Short answer, as of September 2026: track AI visibility per location, not per brand — build a prompt set that carries the city or neighbourhood inside the prompt, sample a tiered subset of locations rather than all of them, and report a rollup that always names the bottom five. A brand-level number is an average, and averages are precisely where a location that never gets named goes unnoticed.
The failure is easy to picture. A 40-location brand sees "cited in 55% of tracked prompts" and calls it healthy. Underneath, twelve flagship locations in big metros are named almost every time, and eighteen suburban locations are named essentially never. The average is real. It is also useless, because nobody operates an average — somebody operates a store in a suburb no assistant has ever heard of.
Why brand-level numbers break down
When someone asks an assistant for a recommendation, the question is nearly always local in shape. "Who should I call near me", "best one in [city]", "is there one open on a Sunday in [neighbourhood]". The engine resolves that to a place before it resolves it to a brand.
That means the unit of competition is the location, not the company. Your brand can be well known nationally and still be absent from the answer in a city where your entity data is thin, your reviews are sparse, or a strong independent competitor dominates the local sources. National authority does not automatically flow down to a branch.
This is the same structural argument behind measuring local rankings on a grid rather than as one number, which the multi-location SEO playbook covers for traditional search. AI answers make it sharper, because there is no ranked list to scan — there is one answer naming two or three businesses, and you are in it or you are not.
Put the place inside the prompt
The mechanical change is small and it is the whole game: the location has to be in the prompt text, because that is what the engine sees.
Tracking "best HVAC company" and hoping the engine infers a city produces noise. Tracking "best HVAC company in Plano" and "emergency AC repair Plano Saturday" produces a result you can act on. Run the same prompt shapes across every tracked location so the answers are comparable — the point is a matrix, not a pile.
A workable per-location set is four or five prompt shapes reused everywhere:
- a plain recommendation prompt ("best X in [place]")
- an urgency prompt ("emergency X in [place] tonight")
- a cost prompt ("how much does X cost in [place]")
- a qualifier prompt ("X in [place] that does [specific service]")
- a comparison prompt ("X in [place] vs [known local competitor]")
Sourcing those shapes from real customer language rather than keyword tools matters more than the count; the method is in how to build a prompt set worth tracking. Reuse the shapes, swap the place, and you have a grid you can read down a column or across a row.
The arithmetic nobody does before buying
Per-location tracking multiplies fast, and this is where most programmes quietly break. The volume is:
locations x prompt shapes x engines x repetitions per cycle
Repetitions are not optional. Assistant answers vary between runs, so a single ask tells you what happened once, not what is typical. Two or three repetitions per prompt per cycle is the difference between a measurement and an anecdote.
Forty locations, five prompt shapes, four engines and three repetitions is 2,400 checks per cycle. Run it monthly and that is your monthly budget; run it weekly and it is four times that. For reference, the tracker on the Agency plan includes 2,500 prompt-checks a month and Pro includes 10,000 — so the design above fits a monthly cadence on Agency and a weekly one on Pro. Do this sum before you choose a plan or a vendor, because the number decides your cadence.
Tier the locations instead of sampling all of them
The instinct is to track everything. It is usually the wrong call, because it forces a cadence so slow you cannot tell a real change from ordinary variance.
Tiering works better. Track a smaller set more often, and rotate the rest.
| Tier | Which locations | Prompt shapes | Cadence | Why |
|---|---|---|---|---|
| Flagship | Top revenue, or newly opened | All 5 | Weekly | Change matters fast; new sites need an entity baseline |
| Core | Established, steady performers | 3 | Monthly | Enough to catch drift without burning budget |
| Long tail | Small, stable, low-margin markets | 2 | Quarterly rotation | Confirms presence; full coverage over a year |
| Watchlist | Any location that lost a citation | All 5 | Weekly until recovered | Temporary promotion, not a permanent tier |
The watchlist tier is the one teams forget to build, and it is the one that earns its keep. A location that drops out of answers gets promoted to weekly until it is back, then returns to its normal tier. That gives you resolution exactly where something is wrong, and costs nothing where nothing is happening.
Read the divergence, not the average
Once the grid exists, the useful analysis is comparative. Because every location runs the same prompt shapes, the differences isolate causes a single-location view cannot.
If one location is weak and its neighbours are fine, the problem is local: thin or inconsistent entity data, few reviews, a Business Profile that is incomplete or mis-categorised, or a genuinely dominant local competitor. Start with entity data — the mechanics are in entity consistency and the knowledge graph — and with the profile itself, since posts and review responses are the cheapest lever and the most automatable, as covered in Business Profile automation.
If every location is weak on the same prompt shape — say the cost prompt — the problem is brand-level content, not local. No amount of Business Profile work fixes "how much does X cost in [place]" if you publish no pricing anywhere. That is a content decision, and the tradeoffs are laid out in publishing prices and cost transparency.
If a whole region moves together, look for something regional: a directory or news source with strong local coverage, a regional competitor with better third-party presence, or a franchise partner running their own sites in a way that fragments your entity.
That three-way split — one location, one prompt shape, one region — resolves most of what a grid shows you, and it is only visible because you kept the prompts constant.
Report the rollup, but never only the rollup
Executives want one number and they should get one. The honest way to give it to them is to pair it with the spread.
A one-page rollup that survives scrutiny has four parts: the brand-level citation rate, the number of locations cited in zero prompts, the five weakest named locations, and what changed since last cycle. The zero-citation count is the most important line on the page and the one a brand-level average erases completely.
Set the baseline before you start optimising, or you will have no way to prove anything moved; the method is in benchmarks and share of voice. And resist converting citation rates into a rank-like score — it invites comparisons the data cannot support, a trap explained in rank tracker versus AI visibility tracker.
Where this stops being a spreadsheet
A grid of four locations run by hand once a quarter is fine in a spreadsheet. Forty locations across four engines with repetitions is not — not because the asking is hard, but because the storing is. Without dated, retained answers you cannot see decay, you cannot prove improvement, and you cannot tell variance from a real drop.
That is the point at which per-location tracking becomes software: scheduled runs, one prompt set templated across locations, retained answer history, and a rollup that names outliers instead of averaging them away. You can see how the tracker handles the location dimension on the features page, and the plan limits map directly onto the arithmetic above — decide your cadence first, then pick the tier that covers it.
Start with one honest count this week: how many of your locations have never been named in an AI answer? If you cannot answer that, the brand-level number you are reporting is hiding it.
Frequently asked questions
How do I track AI visibility for multiple locations? Build four or five prompt shapes that contain the place name inside the prompt text, then run the same shapes across every tracked location so results are comparable. Multiply locations by prompt shapes by engines by repetitions to size the sampling budget, and tier locations by importance rather than tracking all of them at the same cadence.
Can I just track the brand name instead of each location? Not if you operate in more than one market. Assistant recommendations resolve to a place before they resolve to a brand, so national recognition does not guarantee a branch is named in its own city. A brand-level average routinely hides locations cited in zero prompts.
How many prompts do I need per location? Four or five reusable prompt shapes per location is usually enough: a plain recommendation, an urgency variant, a cost question, a service qualifier and a competitor comparison. Depth matters less than keeping the shapes identical across locations, because that is what makes the grid comparable.
How often should I run per-location checks? Tier it. Flagship and newly opened locations weekly, established locations monthly, small stable markets on a quarterly rotation, and any location that lost a citation promoted to weekly until it recovers. A uniform cadence across every location either costs too much or moves too slowly to be useful.
Why is one of my locations invisible when the others are fine? Almost always something local: inconsistent or thin entity data, few reviews, an incomplete or mis-categorised Business Profile, or a genuinely dominant local competitor. Check entity data and the profile first, since those are the cheapest to correct and the fastest to propagate.
What should a multi-location AI visibility report contain? Four things: the brand-level citation rate, the count of locations cited in zero prompts, the five weakest named locations, and what changed since the last cycle. The zero-citation count matters most, because it is the figure a brand-level average erases.
Do I need to track every AI engine for every location? No, and it is usually the first place to economise. Track all engines for flagship locations and a reduced set for the long tail. Engine coverage differs more by site access and source availability than by location, so the long tail rarely tells you something new about which engine to prioritise.
Does this work for franchises where each location has its own site? Yes, and it matters more there, because separate franchisee sites frequently fragment the brand entity. Track per location as normal, but watch for whole regions moving together — that pattern usually points at entity fragmentation or an inconsistent local web presence rather than anything a single location did.