How to verify a research report
before you rely on it

Why this is now necessary

Until recently a report from a reputable firm could be taken on trust. In 2025 and 2026 that changed. Deloitte Australia refunded part of a government fee after a report cited a court judgment that did not exist. Fabricated references have been found in published consulting work and in peer-reviewed papers at leading AI conferences. In August 2026 Source Global Research found that 70 percent of 3,887 companies would not trust a report prepared with AI. The tools that produce these errors are now used, disclosed or not, in almost every research organisation. The reader has to check.

Step one: trace five citations

Choose five citations at random, weighted toward the ones supporting the most important conclusions. For each, find the source. Does it exist. Does the page or paragraph say what the report claims. Is the figure the same. A fabricated citation usually has a plausible author, a plausible title and a journal or publisher that exists, but the specific article does not, or it exists and says something different. If one of five fails, the report has not been checked by its authors, and nothing else in it can be relied on without checking.

Step two: test the sample

If the report rests on interviews or a survey, ask three questions. Who exactly was interviewed: roles, organisation types, tenure. How were they recruited: directly, through an expert network, through an online panel. How was each verified as being who they claim. A report that cannot answer these has an unverified sample. Given that 30 to 40 percent of online panel responses are compromised by Kantar's estimate, an unverified survey sample should be treated as decoration.

Step three: find the AI

Ask the author in writing: where in this report was generative AI used, and where was it not. Acceptable answers include transcription, coding, drafting and formatting. Unacceptable answers include as a source of facts, figures or citations, or any refusal to say. A firm with nothing to hide will answer in a sentence. Look also for the tells of undisclosed AI drafting: uniform paragraph length, hedged conclusions, generic recommendations that could apply to any company, and citations that are all from the last eighteen months.

Step four: find the person

Who signed this. Not the firm, the person. A report with a named author who did the interviews and takes a position is a different document from one issued under a logo. If no individual will put their name to the recommendation, treat it as a summary rather than advice.

Step five: decide the weight

Having done the four checks, decide how much the report should influence the decision. Passed all four: rely on it as you would on your own team's work. Failed the citation check: do not use it without independent verification of every claim you intend to act on. Unverified sample: use the qualitative colour, ignore the numbers. Undisclosed AI: ask for a corrected version with disclosure before paying the invoice.

What a verifiable report looks like

The Theory Standard requires every claim to trace to a verified interview or a checked document, a provenance appendix listing each participant by role and verification method, a written statement of where AI was and was not used, and a named researcher's signature on the recommendation. We publish it so that clients can hold us to it, and so that anyone can apply the same tests to any other report.