The Research Method: From Vague Question to Publishable Answer

Everything factual on this site goes through the same six-step loop. It's slower than guessing, it's slower than summarising what everyone else already wrote, and it's the reason I've published very few corrections.

Here it is in full, including the step most people skip.

The six-step loop

1. Frame the question until the answer could be wrong.
2. Source — gather, then rate what you gathered.
3. Verify the claims you intend to make.
4. Gap pass — find what you still don't know.
5. Write, marking uncertainty as you go.
6. Publish with a correction path.

1. Framing: make it falsifiable

"Is eSIM worth it" is not a researchable question. "What is the cost per usable gigabyte of eSIM versus local SIM across countries I actually visit" is, because there's an answer that could come out against my expectations.

The test I apply: write down what would change your mind before you start looking. If nothing would, you're not researching, you're assembling support. That single sentence, written down, has redirected several pieces before I wasted a week on them.

2. Sourcing, and rating what you found

Gather widely, then rate each source in about 40 seconds on three axes:

• Proximity. Is this the origin of the claim, or a description of someone else's description? Chase every claim back to its source. A surprising proportion of confidently repeated statistics terminate at a blog post citing a blog post citing nothing.
• Incentive. What does this source gain if I believe them? Not disqualifying, but it tells you where to look harder. A vendor's benchmark of their own product isn't worthless; it's just a starting point.
• Method. Can I see how they know? Sample size, dates, and what was measured. If the method isn't stated, the number is decoration.

3. Verification

This is the step that earns the whole method. For every claim I intend to publish:

• Open the primary source. Actually open it. Citations that look right and don't say what they're claimed to say are common enough that this catches something in roughly one piece in three.
• Check the date. A statistic that was true in 2019 may be badly wrong now, and it will still be circulating with no date attached.
• Check the base rate. "Tripled" means nothing without knowing what it tripled from.
• Look for the strongest disagreement you can find. If nobody credible disagrees, that's information. If someone does, you need to deal with it in the text rather than around it.

4. The gap pass

Before writing, list everything you still don't know that a careful reader would ask. Then decide, explicitly, for each one: find out, or state that you don't know.

Those are the only two acceptable options. The third — quietly writing around the hole — is what produces articles that feel authoritative and collapse under a single follow-up question. Readers can feel the difference even when they can't name it.

Stating uncertainty costs you nothing with the readers worth having. "I couldn't establish this and here's why" is a sentence that has never once lost me a subscriber, and it has gained me several corrections from people who did know.

5. Writing, with uncertainty marked

Three rules I hold to:

• Distinguish measured from estimated from assumed — in the sentence, not in a footnote. "I measured", "I estimate", and "I assume" are three different claims and the reader deserves to know which one they're getting.
• Attach numbers to their method. "$4.20 per usable gigabyte, median across 14 countries over 18 months" is a claim someone can check. "$4.20 per gigabyte" is trivia.
• Write the counterargument properly. Not a strawman you demolish — the strongest version, stated fairly. If you can't state it fairly you don't understand it yet.

6. Publishing with a correction path

Every factual piece gets a visible way to tell me it's wrong, and corrections get made in the open with a dated note rather than a silent edit.

This feels exposing and does the opposite of what you'd expect. A visible correction log is one of the strongest trust signals available, precisely because so few people are willing to keep one. Silent editing, if anyone notices, costs far more than the original error did.

What it costs, honestly

The obvious objection to this method is that it's slow. It is. Here's roughly where the time goes on a typical explainer, from the last four I tracked.

Sixteen hours for a piece that takes twelve minutes to read. More than half of that is spent before a word gets written, which is the part that feels wrong and is the part that does the work.

Two things make this sustainable. First, the sourcing and verification for one piece frequently seeds two others — the reading isn't consumed by the article, it accumulates. Second, publishing twelve well-researched pieces a year has produced better results for me than publishing a hundred summaries would have, on every measure I track including the ones that pay the bills.

Three times the method caught something

Abstract methods are unconvincing, so here are three specific catches from the last year.

A statistic that terminated in nothing. A widely-cited figure about remote work productivity appeared in four articles I read. Following the chain: article cited article cited a press release cited an unpublished internal survey with no methodology and no sample size. The number may well be true. It's unusable, and I said so in the piece instead of repeating it.

A source that didn't say what it was cited for. A claim about eSIM adoption rates attributed to an industry report. I opened the report. It gave a figure for eSIM-capable devices, which is a completely different thing from devices with an active eSIM profile, and about four times larger. That's the single most common failure I find, and it's only findable by opening the source.

My own contradiction. The gap pass caught me writing a summary that said twelve automations died from tooling changes, while the breakdown further down the same post said five. Nobody had flagged it; I found it by listing what a careful reader would ask. It's now the third entry in the corrections log, because I'd already published it.

One of those three was caught before publication and two after, which is roughly my real ratio. The method reduces errors; it does not eliminate them, and a corrections log is what covers the difference.

Questions I get about this

Found an error in this piece? Tell me — corrections get made in the open with a dated note, never a silent edit. The log is public.