How to set up coordinate-level AI search tracking
A step-by-step setup for pinning an AI search monitor to an exact point with Promptwatch's Local GEO/SEO targeting: the pin, the radius, which engines it affects, the prompts, the baseline and how to read the results.
City-level tracking answers "does ChatGPT recommend us in Denver?" That works for a law firm that serves the whole metro. It's too coarse for a coffee shop, a clinic on one corner or a franchise with three stores in the same city, where the useful question is what someone standing near the door gets told.
Coordinate-level tracking asks that question. Instead of telling the AI engine the user is in a city, you give it a latitude, a longitude and a radius. Promptwatch shipped this in October 2026 as Local GEO/SEO targeting. This guide walks through the setup in the order we'd do it, based on Promptwatch's llms-full.txt and its academy pages on creating and managing monitors (all read on 28 October 2026).
Before you start: check the plan
Coordinate targeting is on three plans: Business ($579/mo), Agency Scale ($799/mo) and Enterprise. Essential, Professional, Agency Kick-off and Agency Growth don't have it, although Kick-off and Growth do include city-level tracking. Neither Business nor Agency Scale comes with the 7-day free trial, so you can't test coordinates for free. If you first want to know whether local answers move at all in your category, a city monitor on Kick-off (which does have the trial) is the cheaper experiment.
You also need an editor role on the project. Viewers can't create monitors.
Step 1: decide what the point represents
A monitor holds one location. Before you place the pin, decide what it stands for. For a walk-in business it's usually the storefront. For a business whose customers come from somewhere else, such as a dental practice that draws from a residential area or a service company that drives to people, the more honest point is where the customers are, not where the van is parked.
Put that decision in the monitor name, because you'll forget it. "Store 2, Main St, 100 m" says more on a dashboard six weeks from now than "Local test".
Step 2: place the pin and pick a radius
When you set the monitor's location you can search for a place or drop a pin on the map. Promptwatch then asks for an accuracy radius from a fixed list: 20 m, 50 m, 100 m, 500 m, 1 km or 5 km.
A smaller radius pins the location more tightly. Which one fits depends on how far your customers travel, and we've written a separate guide on choosing a radius. If you're unsure between two, test both in parallel rather than guessing.
Get this right the first time. Like the rest of a monitor's location, the point is fixed once the monitor exists. A different pin or radius means a new monitor. For other big changes, Promptwatch's "Copy to new monitor" button pre-fills the location, models and frequency from the original (not the persona), so the old time series stays clean and the new setup starts its own. Check the location on the copy before you save it.
Step 3: fill in country, city and language anyway
People skip this step, and it matters. Coordinates only change the answers from three engines: ChatGPT, Google AI Overviews and Google AI Mode. Every other model in the monitor uses the monitor's country, state and city. So set the country and, since you're on a plan that has it, the city that contains the pin. State is a US-only field.
Set the tracking language to the language your customers ask in. A persona is optional. If you pick one, a persona stacking switch appears, which wraps each tracked prompt in the persona's own prompt at run time.
Step 4: choose models on purpose
The model list shows ChatGPT, AI Overview, AI Mode, Copilot, Gemini and Perplexity, with more behind "Show all models". You have two sensible options.
The first is to keep the coordinate monitor to ChatGPT, AI Overviews and AI Mode. Every answer in it then reflects the pin, and the numbers mean one thing. Track the other engines in a separate city monitor.
The second is to add everything and always read results per model. That's fewer monitors, but a blended visibility score mixes answers given for a precise point with answers given for a whole city. We go through the mixed setup in how to track Perplexity, Gemini and Copilot for local queries.
Watch the budget either way. Responses are prompts times models times runs, and monitors run daily. Twenty prompts on three models for 30 days is 1,800 responses for one location. Business includes 42,000 responses and Agency Scale 65,000.
Step 5: write the prompt set
After you click Next, Promptwatch suggests up to 10 starting prompts based on the monitor name, location and persona. Treat them as a draft and edit until they read like something a customer would type.
The rules we use:
- Keep place names out of the prompt. The pin sets the location, so "best physio near me" tests the location signal, while "best physio in Clapham" mostly tests your wording.
- Include "near me" variants. They carry no place at all and depend entirely on where the engine thinks the user is. More in how to track "near me" prompts.
- Label prompt types. Organic discovery prompts, brand prompts ("is [name] open on Sunday") and competitor comparisons answer different questions, and Promptwatch lets you tag each one.
- Size it at 15 to 30 prompts. That's Promptwatch's own starting suggestion per monitor or market, and it's about as many as one person will actually read.
Step 6: run a baseline before changing anything
The monitor is active as soon as you create it, and the first responses arrive after the first run. Then leave it alone. AI answers vary from day to day, so one run tells you very little. We give a new local monitor two weeks of daily data before reading a trend, and we don't touch models or prompts in that window. Promptwatch's settings page warns that editing a running monitor can affect its analytics. The warning is fair: the chart ends up mixing two setups.
Step 7: read the results
Each monitor has Overview, Prompts, Responses, Sentiment, Citations, Offsite Mentions and Heatmap tabs. For a local monitor we read them in this order.
Responses first. It shows every answer with your brand highlighted, the raw text and its citations, so you can check whether the engine actually answered for the area or gave a generic list. Then Prompts, to see which questions name you and which never do. Then Citations, to see which pages and domains the local answers draw on: your location page, a directory, a local news site, a Reddit thread.
Visibility Score and Position in Answer are the headline numbers. Read them per model. A strong ChatGPT result and an empty AI Mode result at the same pin is a finding, and an average hides it.
When you have many locations
Clicking 40 pins onto a map gets old fast. The REST API and the MCP monitor tools both accept coordinates (latitude, longitude, locationAccuracy and locationName on a monitor), so a store list in a spreadsheet can become 40 monitors from one script.
Where this leaves you
For pin-level AI tracking with the "why" attached, the whole setup above runs inside Promptwatch on Business or Agency Scale. The coordinate monitor tells you what a person at that spot is told. The same workspace then shows whether AI crawlers fetched the location page (Agent Analytics, included on Business and every agency plan) and whether AI-referred visitors converted. That second half is what turns a local score into a list of things to fix. If you're still choosing a tool, our local AI rank tracker ranking covers the alternatives, and plans are on promptwatch.com.