Open the source. Then confirm three things the draft did not: that the identifier is real, that the text says what is claimed, and that the instrument applies to the person you are writing for.
Key takeaways
Fact-checking AI output is not one action, it is four, and skipping any of them is where consultants get caught. Find the primary source. Confirm the identifier actually exists. Confirm the source says what the draft says it says. Confirm it applies to the reader you are writing for. Only the first of those is what most people mean by ‘checking a citation’, and it is the weakest of the four.
The correct source for a rule is the instrument, and for Canadian federal legislation that means the consolidated statutes and regulations at the Justice Laws website rather than a summary of them. The correct source for a Canadian figure is the statistical agency's own table rather than an article about it. A consultancy's report, a vendor blog and a trade-press summary are all downstream, and every hop introduces the possibility that a number has been rounded, rebased or attached to the wrong year.
This is the step nobody is taught and it takes ten seconds. Take the citation, change the identifier to something that cannot exist, and request that instead. If the site returns a page for the fake one that looks like the page it returned for the real one, then the site cannot confirm that any particular section, table or article exists — it is serving a shell and filling it in the browser.
The consequence is specific: on such a site, a successful response is not evidence. You may cite it for general context and you may not cite it as proof that a provision exists. Where the site does return a genuine not-found for a nonsense identifier, you have learned something real and you can rely on the positive result.
Most bad citations in consulting deliverables are real documents that do not say the claimed thing. The test is whether you can point to the sentence. If the draft asserts a threshold, find the number in the text. If it asserts a duty, find the verb.
Statistics is where this fails most often, because a figure travels well and its qualifications do not. A Statistics Canada table carries a table number, a release date and a DOI in its own page — table 33-10-0860-01, Business or organization obstacles over the next three months, third quarter of 2024, released 27 August 2024 is a worked instance. Those three fields are what a citation has to match. Note also the practical trap: on many statistical tables the numbers themselves are loaded into the page after it opens, so an automated reader can retrieve the page and still not have seen the figure. If you cannot see the number, you do not have the number.
This is the check that separates a competent deliverable from a plausible one. An instrument can be real, current, correctly quoted and irrelevant, because its application provision covers someone else. Federal legislation frequently limits itself to federally regulated entities or to a defined class; provincial rules differ by province; a standard may apply only to engagements of a particular type or beginning after a particular date.
So find the section that says who the instrument applies to, and read it. If you are writing for an Ontario owner-managed business and your source governs federal works and undertakings, you have a citation and no support. Name the province in the deliverable, every time.
Consulting output becomes client marketing. Once a claim about performance or benefits is made to the public to promote a product or a business interest, paragraph 74.01(1)(b) of the Competition Act makes it reviewable conduct to make a representation in the form of a statement, warranty or guarantee of the performance, efficacy or length of life of a product that is not based on an adequate and proper test — and the provision states that the proof of the test lies on the person making the representation. Paragraph 74.01(1)(a) covers representations to the public that are false or misleading in a material respect. An invented efficacy statistic that you passed to a client, and the client published, is exactly the shape of that problem.
The same logic runs through private-law claims. Treadstone Law's article on fraudulent versus negligent misrepresentation sets out the distinction that decides which one you are facing, and its explanation of expectation, reliance and restitution damages is the plain-language version of what reliance on a wrong figure is worth.
A strategy consultant drafts a market section with AI assistance. The draft contains a national adoption percentage, a statute section number and a regulator's name. All three look right.
The verification pass takes eleven minutes. The percentage cannot be located in any table on the agency's site and is deleted rather than softened, because ‘approximately’ in front of an invented number is still an invented number. The statute section exists but the Act's application provision covers a class the client is not in; the paragraph is rewritten to name the provincial rule instead. The regulator's name is correct but its former name was used, which would have dated the report to a reader in the sector.
One of three claims survived unchanged. That ratio is normal, and it is the reason verification is a scheduled step in the deliverable rather than a virtue.
None of this argues against AI in research. It argues for using it at the retrieval and drafting end and putting the confirmation somewhere it cannot be skipped. The federal Voluntary Code of Conduct on Advanced Generative AI Systems frames the same idea at the system level through its outcomes — accountability, transparency, human oversight and monitoring, and validity and robustness — and the Privacy Commissioner's generative AI principles make the point that these technologies do not sit outside existing law. Your deliverable is subject to the same rule.
Can I ask the model to cite its sources?
You can, and the citations then need the same four checks. A citation produced by the same process that produced the claim is not independent evidence of it.
Is a government page always a safe source?
The domain is a good signal and not a guarantee. Apply step two before you rely on a page to prove a specific identifier.
What if I cannot verify a figure?
State the mechanism instead of the magnitude. ‘Renewal volumes are concentrated in the first half’ is defensible; an unsourced percentage is not.
A 30-minute call is enough to tell you whether AI pays for itself here.
