A 45-person mapmaker reads Japan in 125-meter squares

Walk a few minutes from the Yaesu side of Tokyo Station. In the morning, bodies pour toward the gates. At lunch, queues form outside restaurants. As office lights begin to dim, signs come on in another alley. Cross a single road and a shopping complex that fills on weekends may sit beside an office block that swells only from Monday through Friday.

An address book calls all of it “the station district.” A census can separate residents from workers. A person opening a shop needs to know more. Will anybody buy coffee at seven? Does lunch demand survive at two? Do families arrive on Saturday? Will residents be home for an evening delivery? A place is not merely a coordinate. It is a combination of time and use.

On Aug. 5, Giken Shoji International Co. began providing c-japan® Dynamics, a packaged local-area dataset intended to express those differences. Using smartphone-location data from SoftBank subsidiary Agoop, it divides Japan into cells roughly 125 meters square and analyzes the average number of people present in each of 24 hourly bands on weekdays and holidays. The company assigns cells to six broad types—entertainment district, office, holiday leisure, transport and logistics, actively occupied residential area, and commuter suburb—and 23 finer categories.

Ino measured the shape of the country

Japan’s history of reading itself through data began by establishing the shape of the land. From 1800 to 1816, surveyor Ino Tadataka traveled coastlines and roads, recording bearings and distances. In 1821, three years after his death, the Tokugawa shogunate’s astronomical office completed the Dai Nihon Enkai Yochi Zenzu. Its large-scale edition consisted of 214 sheets at 1:36,000, each roughly the size of a tatami mat.

The Ino maps answered where the shoreline lay and how far one post town stood from another. They did not show how many people lived there, where they went to work by day or where they returned at night. The map was the skeleton of place; daily life was blood moving across it.

The Meiji government counted population through the family-registration system created in 1871, but registered domicile and actual residence diverged. Missing notices of moves and deaths inflated the records. In 1920, Japan conducted its first direct national census and counted 55,963,053 people. The state had taken a standardized snapshot of who was present where.

A city has one population by night and another by day

As modern cities grew, resident population became insufficient. The 1930 census asked place of work and calculated what Japan calls daytime population. The 1960 census expanded questions about commuting, school and workplace to capture daily movement produced by industrialization and metropolitan concentration.

The distinction was powerful. A central business district such as Chiyoda has far more workers by day than residents at night. A commuter suburb has the reverse. Railways, retailers and local governments could think about transport, stores and public facilities using the people who gathered, not only those who slept there.

Day and night are still just two still images. The morning transfer crowd, lunchtime diners, afternoon shoppers and late-shift workers collapse into one “daytime” category. Dynamics attempts to insert 24 frames between each of those old pictures, one set for workdays and one for days off.

Japan removed administrative borders and laid down a grid

Municipalities and neighborhoods are useful names but unstable units of analysis. They merge, split and vary enormously in area and shape. A mountain village and a city block cannot be compared as if each were one equivalent “area.”

Japan’s Statistics Bureau therefore organized data into fixed cells based on latitude and longitude. The government first produced experimental regional-grid statistics for parts of the Tokyo region in 1969, and the approach entered the census era in 1970. Cells of roughly equal shape and size—one kilometer, 500 meters and later finer subdivisions—make comparison possible even when administrative borders change. The method found uses in planning, disaster prevention, pollution policy, academic research and retail catchments.

A 125-meter cell is approximately a one-kilometer standard grid divided eight ways in each direction. It can distinguish the east and west sides of a station, two sides of a major road, a mall and the homes behind it. Yet a finer grid does not necessarily mean a more accurate observation. That distinction is the key to reading this product responsibly.

A small mapping company’s 30-year view

Giken Shoji is a private company founded in 1976, with headquarters in Nagoya and Tokyo. It has capital of ¥231.125 million and only 45 employees as of April 2026, including ten certified statisticians and five data analysts. It is neither a telecommunications carrier nor an advertising conglomerate. Its specialty is placing population, households, consumption and stores on maps.

Its MarketAnalyzer geographic-information system has a history of about 30 years and, according to Giken Shoji, has been adopted by more than 2,000 companies, chiefly larger businesses. It overlays census and store information, draws catchments by radius or travel time, and helps compare competition with market volume. Work once done by drawing circles on paper maps and assembling statistical tables became repeatable on a screen.

The c-japan series packages more of that judgment. “Home” classifies residential areas through household structure, income and age. “Daytime” classifies business and visitor concentrations. Dynamics adds when people gather. The company says it statistically organizes more than 100,000 data fields into small-area portraits.

The smartphone became a moving census—of a sample

A census is normally taken once in five years and captures a defined moment precisely. A smartphone, after a user grants location permission, becomes a moving observation point. Agoop embeds collection technology in its own and partner apps. It says it takes data only from users who agree to the terms and permit transmission, and that the sample is multi-carrier rather than limited to SoftBank subscribers. It removes direct identifying information and supplies statistical aggregates.

Dynamics uses those records to build average presence by 125-meter cell. Compare each hourly curve and it becomes possible to distinguish a station that spikes in the morning, an office district that peaks at lunch, a commercial destination that fills on holiday afternoons and a suburb whose population returns at night.

The important move is from quantity to shape. Two places may each contain an estimated 1,000 people at noon, yet one may hold commuters in transit while the other attracts families who stay for hours. The product does not directly observe purpose. It assigns use from the temporal pattern.

“AI-ready” translates a neighborhood from words into columns

Before footfall can enter a demand model, an analyst normally handles missing values and outliers, normalizes hours, separates weekdays from holidays, adjusts for density and creates useful features. This preprocessing consumes a large share of practical data-science work.

Giken Shoji says it analyzes the hourly curves and provides not only segment labels but continuous principal-component scores along such axes as work versus residence, weekday versus holiday, entertainment intensity and concentration at particular times. A customer can join those columns to store sales, rent, delivery results or advertising response and use them in regression, classification or clustering.

AI-ready is therefore not a claim that a machine has learned the essence of Ginza or a rural town. It is a practical promise that repeated cleaning and compression have been turned into reusable variables.

125-meter detail is not 125-meter truth

The finer a map looks, the more confidence it inspires. Location signals still have error. GNSS performs poorly indoors and underground. Wi-Fi, cell towers and other signals can help, but urban canyons produce drift. A person on a railway platform may appear in the neighboring shopping cell. A car waiting at a light may be interpreted as a stationary person.

More importantly, the observed population is not everyone. It consists of people who use a participating app, grant permission, carry the phone and produce usable records. Children, older adults, feature-phone users, privacy-conscious users and people running battery-saving settings can be absent at different rates. If an app’s popularity changes by region or month, the sample changes even when the street does not.

Agoop expands app-user observations to the scale of Japan’s total population. Expansion is necessary, but a multiplier alone cannot erase selection bias. The land ministry’s official guidance warns that commercial movement data may be estimates rather than direct counts. Buyers should ask for sample sizes and methods, inspect data from the required locations and periods, and compare them with pedestrian counts and established statistics.

A classification is a useful lie

Six types and 23 classes turn a mass of hourly data into language a meeting can use: office district, so pursue weekday lunch; holiday leisure, so think about families. Labels reduce explanation costs and allow cities to be compared.

They also draw hard lines across a continuous world. A cell may contain offices and nightlife, a university and homes, or a logistics center that becomes an event venue on weekends. The position of the grid origin or a single road can split one functional area into different classes. Statisticians call the broader difficulty the modifiable areal unit problem: results can change when the size and boundary of the unit change.

An area label does not describe an individual. A person inside a “commuter-suburb” cell need not be an office worker; a visitor to an “entertainment” cell need not be drinking or shopping. Inferring individuals from an area average is the ecological fallacy. When used in advertising, convenient categories can become stereotyping or surveillance-like treatment.

What footfall cannot tell a shopkeeper

Whether an area is strongest in the morning or at night matters enormously to a coffee shop, nursery, pharmacy or bar. But presence alone does not create revenue. Can people cross the road? Is the sign visible from the gate? How powerful are competitors? What is the rent? Are passersby rushing? Does the route move under rain?

The responsible use is to treat Dynamics as one explanatory variable, not the answer. Join it to rent, floor area, competitors, direction of flow, counts at the door, sales and average ticket. Train on existing locations and test on unseen ones. Investigate where predictions fail and whether they reproduce in another place and season.

Averages erase festivals—and disasters

An average weekday and holiday can be useful for normal operation. A city’s most consequential days are abnormal: a fireworks festival, concert, typhoon, earthquake, suspended railway, heatwave or tourist season. Depending on the method, averages can wash those events out or allow them to distort the class.

Update frequency is equally important. Move a station gate, open an office tower, close a factory or relocate a university and neighborhood use can change within months. A long observation period responds slowly; a short one is noisy. The release does not state the date range used to produce the average or the planned refresh schedule.

“Dynamics” sounds live. What is sold appears to be a static table of classifications and scores derived from movement. Time-aware data should not be confused with present-time data.

Consent at collection and responsibility after aggregation

According to the companies, Agoop collects only from users who accept participating-app terms and permit location transmission, then strips identifying information and statistically processes the records. That is an important foundation. It is separate from whether a person understood that her movement could help classify retail catchments, advertisements or delivery schedules.

Aggregation is not the end of privacy design. In a sparsely populated rural cell or at three in the morning, a small group’s behavior may remain conspicuous. Minimum-count thresholds, suppression, temporal rounding or merging with adjacent cells may be needed. The release says records are anonymized and aggregated but does not disclose a minimum cell count or use of formal methods such as differential privacy.

Commercial and public uses also carry different consequences. An error in a convenience-store hypothesis mainly risks investment. An error in the placement of transport, medical care or evacuation services can undercount older people, low-income residents or children who are less visible in an app sample. Government use requires stronger transparency, audit, resident explanation and comparison with official statistics.

AI can narrate correlation; it does not know the “why”

An hourly curve shows what appears to happen. A weekday noon population rises; people stay longer on holiday afternoons; population returns after dark. Location alone does not identify the cause. Cheap restaurants, a school timetable, factory shifts, a hospital, construction and tour buses can generate similar patterns.

Two days before the Dynamics launch, Giken Shoji announced an update to a separate MarketAnalyzer “catchment report AI.” It combines statistics with general impressions held by a language model to produce a neighborhood story. The company itself notes that generic generative AI can produce one-sided summaries or hallucinations, while saying its design controls but deliberately permits some stereotypes to add context.

That is a risky line. Add an inherited image—“young people’s district,” “wealthy suburb,” “foreign-tourist town”—to a correlation and the AI may transform it into a causal story. Treat every machine-written explanation as a hypothesis. Verify it by walking the street, interviewing shops, consulting local documents and checking transport schedules. Persuasive prose is not evidence of a correct cause.

What the little company can do—and must explain

A private, 45-person company is turning the rhythms of a country into AI features. Smallness is not a reason to distrust it. Thirty years of catchment analysis may teach the firm practical frictions a giant platform overlooks: which statistics join cleanly, which unit a manager can use, which label survives a meeting. Expertise does not scale directly with headcount.

Yet a provider claiming nationwide consistency and the removal of sparse-area distortions owes customers a method sheet. Observation windows, sample distribution, clustering procedure, outlier rules, low-density treatment, validation and version history should form a geographic data card. “Proprietary” and “patent pending” cannot substitute for testing.

The one-month, production-equivalent free trial creates an opportunity. Hide a set of existing stores, predict them, compare the map with what local managers know and investigate the failures. A small company earns durable trust less by making a large claim than by showing where its map is likely to be wrong.

The map is never the neighborhood

Ino’s maps captured the coastline with astonishing fidelity. Still, the paper carries neither the smell of a fishing port at dawn nor the noise of a post town. The census counted residents and commuters, but did not divide a city’s day into 24 frames. The smartphone filled some of that space with moving dots.

The compelling idea inside c-japan Dynamics is to treat a place not as a fixed address but as a wave through time: morning station, midday office, holiday park, nighttime suburb. It packages that rhythm for a spreadsheet or model.

Humans still draw the grid, the company still names the classes, and estimation still stands in for unobserved people. As the cells become finer, those choices become easier to overlook. The best location data should not make its user feel that the neighborhood has been “understood.” It should tell her more intelligently where to walk next, what to count and whom to ask.