Buy intelligence, not data

Web search already ran this experiment.

Every AI product that needs to know something about the web faces the same choice: scrape Google, or use search infrastructure built for machines. Scraping seems free. The pages are right there. And yet serious teams pay for retrieval APIs instead, for reasons that have nothing to do with laziness.

Scraping gets you documents. It doesn't get you meaning. Google's index answers keyword queries for humans; it isn't designed to answer “companies like Stripe but for logistics,” because that question lives in a semantic index someone had to build on purpose. Scraping also gets you HTML, which means a second engineering project hiding inside the first: parsing, deduping, ranking, all of it yours now. And it gets you a maintenance treadmill, because sites change, markup rots, and blocking gets better every year. What retrieval companies actually sell is not access to the web. It's the web, indexed for reasoning, with the churn absorbed on their side of the API.

Nobody calls this buying data. It's buying intelligence about a corpus.

The physical world has no such layer

The data about the physical world is technically public and practically unusable.

Parcels live in three thousand counties, each with its own schema, update cadence, and coding quirks. Ownership hides behind LLCs that resolve to filings in other states. Building footprints, flood zones, wetlands, traffic, imagery: each is a separate feed with its own idea of where things are. The interfaces are built for a clerk at a desk, one lookup at a time.

You can scrape it. Teams do. We watched one spend well into six figures standing up their own pipeline before concluding the pipeline was never the hard part. The hard part is that the sources disagree. The assessor says vacant, the imagery shows a building, the deed names an entity that took three hops to trace. Knowing which source to trust means knowing how each one fails, and an agent pointed at raw feeds relearns those failures one parcel at a time, on your token bill.

Reading scales with tokens. Reconciling doesn't.

A world model built for reasoning

MAIA is an index of the physical world built for reasoning.

Entities resolved to the operating owner, once, not per query. Geometry conflated across sources. Enrichments computed ahead of time. Small, specialized models doing the classification work a frontier model does slowly and badly: is the yard paved, is the roof viable, is that footprint a building or a shadow. Retrieval shaped so a model can ask “industrial parcels with unimproved yard, out-of-state owners, outside the floodplain” and get an answer, not a reading assignment.

We started with property because “where” questions are worth the most there. Solar developers, industrial brokers, housing groups: people whose entire job is bottlenecked on spatial reasoning. But the layer is general. The same index that qualifies a solar site qualifies a clinic location or a charging corridor. The verticals are proof points. The infrastructure is the company.

Pay for the work, not the records

Every dataset we hold is in every plan. Parcels, ownership, buildings, flood, imagery, all of it. We don't charge for access because access was never scarce, and pretending otherwise is a business model with an expiration date. You pay for the analysis you run.

And because the value is judgment rather than records, the judgment has to be inspectable. Every value MAIA produces carries its sources. Any row opens to the records and imagery behind it. A number you cannot check is not intelligence. It is data with more steps.

The web taught everyone this lesson once already. Data was never the product. The companies that mattered were the ones that indexed meaning and let everyone else build on top.

We're doing that for the ground under your feet.