National AI Visibility Cannot Answer a Regional Sales Question

A national AI visibility score cannot tell you what buyers in Austin, Chicago, London, or Singapore are seeing. If a regional campaign depends on the answer, build a location-sampling brief that separates the words in the prompt from the place of the test, the geography of the audience, and the engine configuration.

Until those fields are explicit, the honest regional result is unmeasured.

Regional AI visibility is a sample-design problem

The mistake starts when a team treats geography as one field.

It is at least four different fields:

Field	The question it answers	Example
Prompt geography	Did the question name a place?	“Best payroll platform for Texas manufacturers”
Test locale	Where and under what locale was the answer observed?	United States, English, Central Time
Audience geography	Which market contains the buyers whose behavior matters?	Texas-based operations leaders
Business coverage	Where can the company actually sell, deliver, or support?	United States and Canada

A prompt can name Dallas while the test runs from another locale. A test can run in Texas while the prompt never names a place. A brand can appear nationally while lacking evidence for the local use case. Those observations are not interchangeable.

NIST’s AI Risk Management Framework recommends documenting test sets, metrics, tools, and conditions and evaluating systems under conditions similar to their expected use. For a regional sales question, geography is part of the expected use. It cannot be added after the result arrives.

Write the sales decision before choosing locations

Do not begin with a map. Begin with the decision.

A weak objective is:

Measure our AI visibility by city.

A useful objective is:

Decide whether our Dallas campaign needs additional local evidence before the October enterprise launch.

The second version identifies a market, a deadline, an audience, and an action. It also limits what the sample needs to answer.

Use this decision block:

Decision field	What to write
Commercial decision	The budget, campaign, market, or sales motion that could change
Target audience	The buyer role and company context
Geography	City, state, region, or country relevant to the decision
Decision date	When the evidence must be useful
Minimum evidence	What must be observed before the team acts
No-result rule	What will remain unmeasured if the sample is too thin

This keeps measurement tied to pipeline. It also prevents a national headline from becoming a local promise simply because the dashboard offers a location label.

Separate a location-worded prompt from a location-controlled test

A location word inside a prompt is not proof that the system answered from that location.

OpenAI documents that ChatGPT search may use approximate location inferred from an IP address and can use more specific device location when a user enables it. Google Search documents that it uses a shared or estimated location to make some results locally relevant. Those product documents do not establish identical behavior across every AI answer surface. They establish the measurement point: prompt wording alone does not reveal every location input that may affect a result.

Consider two hypothetical prompts:

1. “Which cybersecurity consultancies serve healthcare companies in Chicago?”
2. “Which cybersecurity consultancies should a 500-person healthcare company consider?”

The first prompt contains a city. The second does not. Either prompt could be observed under different test locales, languages, account states, dates, or engine configurations.

Your worksheet should preserve both the prompt and the conditions separately:

Observation field	Required record
Exact prompt	Full text, without paraphrase
Named place	Any city, state, region, or country stated in the prompt
Engine and surface	The answer engine and product surface used
Date and time	When the answer was observed
Test locale	The locale declared or controlled by the test method
Language	The language used for the query and answer
Account state	Logged in, logged out, or unknown
Full answer	The complete observed response
Cited sources	Every source URL shown with the answer

Do not fill an unknown field with an assumption. Write unknown.

That one word protects the analysis from pretending the test controlled something it did not control.

Build a location matrix, not a national average

A regional AI visibility sample should keep markets separate long enough to see whether they differ.

Start with a matrix like this:

Market	Buyer context	Location-worded prompt	Non-location prompt	Test locale documented?	Result label
Austin	Mid-market software CMO	Yes	Yes	Yes	Observed sample
Chicago	Enterprise operations leader	Yes	Yes	Yes	Observed sample
London	UK technology buyer	Yes	Yes	No	Incomplete conditions
Singapore	APAC procurement lead	No	No	No	Unmeasured

These rows are hypothetical. They show the accounting rule, not a recommended number of markets.

Do not average the rows until you have a business reason and a defensible weighting method. Austin and London are not interchangeable observations. Neither are a city-level buyer question and a national category prompt.

NIST’s trustworthiness guidance says accuracy measurements should be paired with realistic test sets that represent expected conditions of use. A location matrix makes that representation visible. A national average can hide it.

Label every region as observed, incomplete, or unmeasured

Regional reporting needs three labels.

Observed sample means the team preserved the prompt, answer, date, engine, sources, and declared test conditions for that region.

Incomplete conditions means an answer was captured, but a material field such as locale, account state, or configuration was unknown.

Unmeasured means the team does not have an eligible observation for that region.

Do not turn unmeasured into zero.

Do not turn incomplete conditions into a clean comparison.

Do not turn one observed sample into a market estimate.

The current Machine Relations Index demonstrates the value of publishing evidence boundaries. Its September 12 release reports 15,154 observed answer runs, 863 monitored prompts, six answer engines, and 79 published source segments. Those published segments are organized by category and question shape. They do not provide a representative regional estimate.

The release manifest records 119 observed dates from May 10 through September 12 and all six engines healthy in the observation window. Engine health and longitudinal coverage are useful facts. They still do not establish geographic balance.

The absence of a regional estimate is not a zero regional result. It is an evidence boundary.

Use the regional sampling worksheet before approving a campaign

Here is the full worksheet I would put in the campaign brief.

1. Decision

• Which regional decision could change?
• Who owns the decision?
• What is the deadline?
• What result would cause action?

2. Markets

• Which markets are commercially relevant?
• Why is each market in the sample?
• Which markets are intentionally excluded?
• Are the markets weighted equally or kept separate?

3. Prompt set

• What buyer question is being tested?
• Which prompts name a location?
• Which prompts keep location out of the wording?
• Are category, comparison, shortlist, and problem-first questions being mixed?

4. Test conditions

• Which answer engine and surface were used?
• What date, language, locale, and account state were recorded?
• Can another operator reproduce the setup?
• Which conditions remain unknown?

5. Evidence

• Was the full answer preserved?
• Were cited sources captured?
• Does each source support the claim attributed to it?
• Is the brand mentioned, recommended, cited, or merely adjacent?

6. Reporting

• Which rows are observed samples?
• Which rows have incomplete conditions?
• Which regions remain unmeasured?
• What claim is the evidence strong enough to support?

This is deliberately stricter than “run the prompt from a few cities.” The location is only one part of the measurement contract.

Keep regional sampling separate from broad product strategy

The existing Adobe Brand Visibility CMO strategy brief addresses the broad platform and budget question. This worksheet solves a narrower operating problem: how to stop a national or general visibility result from being exported into a regional sales claim.

The distinction matters because tool selection and sample design are different decisions.

A capable platform can still be used with a weak regional sample. A manual test can still produce useful evidence if its conditions and limits are preserved. The tool does not rescue the question.

If the regional result reaches a buyer and cannot be supported, use the separate sales correction protocol. Sampling should prevent the overclaim. Correction handles it after it appears.

Machine Relations makes the evidence boundary operational

Machine Relations connects earned authority, entity clarity, distribution, and measurement around what machine readers can actually retrieve and cite.

Regional sampling belongs in that system because a city-level answer can depend on local evidence: publications, associations, customer proof, product pages, and third-party sources that establish relevance in that market. Measurement can show which sources appeared. It cannot create regional authority that the evidence base does not contain.

That is why the report needs more than a visibility score. It needs the prompt, test conditions, cited sources, market, and decision.

Before allocating a regional campaign, run the AuthorityTech AI visibility audit as a baseline, then document the location-sampling fields separately. Use the audit to identify what appeared. Use the worksheet to decide what the observation is allowed to mean.

A national answer can start the investigation.

It cannot finish the regional decision.

FAQ

Can a national AI visibility score predict city-level performance?

No. A national or general score can describe the measurement contract that produced it. It cannot establish what buyers in a specific city see unless the test design includes eligible observations for that geography and preserves the relevant conditions.

Does adding a city name to the prompt create a local AI visibility test?

Not by itself. A city name changes the wording of the question. A location-controlled test also records the engine, surface, date, locale, language, account state, full answer, and cited sources. Prompt geography and test locale are separate fields.

How many locations should a CMO sample?

There is no universal number. Choose locations from the commercial decision, audience footprint, and available measurement controls. Keep thin markets labeled as unmeasured or incomplete rather than inventing representativeness from a convenient count.

What should a regional AI visibility report include?

Include the exact prompt, named geography, engine and surface, date, test locale, language, account state, full answer, cited URLs, buyer context, result label, and the decision the observation may change.

What does unmeasured mean in an AI visibility report?

Unmeasured means the team lacks an eligible observation for that region. It does not mean the brand had zero visibility, zero citations, or zero buyer awareness there. It means the evidence does not support a regional result.