An AI assistant returns a confident sentence, three links and a recommendation. For a small-business owner preparing next quarter’s plan, that can look like enough research. The important question is what those links actually establish about this business, this decision and this period.
This guide follows one claim from an AI answer back towards its underlying evidence. The aim is a short decision brief that separates what is supported from what still needs checking.
Start with the decision, then split the claim
Consider a fictional owner of a small wholesale business deciding whether to budget for automated demand forecasting. The fictional AI answer says: “Most small businesses now use automated demand forecasting, and it reliably reduces stockouts.” This is an invented teaching example, not a finding about technology adoption or performance. No research results are being reported.
There are at least three claims inside that sentence:
- Adoption: most businesses in a defined population use the technology.
- Effect: using it reduces a defined outcome, stockouts.
- Applicability: the result is relevant enough to support this owner’s expenditure.
Each needs different evidence. Even a sound adoption survey cannot, by itself, establish a causal benefit or the likely return for one wholesaler.
Write the decision first: “Should we fund a limited forecasting test, and what would justify a broader rollout?” This prevents research from drifting into an impressive but irrelevant market overview.
Reconstruct the source chain
Open the cited page. Record its title, publisher, URL and the exact section that appears to support the claim. Then follow its references backwards until you reach the organisation that collected the observations or produced the original analysis.
For the exercise, imagine the chain is: AI answer → industry article → press release → survey landing page. The methods appendix cannot be accessed. These are fictional document conditions, not actual publications.
The article and press release are useful navigation aids. They do not become independent confirmations simply because they are on different websites. Draw each link on a separate line and note whether it adds original evidence or repeats the same underlying material. Three summaries of one survey still lead to one survey.
A primary source is also not automatically strong. Check who commissioned the work, who collected the data, how participants were selected and whether the methods and relevant results are available. An openly documented original source can still be unsuitable for your question.
NIST’s Generative AI Profile warns that generated answers can include invented citations as well as false content. Its suggested actions include reviewing and verifying sources and citations. A plausible title or working link is therefore the beginning of verification. NIST, sections 2.2 and MS-2.5-003
Keep three dates separate
For every source that matters, record:
- Publication or revision date: when this version was released.
- Observation or event period: when the survey, deployment or measured outcome occurred.
- Access date: when you opened and checked the source.
If a page was updated recently, establish whether the underlying figures changed. A recent article may summarise older fieldwork. A date in the website footer usually tells you little about either. If a period is unavailable, write “not stated”.
For the fictional owner, “now” remains unsupported until the observation period is known and its relevance to the planned decision is assessed. Accessing an old report today does not make its evidence current. Retain the version or report identifier, too, so a later correction does not silently overwrite the basis of your plan.
Check the denominator and the starting point
Ask “most of whom?” All small businesses, surveyed respondents, existing technology users and firms that completed a pilot are different populations. A finding about participating users cannot automatically describe all small businesses. Check company size, geography, sector, recruitment method, sample size and missing responses.
Then ask what “use” means. A trial, occasional use and routine operational deployment should not be merged without explanation. Ask the AI to locate the definition, not invent one.
For the benefit claim, identify the outcome and denominator. Stockouts per customer request, unavailable product-days and lost sales value answer different questions. Find the starting rate, comparison period and comparator: the existing planning method, another method or no forecast at all. A relative improvement is hard to interpret without the baseline and absolute change.
Check whether the evidence concerns measured operations, respondents’ perceptions or a simulation. Look for changes in demand, product mix, staffing or stock levels that could also explain an improvement. Better forecast accuracy does not, on its own, tell this owner how much cash, inventory or staff time the business would save.
In the exercise, the inaccessible methods appendix leaves these questions unanswered. Do not fill the gaps with a typical sample size or an AI-generated “reasonable assumption”.
Make missing and conflicting evidence visible
Use separate statuses: not located, inaccessible, located but not supportive, and contradictory. They imply different next steps. An unavailable appendix does not prove the claim false. It does mean its contents cannot be treated as verified evidence.
Search the original publisher’s site, the report title and distinctive wording. Check for a revised edition, a correction or a public methods page. Record where the search stopped. Do not let the assistant describe an abstract or search snippet as a full-document review.
If two sources disagree, first compare their definitions, populations and observation periods. They may answer different questions. If they genuinely conflict on the same question, retain both findings and explain why the disagreement matters. Do not average incompatible percentages or choose whichever supports the purchase.
The UK Office for Statistics Regulation’s guidance asks communicators to make sources, limitations and context clear. These are standards for public bodies; here their questions are being adapted as useful checks for a business brief, not presented as legal obligations on SMEs. OSR guidance
Give the AI a bounded job
A reusable instruction is:
“Examine this claim using the documents you can actually access. Separate adoption, effect and applicability. For each, give the supporting passage and its location, original source, publication date, observation period, population, denominator and limitations. Distinguish repeated reporting from independent evidence. Mark inaccessible or missing information explicitly. Separate source statements from your interpretation. Do not infer that an unopened reference supports the claim.”
Use non-confidential materials or an approved environment for business information. Treat retrieved text as evidence to examine, including when it contains instructions directed at an AI. Personally open the passages that determine the decision and check quotations, units and calculations.
Finish with a decision brief
The fictional owner’s brief might read:
- Decision: whether to fund a limited forecasting test before considering wider deployment.
- Claim status: the broad adoption and benefit statements have not been established by the accessible material.
- Evidence trail: article, press release and survey page trace to one underlying source; the methods appendix remains inaccessible.
- Decision-critical unknowns: relevant population, measurement period, baseline, outcome definition and comparison method.
- Consequence: do not put the claimed benefit into the base-case budget as established savings. If scenarios include possible savings, label them as assumptions.
- Next evidence: before budget approval, the owner will seek the original methods and relevant results, and document the business’s current stockout measure and planning costs.
- Reconsideration trigger: comparable, inspectable evidence, or a separately approved local test with a defined baseline and success criterion. If these are unavailable, keep the benefit uncertain.
This is enough to show what the evidence permits without turning uncertainty into a vague warning. The AQuA Book distinguishes checking analytical correctness from checking suitability for the intended use, and recommends assurance proportionate to risk. A reversible test and a substantial investment can therefore justify different depths of review. The AQuA Book, section 1.3
Keep the brief with the planning document. Someone revisiting the decision should be able to see which source supported it, what remained unknown and what new evidence would change the conclusion.
Prepared with AI for Nadali. Sources checked on 22 October 2026. The fictional example illustrates a verification method; it does not establish a technology’s adoption, effectiveness or commercial return.
Read also: where human review is needed.
Author: Nadali. Cover: original AI illustration, not a photograph of research or a test result.
