Verify AI Answers Against Your Sources
Paste the AI answer with links to the policies or sources you approved. The check opens each link and quotes the passage that backs or contradicts the sentence citing it.
On 26 September 2026 we ran it on an answer that cited three GOV.UK pages. It caught the one sentence that said 26 weeks where the page says day one. It works on any page a link can reach, such as policy pages, government guidance and papers. It cannot check against a file you upload or a private knowledge base. A page behind a login comes back unverifiable. It is free once a day on one model.
Looking each one up on Crossref and the live page, then reading the source. Ten to twenty seconds.
Deep takes a minute or two: four pro models answer, two critics review, one mediator writes the final call. Leave this page open and the result lands here.
Citation Checker runs the moment you sign in, on the text you just pasted. Nothing to enter twice.
A saved check is kept for 24 hours.
You've used up today's free runs on this check.
Come back tomorrow for another free run.
Checking a reference list or a DOI? Use the AI Citation Checker
Which tools check AI answers against approved sources?
Three kinds of tool do this job. The right one depends on where your approved sources live. If they are public web pages, you can check an answer on this page in seconds.
| Kind of tool | What you give it | What it checks | Best for |
|---|---|---|---|
| Grounding checks in a chatbot platform | Your documents, uploaded or connected ahead of time | Each answer against your own document store | Teams running an AI assistant on internal policy. We do not offer this. |
| Citation checks (this page) | The answer, with links to the pages it relies on | Each cited sentence against the page it cites | Answers that cite public policies, guidance or papers |
| Claim checks | One claim, or a whole draft | Whether each claim matches what is published | Sentences that cite nothing at all |
Only the middle row runs on this page. For sentences with no source, use the free claim check. For a whole draft, the AI fact checker runs both checks together.
How the check reads your sources
An answer can link to the right policy and still get it wrong. The link works, so nobody opens it. This check opens every link and reads what the page says.
Paste the answer with its links
Keep each link next to the sentence that relies on it. A DOI, the permanent ID on a published paper, works too. So does an Author (Year) reference.
The check opens every link
It fetches each page live, the way a visitor with no account would. A page that returns 404 counts as nonexistent. A page behind a login counts as unverifiable. No model decides either one.
A model reads the page against your sentence
It must quote the passage that backs or contradicts the sentence. With no quote, a sentence cannot pass. Numbers, dates and the direction of a rule must match.
You get a verdict for each source
Each source comes back supported, misrepresented, nonexistent or unverifiable, with the quote beside it. The whole answer is marked all supported, needs fixing or unsupported.
What this check cannot read
It reads pages a link can reach and nothing else.
-
Sources must be links, not uploaded files
You can drop a file to fill the box, but its text becomes the answer to check, never a source. If a policy only exists as an internal file, give it a public link or check it by hand.
-
Pages behind a login come back unverifiable
Paywalls and sign-in walls block the fetch. An intranet address our server cannot reach at all comes back nonexistent. From outside, it looks the same as a dead link.
-
Sentences without a link are not checked
This check tests the citations that are there. It does not flag a sentence that should have cited a source and did not.
-
One check covers up to 8 sources
An answer can run to 5,000 characters. For a long page, the check reads the top and then the parts closest to your sentence.
To do the same checks by hand, read how to check if AI citations are fake.
A real check on three GOV.UK policy pages
Council run, 26 September 2026 · GPT, Claude, Gemini and Grok · 11.5 seconds · all four agreed on every source
We wrote a three-sentence answer the way an HR assistant might reply to a part-time worker. Each sentence cites a GOV.UK guidance page. We got the third one wrong on purpose. The free one-model run gave the same three verdicts in 4.6 seconds.
"Here is what UK law gives you as a part-time employee. Most workers on a five-day week must get at least 28 days of paid annual leave a year, which is 5.6 weeks (https://www.gov.uk/holiday-entitlement-rights). If you are too ill to work, your employer pays Statutory Sick Pay for up to 28 weeks (https://www.gov.uk/statutory-sick-pay). You can ask to work flexibly, but only once you have worked for your employer for 26 weeks (https://www.gov.uk/flexible-working)."
Holiday entitlement: Entitlement - GOV.UK
SOURCE SAYS: “Most workers who work a 5-day week must receive at least 28 days' paid annual leave a year. This is the equivalent of 5.6 weeks of holiday.”
The source directly states this requirement. The draft accurately captures both the 28-day minimum and the 5.6-week equivalence for five-day-week workers.
Statutory Sick Pay (SSP) : Overview - GOV.UK
SOURCE SAYS: “It’s paid by your employer for up to 28 weeks.”
The source directly confirms that Statutory Sick Pay is paid by the employer for up to 28 weeks when the worker is too ill to work.
Flexible working: Overview - GOV.UK
SOURCE SAYS: “All employees have the legal right to request flexible working. Employees can make a request for flexible working from their first day in a job.”
The source explicitly states employees can request flexible working from their first day, not after 26 weeks. The draft inverts the actual legal position by adding a 26-week waiting period that does not exist in the source.
In our run, all three links worked and all three were the right GOV.UK pages. A link checker would have passed the whole answer. The 26-week error showed up only when a model read the page.
Run and written up by Arun Agrahri, Founder, TrueStandard, on 26 September 2026.
Frequently Asked Questions
Three kinds. Grounding checks in chatbot platforms compare answers with documents you upload. Citation checks, like this one, open the pages an answer cites and test each sentence against its page. Claim checks test sentences that cite nothing. If your policies are public pages, paste the answer here with its links.
Not as sources. The upload box only fills in the answer you want checked. The check reads the pages that answer links to, and it has no private document store. If a policy is on a public page, link to it. If it only exists as an internal file, this check cannot read it.
Only if the page opens without a login. The check fetches each link the way a visitor with no account would. A page behind a login comes back unverifiable. An intranet address our server cannot reach comes back nonexistent, because from outside it looks like a dead link.
The sentence comes back misrepresented. The source is real, but it never says what the answer claims. If the passage could sit in a part the check did not read, such as a figure, it comes back unverifiable instead. The note says what to read.
No. It checks the citations that are there. A sentence with no link is not tested by this check. To test the claims between the citations too, run the answer through the AI fact checker. It runs the claim check and the citation check together.
Up to 5,000 characters, with up to 8 sources per check. For a long page, the check reads the top of the page and then the paragraphs closest to your sentence.
Yes, once a day on one model, with no signup. Paid plans run four models from different vendors on the same sources. Each source gets the majority verdict, a tie goes to the more cautious one, and you see how each model voted.
Every run makes a receipt that stays private unless you share it. It holds the text you pasted and the result. We fetch sources straight from the sites that publish them. Your content is never used for model training.
Powered by TrueStandard
Four independent models check every claim in your draft.