Data Intelligence Research

A portfolio of websites is not automatically a portfolio. Most multi-site operators end up with a collection of disconnected properties, each competent alone, none of them making the others stronger. Constellation Architecture is our model for connecting them, and the reason we can describe its failure mode precisely is that we have run into it.

The starting problem

We currently have 131 properties reporting into our data warehouse. At that scale the naive approach stops working immediately, and it fails in two directions at once.

Run them as genuinely independent businesses and you get no compounding. Each property earns its own authority from scratch. Nothing corroborates anything. You have done the work of a portfolio and received the benefit of several small sites.

Run them as clones of one template with the topic swapped and you get the opposite problem, which is worse. Now the properties are not distinguishable, and instead of occupying more of a results page they compete to be the same answer.

Constellation Architecture is the model for the space between those two failures.

Independent, not identical

Every property keeps its own identity and its own job. A research property that publishes measured findings should not read like a services property. A founder's personal site should not read like an institutional education site.

These differences are the asset, not friction to be smoothed away. A research property is credible because it is not selling the thing it is measuring. The moment it starts sounding like the services property, it stops being able to corroborate it.

What the model standardizes is not voice or design. It is the technical layer underneath: how properties are structured, how their relationships are declared, how they report into shared measurement.

Three connective layers

Narrative

A shared idea that explains why these properties exist together. Ours is Digital Karma, the argument that consistent, corroborated signals compound. Without something like it, a portfolio is a list of domains someone happened to buy.

This layer is the one most often faked. A tagline applied across unrelated properties is not a shared idea. The test is whether someone reading two properties would independently arrive at the connection.

Structural

Relationships declared in machine-readable form rather than implied by footer links. Each property's structured data states which network it belongs to, which properties it relates to, and how. A human can infer a connection from a link. A machine should not have to.

This is the layer that converts a narrative into something a search or AI system can act on, and it is the one most portfolios skip entirely.

Data

Shared measurement, so the portfolio can be seen at once rather than property by property. Ours ingests raw access logs and search data across all 131 sites into one warehouse.

The value is comparative rather than aggregate. Portfolio totals are mostly vanity. What matters is being able to ask why one property is crawled by 23 distinct AI systems and a much larger one by 11, and then to check whether the answer generalizes.

Why it compounds

When those three layers hold, authority earned by one property becomes corroboration for the others. A citation of the research property strengthens the case studies. A clearly defined founder entity sharpens every property connected to that person. A framework defined once and referenced from several properties is a claim with several independent witnesses instead of one.

The mechanism is corroboration, which is also why it depends completely on the properties being genuinely distinct. Two near-identical sites saying the same thing are not two witnesses. They are one witness and an echo, and systems appear to treat them accordingly.

The failure mode, with a receipt

We have watched this go wrong on our own properties.

Two sibling sites in the same subject area, deliberately overlapping, drifted close enough that they were no longer distinguishable as separate entities. In August 2026 they were performing comparably, at roughly 518 and 427 daily search impressions. Over the first eight days of September, the first fell to about 49 a day while the second held at 479 and edged slightly up.

No penalty. No technical fault. The most likely reading is consolidation: two properties competing to be the same answer, and a system picking one.

We should be clear about the limits. One pair, a short window, and no visibility into the ranking system. It is a well-controlled correlation, not a proven mechanism. It was also enough to change how we build siblings.

Redundancy is deliberate. Duplication is a bug.

We still run multiple properties against overlapping subject areas on purpose. Occupying more of a results page with genuinely different resources is a legitimate strategy and we have no intention of abandoning it.

The distinction that has to hold is between redundancy and duplication.

Redundancy is several distinct entities addressing the same subject from positions a reader would describe differently. Research, commercial, educational, personal. These reinforce each other because they are not interchangeable.

Duplication is several properties a machine cannot tell apart. Same framing, same claims, same structure. That is not coverage. It is one asset with copies, and you are asking a system to choose.

The practical check we now run: could someone who read both properties describe them differently without mentioning the domain names? If not, they are duplicates regardless of what the content plan says.

What we would tell someone starting one

• Write the shared idea down first. If you cannot articulate why these properties belong together, the connective work will be decoration.
• Declare relationships in markup from day one. Retrofitting this across a grown portfolio is far more expensive than doing it at the start.
• Give every property a job it does not share. Not a different topic necessarily. A different function.
• Measure centrally, compare constantly. The value of shared data is spotting a property behaving unlike its siblings while it still matters.
• Audit for drift. Siblings converge over time without anyone deciding to. Ours did.

The honest summary

Constellation Architecture is not a growth tactic. It is a way of keeping a portfolio from quietly turning into either a set of strangers or a set of clones, both of which happen by default.

The model works when the properties are different enough to corroborate each other and connected clearly enough that machines can see the relationship. Miss either half and you have the costs of a portfolio without the compounding.