Free AI Fact-Check Checklist: 42 Checks Before You Publish

Here is the checklist we run on every AI-assisted draft before it publishes: 42 checks in nine gates, ordered so the cheapest checks eliminate the most expensive failures first. It is on this page in full, copy-pasteable, with no email form and no download gate. It came out of auditing 80 of our own published posts, where 68 needed a correction — so every item exists because something specific got past us.

The order is the part that matters most. Gate 1 asks whether a product exists at all. There is no point checking the price of something that does not.

Key Takeaways

• 42 checks, nine gates, run in order. Stop at the first failure and fix it before continuing.
• The gates are ordered by cost. Existence is a 20-second check that invalidates an entire section; arithmetic is last because it assumes the items are real.
• Roughly half of every checkable price in our audit was wrong, and nearly all of the errors flattered the product.
• Gate 6 is the one people skip. Affiliate and tracking parameters travel through AI-assisted drafts invisibly — we found another publisher's affiliate tokens on our own pages.
• Six of the 42 are greppable. The command block at the end runs them in one pass.

How to use it

Run the gates in sequence on one draft. Within a gate, order does not matter. When a check fails, fix it before moving on — a failure in gate 1 usually deletes work that gates 3 through 9 would have spent time on.

A full pass on a 2,000-word roundup covering ten products takes about 40 minutes the first time and roughly 25 once the habit forms. That is more than most people spend, and less than the cost of one correction on a published page.

Nothing here is specific to any particular AI model. The failure modes come from how these systems generate confident, well-formatted text about things they do not have facts for — which is why models make things up in the first place.

Gate 1 — Existence (5 checks)

Does the thing you are recommending exist?

1. Every named product resolves to a live product page on the vendor's own domain — not a search result, not a review site, not an app-store listing alone.
2. The vendor domain is not parked, for sale, or serving a registrar placeholder page.
3. The company behind it shows some sign of current operation — a dated changelog, a recent post, a live status page or a support channel.
4. Every named person, study, report or dataset resolves to something you can open.
5. Nothing is named that you found only inside the AI output and nowhere else.

Caught in our audit: two product names that do not exist anywhere, a parked domain, a domain listed for sale, and a fitness wearable reviewed as a to-do app.

Gate 2 — Category fit (3 checks)

A real product in the wrong article misleads exactly as much as a fake one.

6. Each product actually does the thing the article is about, according to the vendor's own description of it.
7. No product is included primarily because it is well known in an adjacent category.
8. The article's stated scope in the introduction matches what the sections actually cover.

Gate 3 — Price (6 checks)

The single highest-yield gate. Half of every checkable price in our audit was wrong.

9. Every price was read on the vendor's current pricing page today — not carried from a draft, a competitor's roundup, or memory.
10. The billing period is stated and correct: monthly versus annual, and per-seat versus flat.
11. The currency is stated.
12. Claimed free tiers actually exist; free trials are not described as free tiers.
13. Named plan tiers exist under those names, at that level.
14. Retired plans and legacy pricing are not presented as current.

Nearly every price error we found made a product look cheaper or more generous than it is. Errors in that direction are not random.

Gate 4 — Capability (4 checks)

15. Every listed feature is confirmed on the vendor's own documentation or product page.
16. No product is marked down for lacking something it publishes.
17. Limits and caveats the vendor states are carried across — especially where the vendor's own page contradicts the common assumption.
18. Integration, export and API claims are checked against the vendor's integration or developer documentation, not inferred.

Caught: a product ranked below a rival for having "no API" when it sells one, and a product described as scannable when its own page states in italics that it is not.

Gate 5 — Links (6 checks)

19. Every outbound link resolves — no 404, no timeout, no registrar parking page.
20. Redirects are followed to the final destination, and the final destination is what you meant.
21. The anchor text describes where the link actually goes. This is the check that catches the most.
22. No auto-linker artefacts: a sentence boundary or a code token swallowed into a URL.
23. Internal links point at pages that exist and are the closest match, not the first match.
24. Every link to a source supporting a factual claim was opened and read, not just resolved.

Caught: an anchor reading "predictive analytics tools" pointing at an article about image generators, and a link anchored "social media presence" concealing a follower-selling vendor.

Gate 6 — Tracking and disclosure (4 checks)

The gate people skip, and the one with legal exposure attached.

25. No URL carries ref=, via=, utm_, fpr=, rfsn=, aff, partner or gclid parameters you did not deliberately add.
26. No link redirects through an affiliate network — sjv.io, pxf.io, impact, shareasale, partnerstack, firstpromoter, rewardful.
27. Any commercial relationship affecting the content is disclosed in the body, not only in a footer.
28. No commercial domain appears in exactly one post across your whole site with no editorial reason — that pattern is the signature of a sold placement.

Caught: four calls to action carrying another publisher's affiliate parameters. We earned nothing and carried the undisclosed-endorsement exposure anyway.

Gate 7 — Statistics and sources (6 checks)

29. Every number has a named, dated, fetchable source.
30. The source was opened, and the number appears in it as stated.
31. The claim in your sentence is the claim the source makes — a speed finding is not an automation finding.
32. The base year is stated for any market size, forecast or survey, and the figure is not presented in the present tense years later.
33. No weasel attribution: "reports say", "studies show", "according to a study", "experts agree".
34. Any figure you cannot trace is either cut or carries the uncertainty explicitly in the text. There is no third option.

Caught: a consultancy's finding that AI made a task 35–45% faster, restated as AI "automating 45% of coding tasks". Different claim, same number.

Gate 8 — Arithmetic and structure (4 checks)

35. The item count in the title matches the number of items in the body.
36. The introduction's count matches the body.
37. The conclusion's count matches the body.
38. Comparison tables list exactly the items reviewed — no extra row, no missing one, no duplicate numbering.

Caught: a "Top 5" with six entries, a "10 Types" with twenty-five, and a table listing a fourteenth app that the body never reviews.

Gate 9 — Platform documentation and dates (4 checks)

39. Every third-party UI element, report name, metric or setting is checked against that platform's current documentation on the day you publish.
40. No retired product, report or feature is presented as current.
41. Every model name, version number, context window, rate limit and price carries a retrieval date.
42. The post states what could not be verified, and why.

Caught: an analytics guide naming a report renamed years earlier, recommending an attribution model removed in 2023, and instructing readers to use a product retired in 2023.

The greppable subset

Six of the 42 can be run mechanically. This does not replace the gates — it catches what a human skims past on the fourth read.

D=path/to/draft.md
grep -o "](https\?://[^)]*)" "$D"                          # every link: eyeball anchor vs destination
grep -oE "ref=|via=|utm_|fpr=|rfsn=|gclid|sjv\.io|pxf\.io" "$D"   # tracking parameters
grep -oiE "reports say|studies show|according to a study|experts agree" "$D"
grep -oiE "we (spent|tested|ran|timed)|our (test|stopwatch|rubric)" "$D"
grep -oiE "ensure[sd]? compliance|guarantee|eliminates the risk|safe from fines" "$D"
grep -cE "^#{2,3} [0-9]" "$D"                              # numbered sections vs the title's count

One trap worth knowing. Figures in published text often contain narrow no-break spaces (U+202F) between the number and its unit, so an exact-string search for a sentence containing a number can fail silently while the text is right there. Match whitespace as [\s  ]+ when you are searching programmatically.

Two claims that fail a draft outright

These are not "add a citation" problems. Neither belongs in a published page at all.

Research that did not happen. "We spent 30 hours testing 12 tools." "We ran 12 identical trials per tool." Stopwatch timings quoted to the second. If the test was run, publish the dataset, the prompts, the tool list and the dates. If it was not, do not describe one. We found two posts of our own doing this.

Guarantees of a regulatory outcome. "Ensures compliance", "guarantees", "eliminates the risk", "keeps you safe from fines". No vendor control delivers a regulatory result, and on health, finance or legal topics this is the most dangerous sentence pattern there is.

Why 42 and not nine

We published nine checks that catch a fabricated roundup, which is the same structure at gate level. Nine is what you can hold in your head. Forty-two is what you need written down when someone else is doing the pass, or when you are on the fourth draft of the week and your attention is gone.

Use the nine when you are reviewing your own work in the moment. Use these 42 when it matters that the pass is repeatable.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools. We sell software that produces drafts like the ones this checklist is designed to catch errors in. Every example below is a failure in our own published work, not a competitor's.

Where the 42 come from. Between 2026-09-02 and 2026-09-03 we reviewed 80 of our own published posts against their sources. Sixty-eight required a correction printed in the body. Each check above traces to at least one specific defect found in that review; the examples quoted in the blockquotes are all from our own corpus (ANALYSIS, audit dated 2026-09-02/03).

What is not claimed. We have not measured how much of an error rate these checks remove, and we do not publish a figure for it. The honest claim is narrower: each item exists because something got past us that it would have caught.

No download gate, and no file. The checklist is the page. We are not promising a PDF, because we would rather publish the thing than the form in front of it.

No LogicBalls telemetry is used in this guide.

Frequently Asked Questions

Do I have to run all 42 on every post?

Run all nine gates on anything that names products, prices or statistics. On a pure opinion or explainer piece with no external claims, gates 5, 7 and 9 are usually the only ones with anything to check.

How long does a full pass take?

About 40 minutes the first time on a ten-product roundup, roughly 25 once you know the vendors' pricing pages. Gate 3 is most of it.

Can AI run this checklist on its own output?

Partly. It can be asked to list every claim and every link for checking, which is useful. It cannot reliably confirm that a page exists or that a price is current — that requires actually opening the page, and a model asserting it did so is not evidence it did.

What if a check fails on a post that is already live?

Correct it in place rather than republishing, so the page keeps its history, and say in the body what was corrected. Silent edits to factual claims are the thing that makes a correction untrustworthy.

Is 42 an arbitrary number?

It is what the nine gates came to when each defect class in the audit got its own line. Gate 8 exists specifically to catch counts that do not match, so it would be a poor look to round it.

Related reading

• Verified AI Writing: How to Publish AI-Assisted Content You Can Stand Behind
• Nine Checks That Catch a Fabricated Round-Up
• AI Content Errors: The 14 Mistakes That Reach Published Pages
• We Audited Our Own Blog: Products That Don't Exist
• How to Stop ChatGPT From Making Things Up