itsbydavid
    Case study · Real estate

    I made a country's property market readable by AI.

    Guatemala's land data lived in public registries, municipal offices, spreadsheets and the memory of local brokers. There was no single machine-readable source. Now there is one, and any AI can query it.

    Municipalities scored — the whole country
    340Municipalities scored — the whole country
    Official sources and global datasets
    22+Official sources and global datasets
    MCP tools any AI client can call
    8MCP tools any AI client can call

    The problem

    You cannot analyse what nobody has written down.

    Buying land in Guatemala meant asking around. Whether a plot floods, how much of it is actually buildable, how long the drive to the capital really takes, whether the municipality has reliable electricity — that information existed, but scattered across public registries, municipal offices, spreadsheets and people's heads.

    The consequence was not just inconvenience. A first-time buyer and an institutional investor were both making decisions on anecdote, and neither could compare two places honestly. There was no dataset to build a tool on, because the dataset did not exist.

    What I built

    One engine, two front doors.

    The scored data layer

    All 340 municipalities rated from 12+ official sources, 10+ global datasets and 5 purpose-built tools. Buildable land, hazard exposure, electrification, population, drive time to the capital — normalised so two places can be compared honestly.

    Nova Scout

    A free nine-step wizard for individual buyers. It asks who you are first, then weights zones on budget fit, location, purpose, required features and growth potential. Most people finish in five minutes and leave with a ranked shortlist, a map, a PDF and a GPS pin.

    Inteligencia

    Bespoke diagnostics for institutional capital — developers, family offices, foreign investors, project finance. Same data, same engine, scoped to a specific decision.

    The MCP server

    Eight read-only tools exposing the whole knowledge base to any MCP-compatible AI client. Open, free, no API key, no registration.

    The part that matters

    The first source in Guatemala built to be queried by AI.

    Most businesses trying to be useful to AI write better copy and hope. Parcela Nova takes the other route: the data is exposed directly, through the protocol AI clients already speak.

    Point Claude, Cursor, Codex or any MCP client at the endpoint and ask about a place. The model picks the tools it needs on its own — starting with search_municipios(q, department, tier, limit) to resolve a location, then calling for prices, demographics, risk, geography or costs. No scraping, no guessing, no hallucinated statistics about a country most models know little about.

    It is read-only and requires no key, which is deliberate. The value is in being the source an AI reaches for, not in metering access to it.

    What it means

    This is the thing the AI-visibility market keeps describing.

    Every agency selling generative engine optimisation is promising some version of "get recommended by AI." Almost all of them are SEO shops that renamed a services page, and what they can actually deliver is better wording.

    Wording is a competition you can lose to anyone with a copywriter. Being the structured source an AI queries directly is not. That is the difference between optimising for a system and being part of it.

    I built this one solo, end to end — the data pipeline, the scoring, the product surfaces and the MCP layer. It is the same work I do for clients, at a smaller scale and on a shorter clock.

    Bilingual from day one
    ES / ENBilingual from day one
    No key, no registration on the MCP server
    FreeNo key, no registration on the MCP server
    In production and versioned
    LiveIn production and versioned

    What transfers

    What I take to other clients.

    The data layer is the project

    Structuring scattered, inconsistent sources into something worth querying was most of the work. It usually is. The AI part is comparatively short.

    Coverage beats cleverness

    Scoring all 340 municipalities rather than the popular twenty is what makes it trustworthy. Partial data invites the question of what is missing.

    Build the interface once

    One MCP server serves every AI client that speaks the protocol, instead of a bespoke integration per assistant.

    The service this became

    Could your data be queried like this?

    Most businesses are closer than they think — the hard part is usually structure, not AI. Twenty minutes and I'll tell you which one you're facing.

    Book a 20-min call