Primary vs secondary sources
- research
- verification
Sources sit on a ladder of distance from the evidence. A primary source is the thing itself: the study, the law’s text, the official documentation, the original dataset, the vendor’s own spec sheet. A secondary source describes or summarizes a primary one — reporting, whitepapers, explainers. A tertiary source aggregates summaries — newsletters, listicles, and, notably, an AI answering from memory. Every hop up the ladder loses a caveat, rounds a number or adds an interpretation.
How meaning bends on the way up
A study abstract might say: “a pilot with 40 participants measured up to a 12% improvement on one task measure”. One level up: “teams got 12% better”. Another level up: “‘big’ performance gains found”. Same research, three different promises — and only the first one is what the researchers claim.
When secondary is fine
Orientation. Deciding what matters, what the vocabulary is, where to look. Secondary sources are how you find the primary ones. Being maximally skeptical of everything is not the skill; directing skepticism is.
When you must trace
- Numbers you will repeat — in a deck, a report, a purchase decision.
- Anything legal, medical or financial.
- Claims that start with “studies show” or “research suggests”.
- Anything a stranger’s money depends on.
How to trace a claim in practice
- Get a pointer: title, authors, year, publisher — ask for the full citation if it is missing.
- Search the exact title; open the abstract or first pages.
- Find the number and read the sentence around it: what does the original actually claim?
- If every chain you follow ends in “an article about a report about a study”, the claim is not research — it is a telephone game. Say so.
A bad example
Statement for your board memo: “The pitch deck says teams using AI assistants ship 40% faster.”
Deck → whitepaper → study → survey. Each hop chose its favorite number. Shipping “40% faster” under your name inherits all four hops’ optimism.
A better example
After tracing: “In one 2024 study of 2,400 teams, routine-task completion improved 40% during the pilot; the authors note the effect may not generalize.”
Same number, honest boundaries, and you can name the document it lives in. That sentence survives a skeptical reader.
Practice
Trace claims to their origin
A vendor's pitch deck cites research about AI-assisted engineering teams. Your manager asks which statements can go into the board memo as-is, and which need tracing first.
Flag every statement that must be traced to its origin before you repeat it.
The AI answer
AI in engineering, as summarized in the deck: A 2024 study of 2,400 engineering teams found a 40% speed-up on routine tasks. TechDaily's coverage of that study called it 'the strongest effect measured in a decade'. AI coding assistants are widely used across the industry. The deck invites you to read the study's abstract before discussing the numbers. An industry newsletter, summarizing the whitepaper about the study, reports that adoption doubled last year.
Hint
Count the hops between each statement and a document you can open.
One item asks you to go read the original — does that need checking?
Why this is the answer
The flagged statements all rest on chains of retelling — a number with no pointer, coverage about a study, a newsletter about a whitepaper about a study — where each hop can bend the meaning. The statements left alone assert nothing checkable or point straight at the primary artifact. Rule of thumb: trace numbers to origins; let generalities stay general.
Transfer
- Negotiations: the other side’s “industry data” gets traced before it constrains your offer.
- Policy and compliance: read the rule, not the summary of the rule.
- Recommendations: “according to research” is not a source — it is an invitation to go find one.
Next
Next: Find the unsupported claim — some claims have no source to trace at all. Those are the ones to catch fast.