrections list.