PlaceRouter is built on one bet: models converge, information does not. This essay lays out the reasoning, and then does the less comfortable part, which is to check it against what we have actually measured. Some of those measurements support the bet. One of them went against our own product. Both are here.
Intelligence is getting abundant
This section is opinion, and we are labeling it as opinion.
Models get better at reasoning every few months. They are also getting easier to swap. More and more applications are written so that the model underneath can be changed without rewriting the product around it.
When something improves quickly and can be replaced easily, it is a hard place to build a lasting advantage. We may be wrong about the pace. We do not think we are wrong about the direction. If reasoning becomes abundant, the useful question is what stays scarce.
What cannot be reasoned into existence
Whether the road is open. Whether the room is available tonight. What the air sensor read ten minutes ago.
A model can estimate, summarize and infer. It cannot know the present state of a place unless someone measured it and made that measurement reachable. No amount of reasoning produces a reading that was never taken.
So models can reason, and they still depend on outside sources for current, authoritative information about the world. That dependence does not shrink as models improve. A better model asks for more of the world, more often.
Someone paid to create the truth
A weather reading exists because someone installed a station. Flight status exists because an airline runs an operations system. Grid status exists because a utility maintains a network.
These are not passive datasets. Sensors have to be installed, networks operated, infrastructure maintained, information updated and verified. The AI system that consumes the result carried none of those costs.
We think this matters more as machine demand grows. Humans browse. Machines request, continuously. Intelligence does not remove the value of the source. It raises it.
Models can reason.
Someone still has to measure the world.
The real problem is fragmentation
If the information exists, why is it hard for a machine to use? Because it is scattered, uneven and unstable. Three things we measured:
Sources break quietly. We probed 32 real-world information sources across 17 verticals.
- 15 open, no credential
- 4 needed a free key
- 5 private
- 8 unreachable, or changed shape since they were documented
A quarter of tested sources being unreachable or changed is not an edge case. It looks like the normal condition of real-world data.
Availability is uneven at the field level, not the source level. We audited 44 hotels across 110 pages and asked, field by field, whether a machine could read it.
- Address 100% · Phone 90.9%
- Check-in time 9.1% · Pets 2.3%
- Parking fee 0% · Breakfast 0%
The facts an agent needs to complete a booking are mostly written for human eyes, or not published at all. This is one sample in one domain. The pattern is the point: a single source can be excellent for one field and useless for the next.
Age has a cost. We replayed 15,318 paired transit observations with controlled delays. As the information aged, error rose. The direction was consistent in this replay. The size of the effect depends on the feed and the field. Finding a source that contains the field is not enough. It matters how old the answer is.
What a router does, and where ours lost
The layer we are building does five things: discover the sources that can answer a need, evaluate them, route each field to the best one for that place and moment, verify the answer, and learn from the outcome. The application asks for information. It should not have to depend on any single provider staying up, staying fresh or keeping its format.
That is the design. Here is the evidence, including the part we would rather not have found.
On PM2.5 air quality, across 45 stations and 10,459 observations, our routing lost. The raw best source scored 1.284 mean absolute error. Routing plus correction scored 1.356. Worse. One source was best at all 45 stations, and when one source dominates like that, clever routing adds noise. Two things were worth keeping. A constrained oracle reached 0.328, so the headroom is real even though we did not capture it. And calibration learned per sensor hardware class transferred to sensors it had never seen, with negative transfer in 6 of 43 cases. Single dataset, single window.
On weather, across 20 stations and 580 observations, local information helped and naive transfer from the nearest station hurt. At the 24-hour checkpoint PlaceRouter scored 1.2254 against 1.2632 for the best global fixed source, with a confidence interval that includes zero. That is suggestive, not conclusive. We do not claim a win.
What we take from this: per-hardware calibration was useful in air quality, per-station personalization could hurt in weather, and freshness cost accuracy in transit. Source behavior is specific to the domain. A router has to learn it rather than assume it. These results are shaping how we build.
Where this goes: the provider's terms
This section is about where we think this goes.
The question used to be: which website should I visit? For AI systems it becomes: where should this information come from? And eventually: who can access it, and under what terms?
Who, for what purpose, under what rights, how often, at what freshness. Some information will stay open or free. Some will be first-party, restricted, licensed or paid. None of those is better than another. The point is that they differ, and machines will need a layer that understands both the request and the terms.
If the organizations producing real-world information carry the cost of creating it, they should have a say in how machines access it. Today PlaceRouter routes fragmented real-world information. The terms layer is where we think this leads.
AI decides what it needs.
PlaceRouter decides where to get it.
Every number in this essay comes from our research page, where each result is reported with its method and its limits. If you produce real-world information, or your systems depend on it, we would like to talk.