Why is the consensus answer not enough?
The consensus answer is not enough because AI engines already hold it. When a thousand pages say the same thing about a topic, the engine can answer without citing any of them in particular, or cites the most authoritative. A new page repeating the consensus adds nothing the engine needs and earns nothing.
What earns a citation is a page that extends the answer: a specific mechanism, an honest insight, a precise example, a comparison the others did not draw. That extension is information gain, and it is the only content strategy that compounds as engines get better.
What are the honest sources of information gain?
The first source is a named mechanism from your own product or practice: how the fact gate works, what the worth-it verdict compares, why exact matches earn autonomy first. Nobody else can describe your mechanism, and an engine asked how such things work can cite you for it.
The second is a structural insight from experience: that the month-end close shrinks because matching happened daily, not because the close was automated; that the junior hire comparison is the wrong comparison. The third is a worked example with the arithmetic shown and labelled as an example. The fourth is a comparison nobody has drawn clearly, such as a Digital FTE against an agent framework.
- Named mechanism: how your product actually does something
- Structural insight: why something works the way it does
- Worked example: arithmetic shown, labelled, inputs left honest
- New comparison: two things set side by side clearly for the first time
What is not information gain?
Invented statistics are not information gain; they are novelty without truth, and engines that cross-check sources will find no corroboration and lower trust in the page and the domain. Fabricated case studies, unnamed studies, and quotes from nobody are the same failure. Restating the consensus in different words is not gain either; it is the same information with a different surface.
The test is whether a careful reader could verify the new thing. A mechanism can be verified against the product. An insight can be tested against experience. An example can be recomputed. A comparison can be checked. A statistic with no source cannot, and it is the one that does the damage.
How do you produce it reliably?
Produce it reliably by starting every piece with the question: what does this piece say that the index does not already have? If the answer is nothing, either find the mechanism, insight, example, or comparison that would change that, or do not write the piece. The SPEC-035 standard on ftes.ai makes this a checklist item for every article.
The richest source is the product and the practice around it. Every real mechanism, every honest limitation, every design decision with a reason behind it is gain that only the people who built or run the thing can supply.
How does it connect to citation?
It connects to citation because an engine assembling an answer from many sources attributes specific claims to the sources that made them. The consensus is attributed to the most authoritative general source; the specific mechanism is attributed to the page that described it. Pages with gain collect the specific attributions.
Measure it by sampling prompts that ask for specifics, such as how does an AI employee earn autonomy, and recording whether your page with the mechanism is cited. Over time, a library built on gain is cited for the specifics across its whole cluster, which is the position worth holding.
What does information gain look like in one piece?
As a labelled example, a piece about measuring AI employee productivity could restate the consensus that AI should be measured on outcomes rather than activity, which every source already says. Its information gain is the specific mechanism nobody else can describe: that on ftes.ai productivity is three measures read together, outcomes against explicit targets with evidence per item, the approval rate with send-back reasons, and cost per outcome from tokens per role, and that the performance view shows only what the platform measured, marking the rest as unmeasured.
That mechanism is checkable against the product, it is not in the index anywhere else, and an engine asked how an AI employee's productivity is measured can attribute it to the page that described it. The structural insight beside it, that task counts fail because activity is cheap and a hundred drafts nobody uses is a hundred tasks and no value, is the second kind of gain: an honest observation from running the thing. The piece adds no statistic, because it has none it can verify, and it loses nothing by that; the gain is in the mechanism and the insight.