Thirteen keywords in Mumbai, six for the rest of India

India is too big to sweep with every keyword I could think of. So I swept Mumbai with all thirteen first, and let the results decide what the national run could leave out.

Mumbai at full depth

The Mumbai box covered the metropolitan region, from Mumbai and Thane to Navi Mumbai, Vasai-Virar, Kalyan-Dombivli and Panvel. Thirteen keywords ran over it: automation, home automation, smart home, industrial automation, system integrator, building management system, PLC SCADA, robotics automation company, control systems, KNX, lighting automation, gate automation and curtain automation.

The settings were deliberately greedy. Each keyword started from a grid of 42 squares, most about 11 km across, and any square that returned even one result was split into four. The four broadest keywords went down to about 1.4 km, the other nine to about 2.8 km. That came to 7,210 squares and 5,911 distinct places. 24 squares came back with the full 60 results, and every one of them was split again, so no Mumbai square was left at the cap.

Raw counts mislead

My first look was places per keyword, and it pointed the wrong way:

• “Smart home” found 2,170 places, and 1,554 of them turned up in no other search. Big and unique sounds valuable, until you see that this search mostly returns electronics shops.
• “Automation” on its own found 2,174, and 680 of those no other keyword found.
• “PLC SCADA” found 91 places, only 15 of them new.
• “KNX”, a wired home-automation standard, was searched in 50 squares and found one place.

Two thirds of all the places (3,937 of 5,911) were found by a single keyword, so raw volume says little about value.

So keywords were scored on a better question: of the firms the rules actually listed, how many did this keyword find, and how many of those did no other keyword find? The script that answers it reads the saved sweep and the fetched names, and makes no new calls. Six keywords found almost every listed Mumbai firm: automation, control systems, smart home, home automation, industrial automation and system integrator. Those six went national.

“Smart home” stayed despite its noise, because it also finds real firms. The noise was handled later: before any national purchase, places that only “smart home” found outside the widest searches were ranked out.

Replaying split rules offline

The saved sweep doubles as a simulator. Every square’s results are on disk, so a stricter rule (split only at 3, 5, 10 or 20 results; stop at 1.4, 2.8 or 5.5 km) can be replayed by walking the saved squares and skipping the ones that rule would never have searched. Each replay reports how many listed firms it still finds and how many squares it costs, without a single API call.

The national run used:

• 595 start squares per keyword, most about 89 km across;
• a split whenever a big start square returned anything, and below that only at 10 or more results;
• a floor of about 2.8 km for “automation” and 5.5 km for the other five;
• no squares inside the Mumbai box (already done), and none inside 66 circles over cities in neighbouring countries, with a country check on every result as a backstop.

That came to 36,211 squares and 78,624 distinct places.

Where the national run is blind

A coarser rule has a cost, and it can be counted. Nationally, 1,049 squares returned the full 60. All but 65 were split. Those 65 were already at their smallest size, so they may hide more firms. Specialists that only a dropped keyword would find, such as gate or curtain automation firms, may be missing outside Mumbai too. The workbook’s own notes say it plainly: the national list is broad, but less exhaustive than Mumbai.

Rules I’d reuse

1. Calibrate on ground you know before you scale.
2. Judge a keyword by what it finds that nothing else does, among the results you keep.
3. Save raw results, so new rules can be tested without new calls.
4. Count where the cheap run is blind, and say so.