AI is confidently wrong about neighborhoods. We measured it, then fixed it.

Ask a popular AI what a neighborhood is like to live in, and you will get a fluent, confident answer in seconds. Ask it something checkable and that confidence starts to cost you.

In our testing, one model described a park in Montreal’s Plateau called Parc des Compagnons-de-la-Poste. It does not exist; the real park is Parc des Compagnons-de-Saint-Laurent.

Another recommended the “Laurentian International School” near the Plateau. There is no such school, anywhere.

A third told us Seattle’s Capitol Hill is served by the “L Line.” There is no L Line.

None of those answers hedged. They read like local expertise. They were invented.

This is fixable, and we measured the fix. In a formal evaluation this July, connecting an AI to verified location data, a technique called grounding, sharply cut the invented facts and roughly doubled the verified ones. The rest of this post is what we found, why location trips up AI in particular, and what it means for anyone putting AI in front of a home consumer.

Why AI gets neighborhood facts wrong

AI is not useless here. It is convincing, which is worse. Real estate has gone all in on it, and agents, portals, lenders, and their customers are now asking the same handful of models the same questions. The first one is almost always the oldest question in the business: what is it like to live here?

On the details that decide a purchase, the commute, what is nearby, the schools, how walkable a place really is, the model is often wrong, and it is wrong with the same confidence it is right. A buyer cannot tell a real answer from a fabricated one. Neither can the professional who passed it along. That is the trust problem underneath the AI rush, and it is measurable.

What we tested

We ran a paired evaluation of seven AI models: GPT-5, GPT-5 nano, Claude Opus, Claude Sonnet, Gemini Flash, GPT-4o mini, and DeepSeek Flash. Each model answered the same 490 questions across 46 neighborhoods, twice. Once from memory alone, and once connected to Local Logic data. Then we fact-checked every single claim against live data, not against a pre-written answer key.

We fixed the pass marks before the run and report the conservative end of each result, so the numbers hold up under an audit instead of flattering us. We also chose the neighborhoods on purpose to span the famous and the obscure, from Williamsburg in Brooklyn to Butchertown in Louisville, because the obscure end is where the real test is.

What grounding means

Grounding means the AI answers from verified data handed to it at the moment of the question, instead of from patterns in its training. An ungrounded model answers from memory and fills the gaps by guessing. A grounded model looks up the verified fact first, then answers.

That is what our data does once it is connected to an AI. Same model, same question, except now it is reading real numbers instead of reaching for them.

What changed when we grounded the models

Grounding improved every model we tested, and the biggest gains went to the cheapest ones. Three results stood out.

The answers got richer. With Local Logic data connected, the models produced about twice as many verified facts per answer, and up to 2.2x as many on the strongest models. More real, checkable detail, not more filler. Much of that detail is the measured neighborhood characteristics a listing never describes.

The made-up facts dropped sharply. Hallucinations fell by up to roughly five times on the cheaper models. DeepSeek Flash went from about 19% of its key claims being contradicted by live data down to about 4%. Across the whole study, ungrounded models made 1,861 claims that live data flatly contradicts.

Cheap plus data beat expensive alone. This is the finding that reorders budgets. GPT-5 nano, a small and inexpensive model, reached 97.6% accuracy once it was grounded in our data, edging out full GPT-5’s own 96.9%, at roughly one-seventh the cost per answer. You do not need the biggest model. You need the right data feeding it.

Here is the whole story in one question. We asked how walkable Central West End in St. Louis is. From memory, a top model called it one of the most walkable neighborhoods in the city and cited a walk score up in the 90s. Both overstated it. Grounded in Local Logic data, the same model returned the real figure: a walkability of 8.6/10, meaning most day-to-day needs are within a few minutes’ walk. Solidly walkable, but not the outlier the model had invented. Every claim then checked out.

Why location is the hard case

Neighborhoods trip up AI for reasons that will not resolve on their own. The facts that matter are not sitting in the training data in any reliable way. You cannot crawl your way to a calibrated walkability score. And a model can tell you a place feels lively without knowing how lively, on a scale you can actually compare. The same gap shows up in points of interest and schools: the model knows the category exists, then invents the specifics.

It also gets worse exactly where it matters most. In our test, accuracy from memory slid as neighborhoods got less famous, while the grounded answers stayed high across the board. The model is least reliable in the long tail of lesser-known neighborhoods, and that long tail is where a large share of real transactions happen.

That is the case for grounding in one line. The four walls are on every website. Everything outside them, measured and comparable, is the part a model cannot reproduce on its own.