Restaurant GIS and location intelligence: what changed between a hunch and a territory

The territorial layer keeps getting cheaper while the siting mistake costs the same

A territorial data layer now runs between 300 and 1,200 USD per evaluated site, and that range is what turns GIS into the cheapest filter in the trade, because closing a badly located restaurant in month 14 takes the whole investment plus severance. The measurable trend of 2026 is the falling price of the layer, not better modeling: mobility and cadastral data that only a twenty-unit chain could buy in 2019 are now within reach of a single-location operator. The hard signal comes from the survival side: in the United States, 51.4% of restaurants pass five years and barely 34.6% reach ten, according to the U.S. Bureau of Labor Statistics business survival analysis from 2024. If you run one to three units, buy the layer per site instead of the annual platform. Past eight units, the subscription pays for itself in the first rent renegotiation.

Where the foot traffic comes from now outranks how much of it there is?

Two blocks with identical pedestrian counts behave in opposite ways, and what separates them is the ORIGIN of those people, not their number. Someone walking to an office sustains low-ticket corporate lunch at four visits a week;

someone crossing the block toward somewhere else sustains nothing, not even with a flawless storefront. Origin-destination models built on anonymized telecom data have been getting cheaper for three years and are already standard in any serious 2026 study. That Bogotá chain that opened three units in 2024 shows it without decoration: two closed before month 18, and the third one, nine blocks from the second, bills 2.3 times more today. Same concept, same team, same supplier. Always ask for an origin-destination matrix broken out by time band, and if the study on offer carries nothing but one Saturday's count, refuse to pay for it. GIS does not predict what you will bill; it bounds the odds of not reaching month 24 alive, and that distinction rewrites the whole negotiation.

From forecasting success to bounding the failure

A serious territorial model hands you a sentence like this one: in this polygon, 78% of the food-service locations opened since 2021 closed within two years. With that figure on the table you stop negotiating a rent and start negotiating a risk premium, collected in grace months, in stepped lease escalation, or in landlord participation in the buildout. The industry spent years selling 0-to-100 suitability scores nobody could audit. The 2026 correction runs toward risk intervals with visible assumptions. Demand from your provider the polygon's base closure rate, the observation window, and the number of locations counted, or the score is worth nothing at all. The claim that nine of ten restaurants die in year one is false, and dragging it around gets expensive inside any location model. Real first-year failure among independents in the United States is 17%, according to the study by UC Berkeley economists (Parsa and colleagues) circulated by Oregon State University in 2024.

The 90% myth collapses, and with it every model calibrated on fear

A model calibrated on the myth punishes polygons that work and pushes the operator toward expensive rents in oversupplied corridors, which is exactly where the margin evaporates. The healthy trend of 2026 moves toward recalibration with real base rates per country and per format, because 17% a year does not behave like 90%. Before accepting any report, ask which base rate the model was trained against. If the answer is vague or quotes a percentage with no published source, you are buying superstition dressed up as a map. When a development agency hands seed capital to forty food-service MIPYMEs without filtering location, it accepts in advance that part of the portfolio will die for a predictable reason, and that mortality later gets read wrong. It gets logged as a failure of youth employability in hospitality when it was a geography error. Context matters here: youth informality in Latin America reaches 62.4% and female informality 54.3%, according to the ILO and ECLAC Labour Overview of Latin America and the Caribbean 2024.

Location reaches public policy because formal employment depends on it

Every closing restaurant sends people back into that statistic. Masterestaurant S.A.S. contributes the operating layer that turns a map into a decision: expected average ticket per polygon, sustainable food cost against the local basket, and break-even against that block's real rent. Diego F. Parra presses an uncomfortable point: filtering location before disbursing cuts beneficiaries on paper and saves jobs on the street. Siting is costing, and the 2026 trend that protects the most margin puts the polygon's labor cost inside the territorial model instead of computing it afterward. In the United States, median wages reach 16.23 USD per hour for waiters and 16.12 USD for bartenders, according to the U.S. Bureau of Labor Statistics with May 2024 data, while the federal direct tipped wage has sat at 2.13 USD since 1991, according to the Department of Labor. That distance between legal and real varies by neighborhood and decides whether your break-even exists.

Payroll and rent belong on the map, not in the spreadsheet at the end

A polygon whose market wages run 18% above the metro average demands a different ticket, a different menu size, and a different staffing plan. Load the local hourly cost onto the map before you sign. And keep the house rule: food cost up to 32% per dish is a ceiling, never a target, and payroll gets paid at break-even, not on the plate. Adopt three things now and leave the rest under observation. First, the origin-destination matrix by time band, where 70% of a study's value lives and which already costs about the same as two slow service nights. Second, the polygon's base closure rate with a declared window. Third, local labor cost inside the model. Leave under watch the cannibalization modeling between your own units, which only matters above six locations, and sentiment maps built from reviews, whose correlation with actual revenue is still weak.

2026 horizon: what to adopt now and what to watch from a distance

Demographic profiling helps you read a neighborhood: 47% of U.S. restaurants are at least 50% women-owned versus 43% of the private sector, according to the U.S. Census Bureau cited by the National Restaurant Association in 2022. Running fewer than four units, do not buy a platform. Buy studies per site and keep the raw data. The single AI-generated suitability score is the trend I would ignore this year, and I say it having argued for years that one number summarizes better than a report. I was wrong there. An 82 out of 100 tells you nothing about which variable carried the weight, cannot be audited, and will not survive an investment committee or a bank studying your credit file. Worse still: those models train on corpora from mature markets where informality sits in single digits, then get applied to Latin American cities where youth informality reaches 62.4%, according to the ILO and ECLAC in 2024, so observable demand and real demand barely resemble each other.

The overrated trend: the AI-generated suitability score

If the provider will not hand over the variables with their weights and sources, ask for the raw data and build the criterion yourself. Watch Spanish hospitality as a saturation thermometer: 1.32 million workers and 4.8% of GDP, according to Hostelería de España in 2024. GIS does not predict success, it bounds failure. That distinction looks semantic and is not: a territorial model never tells you that you will bill 40,000 USD a month, it tells you that in this polygon 78% of food venues opened since 2021 failed to reach month 24. With that figure on the table the conversation with the landlord changes entirely, because you are no longer negotiating a rent, you are negotiating a risk premium. The heaviest variable is almost never flow, it is the ORIGIN of that flow. Two blocks with identical pedestrian counts behave in opposite ways when one receives people who work there and the other receives people passing through toward somewhere else; the first sustains corporate lunch at low ticket and high frequency, the second sustains nothing.

Where the real difference sits?

For years I looked at volume instead of vector, and that mistake cost real money. There is a genuine tension between precision and programme speed.

A flawless territorial study takes six weeks and by then the unit is leased; a three-layer filter takes four days and discards 60% of the bad cases. We resolved that tension with two gates: a fast filter across the whole pipeline, a deep study only for sites that pass and will receive more than 25,000 USD of capital. Geographic data now enters the credit file, and that shifts the cost of money for the gastronomic MSME. A restaurant arriving with a baseline per polygon, a sales curve by time band and a documented rent ceiling stops being an opaque risk; commercial banks with MSME portfolios can price that, and pricing is the opposite of declining. Whenever menus come up, the house rule holds: keep the PHYSICAL menu always and add the QR menu as a complement.

Where the real difference sits — in practice?

The printed menu governs service pace, menu narrative and suggestive selling; the QR adds delivery, accessibility, price updates and analytics on what guests browse in each area.

QR data crossed with the polygon feeds the GIS; the physical menu is what holds the ticket.