Turn fragmented evidence into a decision-ready research system.

Research protocol: from question to a reproducible conclusion

Start every commercial research assignment with a decision statement: what choice will be made, by whom, within what time frame and using which evidence. Define the primary entity and comparison set before collecting sources. A study of AI business models needs a different evidence base from a study of public-company margins or software procurement rules. Select an observation period, inclusion and exclusion criteria, relevant markets, core terms and a versioned source register. This planning step makes gaps visible and protects the research from becoming a collection of whichever sources were easiest to find.

Preserve a claim-level record with the original source URL, authoring organization, publication date, extraction date, proposition, qualifying language and any known conflict. For a financial metric, capture the denominator and period. For a company-reported product feature, state that its origin is the company. For a strategic interpretation, label it as analysis rather than measured fact. A source can be credible for one proposition but unsuitable for another: official price lists do not establish realized revenue, while an anecdotal customer report does not establish industry prevalence. Reproducibility depends on matching each claim to the type of evidence it can legitimately support.

Information-gain test: what does the synthesis add?

A research page should not stop at describing known terms. Construct a comparison that makes unlike claims comparable, derive an explicit decision tree, expose a counterexample, or show how assumptions change a scenario. For instance, comparing product-led and enterprise-sales businesses becomes more useful when it separates who buys, how value is measured, what onboarding costs and why renewal occurs. State which observations are missing and which conclusions remain conditional. The resulting synthesis should clarify a choice even when available data do not permit a numerical ranking.

Use a counterevidence table with four fields: working hypothesis, evidence supporting it, strongest competing explanation and test needed to distinguish them. Suppose an AI startup reports strong product use but weak revenue conversion. The problem may be packaging, buyer identity, free-tier boundaries or a costly workflow that users value but budgets do not cover. A general statement that interest is high does not resolve those alternatives. A research product should show the next observation that would change the recommendation and identify any irreversible decision that should wait.

Research freshness and citation controls

Classify sources by volatility. Definitions and established methodological documents may remain valid for years; prices, product capabilities, event programmes, rules and public-company figures may require active review. When a substantive claim changes, keep a short change log recording the old position, replacement evidence and resulting decision implication. Do not refresh dates without checking the supporting material. For passages intended to be quoted or retrieved independently, include entity, geography, period and scope directly in the paragraph so the claim remains meaningful outside its page context.

A final research record should state what is known, what is inferred, what remains unknown, what would falsify the analysis and which business decision is now better informed. This makes the output useful to founders and executives even if they never engage a consultancy. It also supports later updates, because the evidence trail reveals which passages need revision when a source changes.