What makes a source worth citing

A useful way to think about this: imagine twenty pages answering the same question, nineteen of which say roughly the same reasonable thing. Attributing the consensus to any one of the nineteen would be arbitrary, so it is stated without a source. The twentieth page, which contains a specific number, a named constraint or an account of what happened when someone tried it, is the one that gets cited — because the claim belongs to it.

Specificity, which is the whole game

Generic claims are unattributable by construction. There is nothing to cite in a sentence that fifty other pages could have written.

The right-hand column is checkable, which is precisely what makes it usable. A system can corroborate a specific claim against other sources; it has no way to evaluate an adjective.

First-hand detail cannot be synthesised

A model can produce a competent summary of any well-documented topic. What it cannot produce is what actually happened when a specific thing was attempted — and that is the category of content with no substitute.

• The failure mode you hit — "the sibling-link test caught an article linking only outside its own cluster" is a sentence no summary generates
• The thing the documentation omits — every non-trivial tool has behaviour that is only learnable by using it
• The decision you reversed — an explanation of why an approach was abandoned is worth more than an argument for it
• The measurement you took — your own numbers on your own system, presented as what they are rather than as a general claim
• The trade-off you accepted — named, with the reasoning, including the part that is unattractive

The second item is the most reliably valuable. Pages that document undocumented behaviour become the canonical reference for it, because there is nowhere else for that information to come from.

Stating limits raises credibility rather than lowering it

This is counterintuitive to anyone writing marketing copy, and it is consistent: the sources that read as authoritative are the ones that say where their claim stops applying.

Agreeing with verifiable facts

A claim that can be checked against other sources and survives is worth more than one that cannot be checked at all. A claim that contradicts well-established facts reduces confidence in everything else on the page.

1. 01Where a standard or specification governs the point, name it — a criterion number, an RFC, a documented threshold.
2. 02Where a figure comes from someone else's research, attribute it and link it, rather than presenting it as your own finding.
3. 03Where your own experience contradicts the common advice, say both: what is generally recommended, what you found, and under what conditions.
4. 04Where you are uncertain, say so. An acknowledged uncertainty is a statement about your reliability that is worth making.

What does not work

The test to apply before publishing

One question, asked of each substantive claim on the page: could this sentence have appeared on any of the other pages answering this question?

If yes, it is consensus text — necessary as scaffolding, and contributing nothing to why this page would be the one named. If no, because it contains a specific threshold, a named condition, an account of a particular failure or a stated limit, then the page has something that belongs to it. A page with five such sentences is in a different category from a page with none, regardless of length.

This is also, not coincidentally, the test for whether the page was worth writing.