Medicare Options began with a problem that would be familiar to almost any publisher working with a large information domain.
People kept asking different questions about substantially the same thing.
In our case, that thing was the Medicare coverage available within a county. A beneficiary might want to know how many Medicare Advantage plans were available, whether PPO plans were offered, which carriers participated, whether Special Needs Plans were available, how many plans had a $0 monthly premium, or what prescription drug options existed.
Change the county and many of the same questions appeared again. Change the plan year and the questions remained while the answers changed.
For years, we had approached this as a content problem. A recurring question represented an information need, and an information need could become another publishing opportunity. Research the question, write the answer, publish the resource, and make it discoverable.
There was nothing inherently wrong with that model. Many Medicare questions genuinely require explanation, and a good publisher should provide it.
But Medicare Options began exposing something underneath the content problem.
The questions were multiplying much faster than the knowledge required to answer them.
16.1 The Questions Were Larger Than the Knowledge
Consider only a few questions someone might ask about Medicare Advantage in a county.
How many plans are available?
How many are PPOs?
How many have a $0 monthly premium?
Which carriers offer them?
Are Special Needs Plans available?
Those sound like separate information needs because they are expressed as separate questions. From the publisher’s side, however, they depend on much of the same underlying knowledge.
We need to know which county is being represented and which plan year applies. We need an inventory of the plans available there. For those plans, we may need plan type, premium, carrier, and other classifications. We need to know which authoritative sources establish those facts and which source periods apply.
Once that knowledge exists coherently, the questions begin to look less independent.
To answer how many Medicare Advantage plans are available, count the applicable plan population.
To answer how many are PPOs, select the PPOs from that population and count them.
To answer how many have a $0 premium, select on another attribute.
To identify the carriers, follow the relationships between the plans and the organizations offering them.
The questions differ, but they operate on the same underlying state.
We did not yet call this the Question Space and the Knowledge Space. Those terms came later. What we were seeing in production was the distinction that would eventually require those names.
The number of questions that could be asked about the county was large.
The body of knowledge required to support many of them was considerably smaller.
That changed the publishing problem.
Instead of asking how many Medicare questions we could anticipate and publish, we could ask whether we had represented the underlying Medicare environment well enough for different questions to operate against it.
16.2 We Had Started Building the Knowledge
That required us to represent more than a collection of county statistics.
The county itself mattered because the coverage environment was geographic. The plan year mattered because the same county could have a different set of plans and costs the following year.
Within that geographic and temporal boundary were different forms of Medicare coverage. There were plans, plan types, premiums, carriers, classifications, and relationships among them. Some observations came directly from authoritative records. Others, such as counts or averages, could be derived from those records.
Taken together, the thing we were trying to represent was not simply “Mohave County” or “Medicare Advantage.”
It was something closer to:
The Medicare coverage environment available within Mohave County during the 2026 plan year.
We would now recognize that as a Knowledge Object whose identity depends partly on context. Change the county and we are describing another coverage environment. Change the plan year and the state of that environment changes again.
We did not understand the work in those terms at the time.
We were trying to make Medicare information more explicit and more usable to machines. The implementation evolved as we experimented with different ways of doing that.
But underneath those implementation changes, something more important was happening.
We had stopped organizing the factual problem around individual questions.
We were organizing the knowledge that existed before those questions were asked.
That gave us a different experimental proposition:
If the publisher represents the underlying knowledge coherently, can questions arrive later?
16.3 The Question Could Arrive Later
Suppose the publisher has represented the Medicare Advantage plans available in a county together with their plan types, premiums, carrier relationships, and other relevant attributes.
Someone asks:
How many Medicare Advantage plans are available?
The knowledge needed to answer already exists.
Someone else asks:
How many are PPOs?
No new factual research is required. The question asks for a different projection across the same plan population.
Then:
How many have a $0 premium?
Again, the question operates on knowledge that already exists.
Or:
Which carriers offer plans in the county?
Now the question follows relationships within that same knowledge environment.
The wording of each question is new.
The underlying knowledge is not.
This was the production equivalent of the camera catalog discussed earlier in the book. A camera publisher who maintains weight, price, sensor type, stabilization, video capability, and lens mount does not need a separately maintained factual universe for every possible combination of those attributes.
A new question can select and combine knowledge that already exists.
Medicare was messier, but the principle was the same.
That did not reduce every Medicare question to a database query. A beneficiary still needs explanation. Someone choosing coverage may need help understanding the difference between plan types, interpreting costs, understanding terminology, or recognizing tradeoffs that cannot responsibly be reduced to a handful of fields.
The point was not that structured knowledge could replace communication.
It was that a new factual question did not automatically require another independently authored factual answer.
The knowledge could exist first.
The question could arrive later.
16.4 Then We Put It on the Public Web
This distinction became more interesting once the county knowledge was no longer confined to our internal systems.
We published it.
That gave us a limited but useful opportunity to observe what happened when external generative systems encountered public resources containing the knowledge.
The limitations of that observation were important from the beginning.
We knew what we had published. We knew which facts were present. We knew the questions we asked. We could observe generated responses and the public resources cited with them.
We could not see what happened between those points.
We could not inspect a proprietary crawler’s internal representation of the page. We could not see retrieval scoring. We could not determine exactly how a model had associated a particular assertion with a particular question. We could not establish that one element, format, or publishing change had caused a resource to appear in an answer.
We were observing from the publisher side.
Within that boundary, something useful began happening.
County resources containing the underlying Medicare state appeared in generated answers to factual questions for which we had not independently authored corresponding answers.
A system might return a count, classification, premium observation, or another projection supported by the county knowledge.
The interesting part was not simply that our resource received a citation.
A citation can be influenced by many things.
What mattered was that the information product being returned did not always have to exist on our site in that form beforehand.
We had published the ingredients.
The external system could sometimes construct the projection.
Later experiments gave us an additional way to examine that observation. In some cases, we could change the structured knowledge available through a resource while leaving the human prose substantially unchanged.
That was useful because it removed one obvious explanation.
If we rewrote the article, changed its headings, expanded its coverage, added direct-answer passages, and changed the knowledge representation simultaneously, any subsequent difference in behavior would tell us very little. Too many things would have changed at once.
Instead, we could leave the human publication substantially stable while changing how much of the underlying knowledge was explicitly available.
Then we could wait for the resource to be indexed and ask questions again.
The result was not a controlled experiment. External systems changed independently of us. Indexes changed. Competing resources changed. Generated answers varied.
But we had changed something other than the prose.
We had changed the knowledge surface.
That was enough to keep investigating.
16.5 The Durable Asset Had Changed
Looking back, the most important result of Medicare Options was not a particular answer-engine response.
It was not a citation.
It was not a ranking.
It was the realization that we had begun maintaining something more durable than any of those outputs.
The county knowledge had a geographic identity and temporal scope. It contained a plan population and classifications describing that population. It contained premiums and carrier relationships. It could contain source assertions and publisher-derived observations.
Different questions could operate on different portions of that same maintained state.
That changed what we considered the publishing asset.
Under the older model, a newly discovered question naturally suggested another answer artifact. The question became the organizing unit of the work.
Under the emerging model, the first question became different:
Do we already possess the knowledge required to support this information need?
If the answer was yes, another independently maintained factual publication might be unnecessary.
If the answer was no, the question had revealed something more valuable: a gap in the Knowledge Space.
This was the origin of a principle we would later describe as:
Engineer once. Answer many.
It is not a promise that one Knowledge Object can answer every question. Nor does it mean publishers should stop creating explanations, guides, comparisons, or other human information products.
It describes where the durable factual investment can live.
The publisher maintains the knowledge.
Questions can project across it.
Answers can be produced from it.
And when the underlying reality changes, the publisher has something more coherent to update than a growing collection of independently authored answers.
Medicare Options began as an experiment in making Medicare information more useful to machines.
It ended by changing our understanding of what we were publishing.
We had stopped publishing the questions one at a time and started publishing the knowledge beneath them.