A national answer can make a network look healthy while one branch is invisible, another is misdescribed, and a third survives only through an old directory page.
The ledger looked tidy until I sorted it by town. Six branches, same brand name, same broad service category, same western-France service area. In the national prompt, the company appeared often enough to calm the room. In the city prompts, the picture cracked. Rennes was named. Vannes appeared only when the prompt included emergency repair. Two smaller towns were missing, and one branch was described as a boiler-only installer even though commercial maintenance was the higher-margin work.
This is a composite scenario, assembled from several regional service networks I have seen around heating, plumbing, electrical work and building maintenance. The rough detail matters: the company’s website had pages for each branch, but one page carried an old phone number in a footer block copied from a previous template. An AI answer did not quote the phone number, but the wrong branch association kept showing up. That is how multi-location visibility often fails. It does not fail in one dramatic way. It frays at the edge of the map.
A national prompt hides the branch problem
A prompt like “best heating maintenance company in western France” feels useful because it sounds like category demand. It may even match the way a management team talks. For measurement, though, it is too wide to diagnose branch visibility. The answer engine has to compress a geography, a service type and a buyer need into a few names. If the business appears, everyone is tempted to read the result as general visibility. I would not trust it for branch decisions.
A multi-location business is made of separate evidence piles. Each branch has its own local pages, review footprint, directory records, service wording, photos, opening hours, partner mentions and customer questions. Some evidence is shared across the network. Some is painfully local. The AI answer may blend the two without warning.
When I test a regional network, I treat the national or regional prompt as only the roof view. It tells me whether the entity has category weight in broad answers. It does not tell me whether a buyer in Saint-Brieuc asking for commercial heating maintenance will see the same business, the right branch, or the right offer. The engine may know the brand and still fail the town.
That is the first uncomfortable lesson. A company can be visible as a network while one branch is absent from the buyer’s answer. The dashboard line looks stable; the branch phone stays quiet.
Multi-location AI visibility is the repeated appearance of the right branch for the right city and service, because buyers ask from local intent, not from the company’s org chart.
That definition is dull on purpose. I want it to force the measurement down to the place where the sale actually begins.
The prompt must carry city, service and buyer intent
For a business with several agencies or branches, I do not start with the brand. I start with the buyer’s small problem. “Who can handle heating maintenance for a small hotel near Quimper?” gives me a different kind of answer from “plumber Quimper” or “best HVAC company Brittany.” The first prompt contains service, customer type, location and use case. The second is closer to a search box. Both can be useful, but they should not sit in the same field of the ledger as if they measure the same thing.
The basic unit is the per-location prompt cell. It combines one city or service area, one service line and one buyer situation. For a heating and plumbing network, I would separate emergency repair, scheduled maintenance, commercial contracts, installation, compliance checks and replacement work. If the company serves both households and businesses, I separate those too. A branch that appears for emergency leaks may disappear for facilities maintenance.
This is where many teams under-sample. They build three prompts for the whole network and call the result a baseline. That gives the pleasant feeling of order. It also misses the point. Each branch needs enough observations to show whether the answer pattern is repeatable. I do not mean hundreds of prompts. I mean a disciplined set that covers the decisions a buyer might actually make.
The city field should also be handled with care. A branch may serve a town without being physically located there. Some AI answers are biased toward the physical office address. Others accept service-area language if the cited source is clear enough. So I test both: “in [city]” and “serving [city]” when the distinction matters. In French, the phrasing changes the feel of the request. “À Rennes” is not the same as “autour de Rennes” or “intervient à Rennes pour…” The ledger should preserve that wording.
I call this the branch-grid test: each branch is tested across city, service, buyer type and language so local visibility can be read without averaging away the weak cells.
French and English should not be folded together
A French SMB with bilingual pages has two evidence paths, even when the business is fully local. English pages may exist for vendor explanations, tourism customers, industrial clients, expat buyers or old SEO experiments. The engine may cite them when the prompt is English, then switch to local directories, review profiles or French landing pages when the prompt is French.
In the composite heating network, the English prompts produced one strange result. The company appeared for “heating maintenance company western France,” but the cited source was not the branch page. It was a directory-style page that had scraped a short description and emphasized emergency plumbing. In French prompts, the engine more often cited the company’s own local pages, but only for cities where the page title and opening paragraph named both the town and the service clearly. Same business. Different language. Different source path.
This is why I keep French and English in separate ledgers before I compare them. A combined score hides too much. If the French prompt finds the right branch but the English prompt cites a weak directory, the fix is not simply “write more content.” The fix may be to strengthen the English evidence for the right service or to reduce the ambiguity that lets a directory define the branch.
There is another trap. Some teams run English prompts because they feel cleaner or because the interface is set that way. Then they make French business decisions from English evidence. For a local French service business, that is usually backwards. French buyer prompts should carry more weight unless the company genuinely sells to English-speaking buyers.
I still run English tests when the site has English pages or the market has bilingual customers. I just do not let the English result vote twice. It gets its own column, its own cited source and its own description accuracy score.
The local failure has several shapes
A branch can fail visibility in more than one way. It can be absent. It can be named but tied to the wrong city. It can be named for the wrong service. It can be cited through a weak source. It can be described with an old business fact. Those failures should not be collapsed into a single red mark.
In my notes, I use a small classification called local entity drift. It has four common forms. The first is branch absence: the network appears elsewhere, but this town does not show the business at all. The second is branch substitution: the answer names the brand but points the buyer toward another nearby branch. The third is service narrowing: the engine describes the branch through one service, usually the most repeated one, and misses the margin work. The fourth is source misassignment: the answer cites a page that belongs to the network, a directory or a partner, but not the best local proof for that branch.
I am cautious with neat classifications because real ledgers are messy. A single prompt can show two forms at once. In one run, the answer may name the company under the right town but cite a national page. In another, it may cite the right branch page but describe the offer too narrowly. The value of the classification is not beauty. It keeps the correction from becoming a vague content task.
Absence asks one question: does enough local evidence exist for this branch and service? Substitution asks another: are branch pages, addresses and service areas clearly separated? Service narrowing points toward offer wording and proof. Source misassignment asks why the better page is not being selected.
If you skip this separation, every problem becomes “improve the website.” That is too blunt. The ledger should tell you which page, source, phrase or branch relationship deserves attention.
Competitors belong in every city row
A branch is not visible in empty space. It is visible against the other names the engine chooses. That sounds obvious until the first audit arrives with only the company’s own presence score. A branch that appears once in five runs may look poor. But if all competitors are unstable too, the market may be thinly understood by the engine. A branch that appears three times in five runs may look healthy. But if two competitors appear in every run and occupy the first cited source, the branch is not leading the answer.
For multi-location work, I keep competitors inside each city row. The competitor set can change by city. A national chain may matter in Paris and mean little in a smaller western town. A local specialist may appear only around one branch. That local competitor is not noise. It is exactly the pressure the branch faces in the answer.
The roughest part is naming competitors without turning the ledger into gossip. I prefer to record what the engine repeats: names, answer order, cited source and description quality. Then I look at the cited pages. Does the competitor have a clearer local service page? Are they cited through a trade directory? Are their reviews more locally tied to the city? Does their page contain a simple description that the engine can lift without confusion?
This is where the work becomes practical. If a rival is winning the answer through a clearer city-service page, the correction path is visible. If they are winning through a strong third-party source, the company’s own page may not be enough. If nobody is stable, the first goal may be to make the branch unambiguous rather than to chase a full answer share.
The competitor row also protects the business from vanity. A branch mentioned in an answer with four stronger competitors has a different problem from a branch missing from an answer where all named businesses are weakly cited.
The correction follows the branch, not the brand
After the first per-location run, the temptation is to make a grand brand correction. Rewrite the main service page. Add a new paragraph to the home page. Push a broad “we serve western France” line. Sometimes that helps. Often it misses the local wound.
For the composite heating network, the useful corrections were smaller. One branch page needed its commercial maintenance work named earlier, with a concrete customer type and service area. Another needed internal links from the regional service page because it was orphaned in the site structure. A third needed the French page and English page to stop using slightly different service labels. The old footer phone number had to be removed, of course, but the more important fix was clearer branch-source alignment.
I retest the same prompt cells after a correction. I do not change the prompt to make the answer nicer. That is a quiet form of cheating. If the buyer wording was valid before the edit, it remains valid after the edit. The point is to see whether the answer pattern changes, not whether I can write a prompt that flatters the page.
Monthly monitoring does not have to repeat every exploratory prompt. Once the baseline has exposed the weak cells, the recurring set can be smaller. I keep the branch-grid core: priority cities, priority services, French and English where relevant, and the main competitors. The ledger should be light enough to run again and detailed enough to catch local drift.
The branch is the unit of repair. The brand may be the unit of reputation, but the branch is where the buyer’s question lands.
The Measurement Note — Signal: a branch appears for its own city, service and buyer situation, not only in a national brand answer. Distortion: averaging six locations into one visibility score. Ledger: record city, service, language, branch named, answer position, cited source, competitor names and description error. Next Test: choose three branches, write five buyer prompts per branch in French, then rerun the weakest two in English before editing any page.