Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Knowledge Exists and Will Not Be Written Down
Your best content is currently being explained on the phone by someone who will never open a document. Asking them to write an article produces nothing for three weeks and then produces something worse than what they say out loud, because writing makes people formal and formality removes the specifics.
This is the actual constraint on distinctive content, and it is almost never named as one. Teams discuss topics, briefs, and tools, and the binding limitation is that the person with the answer has no mechanism for getting it out of their head that fits their week.
The mechanism that works is asking questions and recording the answers. Thirty minutes produces more usable material than a day of drafting, it requires nothing of the expert afterwards, and it captures the things they would have cut as too obvious to mention, which are usually the things nobody else has published.
Why This Is the Durable Pattern
Content that survives has something a competitor cannot regenerate. A model packaging general knowledge produces what everyone else's model produces. A model packaging what your engineer knows produces something that exists nowhere else, and the difference is entirely in the input.
That is the same distinction that separates content which accumulates positions and then disappears from content that holds through reassessment, described in AI content decay. The model is not the problem and never was. Feeding it nothing but its own general understanding is.
It also produces the specifics that make a page quotable rather than merely accurate. Numbers you observed, failure modes you have seen, the exception that catches people out: these are the things a third-party source cites and a reader remembers, and they only exist in someone's head until somebody asks.
Running an Interview That Produces Material
The failure mode is a pleasant conversation with nothing usable in it. Usable material comes from specific questions about specific situations, not from asking someone to explain their area.
- Ask about the last time, not about generally.
"How do you handle X" produces an abstraction. "Tell me about the last customer who had X" produces a story with numbers, constraints, and an outcome in it.
- Ask what people get wrong.
Experts have a ready list and enjoy giving it. This single question reliably produces the most distinctive material in the whole interview, because the corrections are specific to what they actually see.
- Ask what they would not do.
Where a common approach fails, which shortcut backfires, when the obvious answer is wrong. Negative expertise is rarely published and is what makes advice trustworthy.
- Chase every number.
When they say usually or most of the time, ask what that actually is. Frequently they know, roughly, and that rough figure is original data nobody else has.
- Record it, and keep the recording.
The transcript is the source of truth for everything downstream. Notes are already a paraphrase, and a paraphrase is where invention starts.
- Stop at thirty minutes.
It is the number that gets agreement and it produces enough. An hour gets declined, rescheduled, and eventually cancelled, which yields nothing at all.
Keeping the Draft Honest
The characteristic failure of this workflow is drift: the draft says something adjacent to what the expert said, more confidently, with a number that has been rounded into a different claim. The control is checking assertions against the transcript rather than against memory.
Practically that means every specific claim in the draft should be traceable to something actually said. Where it cannot be traced, cut it rather than softening it, because a hedged version of an unsupported claim is still unsupported. This is the same discipline as verifying an AI draft against primary sources, set out in editing an AI draft into something publishable.
Send the draft back to the expert before publishing, with the specific claims marked. This takes them five minutes, catches the drift you missed, and has a secondary benefit worth more than the correction: it makes them willing to do the next one, because they have seen that what they said was represented accurately.
What This Unlocks Beyond the Article
One interview is rarely one article. A thirty-minute conversation usually contains a main piece, several answers to specific questions, and at least one number worth publishing on its own.
The numbers are the most valuable output and the most frequently discarded. An expert's rough sense of how often something happens, checked against whatever records exist, is original data, and original data is what earns references rather than merely informing readers. The method for turning it into something citable is in publishing original research.
It also solves the authorship problem honestly. An article built from a named person's knowledge, reviewed by them, can carry their byline without fiction, which is what the accountability expectation actually asks for and what a fabricated author profile pretends to satisfy. That distinction is in author entities.
Questions About Expert Interviews
- Why interview instead of just writing?
Because the people who know the answer usually cannot write it down and will happily say it out loud. Asking someone to write an article produces nothing for three weeks; asking them thirty minutes of questions produces more usable material than they would have written in a day, and it captures the specifics they would have edited out as obvious.
- What makes this different from AI-assisted writing generally?
The difference is first-hand expert input. A model packaging real human knowledge produces something a competitor cannot regenerate because they do not have that knowledge. Search engines and assistants otherwise receive the same general material everyone else's model produces.
- How do I stop the draft inventing things the expert did not say?
Keep the transcript and check claims against it mechanically rather than by memory. Every specific assertion in the draft should be traceable to something actually said, and anything that cannot be traced gets cut rather than softened. Paraphrase that drifts into invention is the characteristic failure of this workflow.
- Who should I interview?
Whoever answers the question when a customer asks it. That is usually not the marketing team. Support staff, engineers, the founder, and whoever handles complaints hold the specifics that make content credible, and they are rarely asked because they are not thought of as content contributors.
- How long does this take per article?
Roughly thirty minutes of the expert's time, plus editing. That is the trade worth naming to the person you are asking: half an hour of talking, and they never have to open a document. Most experts who refuse to write will agree to that immediately.
