AI Search Visibility by Country, State, and City with Personas and Languages

A national prompt is a different question from the same words asked as a buyer in Austin. Generative engines take location and language as context. They do not return a Google map pack. If you only track one English prompt from one country, you will average away the city where you sell. The reason is mechanical. The model reads the location signal you give it, weights nearby and regionally relevant sources higher, and writes an answer that fits a buyer in that place. Run the same words without a location and it writes a generic answer for no one in particular. The national row is an average of cities you do and do not compete in, and the city you care about can be invisible inside it.

Promptwatch is the tool that attaches country, state, city, language, and personas to prompt checks. Country tracking is on all paid plans. State and city tracking are on Professional ($245/mo) and above, and on the agency plans (Kick-off $199/mo, Growth $399/mo, Scale $799/mo). Essential is $95/mo (7-day trial) with country-level tracking. Explore is free (ChatGPT, 10 prompts) and is not a geo desk. Site: promptwatch.com. The plan split is deliberate. Country is cheap because it is one signal. State and city are more expensive because each location copy is a separate run that consumes response quota. The agency tiers carry state and city because agencies run geo for many clients at once and need the targeting on every project.

Duplicate the prompt per city

Do not reuse one national prompt and "think about" cities in the analysis. Duplicate the prompt, set the city, and read the answers as separate rows. A plumber in Denver and a plumber in Phoenix can get different vendor lists from the same wording. The model is not BrightLocal. There is no pack, no pin, no review-star overlay. What you get instead is a prose answer that names the businesses the model associates with that city for that query. If you sell in Denver, the Denver row is the one that tells you whether you are in the answer. The national row tells you nothing you can act on, because you cannot optimize for "the United States."

This is generative GEO location context: what the model says when it believes the user is in that place. It is not BrightLocal map-pack tracking. Keep map-pack tools for Maps. Keep Promptwatch for the prose. The two systems measure different surfaces. A map pack is a ranked list of pins with reviews and distance. A generated answer is a paragraph that may or may not name you, and may footnote you even when it does not name you in the prose. Mixing the two in one report produces a number that means neither.

Country-level is enough when you sell one market. Add state and city when the answer names local competitors, regulations, or retailers that change across a border or a metro. The test is whether the answer changes. Run the prompt once with the country set, then again with the city set. If the vendor list shifts, you need the city row. If it does not, the country row is enough and you save the quota.

Personas and languages

Personas prompt like a customer segment: a first-time buyer, an IT admin, a clinic manager. The same product question asked as those people produces different citations. Build two or three personas you can defend. A dozen fictional characters will burn response quota. The point of a persona is not to be exhaustive. It is to cover the buyer types that convert, so the citations you track are the citations that lead to pipeline. A first-time buyer asks "what is the best X for someone new." An IT admin asks "how does X integrate with our stack." Those are different prompts and they pull different sources. Three personas is usually enough to cover the spread without spending the response budget on noise.

Multi-language monitoring is a separate axis. Promptwatch can run conversations in languages including English, Dutch, German, French, Portuguese, Japanese, and others. Translate the prompt the way a native buyer would type it. A wooden English calque is a different query. The model reads the calque as an English query wearing a foreign language, and it answers from English-leaning sources. A native phrasing pulls native sources. If you sell in Germany, the German row only counts if the prompt is German the way a German buyer types it, not German the way a translation tool outputs it.

AI Overviews can be tracked per region. AI Mode is available across supported countries. Still check ChatGPT, Gemini, and Claude in the same locations. Overviews are not a proxy for chat. A buyer who uses AI Overviews and a buyer who uses ChatGPT are on different surfaces, and a citation in Overviews does not mean a citation in ChatGPT. Run the same location cluster across the engines you care about and read the rows side by side.

Quota math

Essential includes 6,000 responses across models and country settings. It does not include state or city targeting. City copies multiply fast once you move to Professional, which includes 18,000 responses and 150 prompts. Business is $579/mo with 350 prompts and 42,000 responses. Agency plans have unlimited prompts and include state and city targeting. The quota math matters because every city copy is a separate run. Eight prompts across five cities is forty runs per check, and a daily cadence turns that into forty runs a day. Essential caps at 6,000 responses and country only, so it is not the plan for a city grid. Professional at 18,000 responses is where a real city cluster fits. Business is for teams running many prompts across many locations. Agency plans exist because agencies cannot cap prompts per client without starving the work.

Otterly.AI from $29/mo lists 65+ countries on mention checks, with up to a 7-day lag. Peec AI from $95/mo is three models. Scrunch AI from $250/mo annual is weekly. Method: how we rank. Tools list. The lag on Otterly is the tradeoff for the cheap country row. Peec is three models, which is enough if those three are the engines you report on. Scrunch weekly is a reporting cadence, not a daily monitor.

FAQ

Is city tracking the same as rank tracking in Maps?

No. You are measuring generated answers with a location set, not pack position or review velocity. The two answer different questions.

Can Essential see Austin vs Dallas?

Country, yes. State and city start on Professional and the agency tiers, because each city copy is a separate run that consumes response quota.

What to do this week

1. List the three cities that close revenue, not the ones that look good on a slide.
2. Duplicate 8 buyer prompts per city in Promptwatch on Professional or an agency plan.
3. Add one persona that matches a real buyer role, so the citations you track are the ones that convert.
4. Run the same cluster in the language those buyers type, with native phrasing rather than a calque.
5. Report city rows separately. Do not average them into a national score, because the average hides the city you sell in.