TL;DR
- An AI visibility result belongs to the market and test setup that produced it. One city's answers do not establish coverage in every market you serve.
- An October 2026 ChatGPT API study found that declared location changed cited sources, mainly for nearby and place-dependent questions. Network changes stayed within repeat-run noise in that setup. Its index is not a brand mention rate.
- API findings and guidance for consumer apps need clear labels. OpenAI says ChatGPT search may use IP-based location; an API experiment cannot settle how every app session behaves.
- Ask the firm running the tests which place signals it records and controls. Label unknown settings and keep a new market sample separate from a trend in an existing one.
Name the market a result can describe
This is a scope question. It is not a reason to test every city at once. A firm selling software worldwide and a firm supplying on-site repairs face different market questions. The useful sample follows the buying task. A large answer count cannot fill a missing place field.
Declared Location Changed Sources in a Controlled API Audit
Chen and Koudas's October 2026 preprint used 300 English commercial questions. None of the questions named a place. The figure compares declared-location settings in ChatGPT API requests. Its localization index subtracts cross-location source-domain overlap from repeat-run overlap. A positive estimate means sources changed more across settings than across repeats. Small values can still sit within noise, as the authors say of the network results below. The index is not a percent or mention rate.
| Question group | Between US cities: index | Across countries: index |
|---|---|---|
| Near me | 0.14 | 0.28 |
| Place-dependent | 0.17 | 0.43 |
| Not local | 0.00 | 0.03 |
Source: Chen and Koudas, October 16, 2026, RQ1. ChatGPT API, gpt-5.4-mini, declared locations; 19 cities, repeated requests. Index values, not percentages. Paired settings, not a time trend. Confidence intervals are not drawn here.
The question groups help a buyer read the report. A request for a nearby service depends on where the buyer is. A broad explanation of a software category may not. Do not assume a change of place has the same effect for both tasks.
For a B2B team, the lesson is to name the market a result can describe. “We appeared in this sample” is a checkable statement. “We are visible to buyers everywhere” needs much wider evidence. Keep the source question in view too. A new set of citations does not prove that a different vendor won a place on a shortlist.
Chen and Koudas preprint, October 2026, RQ1. ChatGPT gpt-5.4-mini API; declared locations across 19 cities. Point estimates: cross-setting source change relative to repeat runs. Small estimates can lie within noise. Paired settings, not a time trend. Confidence intervals not drawn. Location audit.
API Results and Consumer Guidance Describe Different Conditions
In the same preprint, changing the network exit of ChatGPT API requests produced indices of 0.00, 0.01 and 0.01 across the three question groups, within noise. The work used APIs and small models. Its network tests used data-center VPN exits. The authors caution that consumer apps and residential connections may differ.
OpenAI's current search guidance says ChatGPT may use approximate IP-based location. It also describes optional device location and possible use of saved memories. This is product guidance. It is not a fresh test of the API study. Read the two sources with their different conditions intact.
| Evidence | Actual result or stated behavior | Basis | Safe reading |
|---|---|---|---|
| Chen and Koudas API audit | Network indices: 0.00, 0.01, 0.01; within noise | Three question groups; tested API and network conditions | No extra source-localization effect detected in that arm |
| OpenAI consumer search guidance | Approximate IP location may shape local results | Product description; device sharing and memory may also matter | Do not infer that network location never matters in ChatGPT |
Source: API audit, RQ2 and limitations and OpenAI search guidance, checked November 3, 2026. The documentation row is a capability statement, not a measured effect size.
The firm running the tests should be able to say which of these settings its tool uses. A dropdown labeled “United Kingdom” is not enough detail on its own. It might change a request field, the wording of a prompt, the network used for a request, or some combination. Those are different test setups.
Do not treat an unknown setting as a known default. Record what the service can attest and what remains unclear. If the firm cannot show how it ran the tests, a report can still describe the outputs it saved. It should not claim to match a stated consumer app and session.
Our guide to API versus web monitoring covers the wider product distinction. Here, ask whether the stated market matches the place signals used in the test.
Keep Market Selection Separate From Page Translation
A market and a language are different choices. Buyers in two countries can ask in English yet need different service coverage, delivery terms or local proof. A page in a new language does not prove that a test used the right market settings.
Our multilingual GEO guide addresses when translation serves a real buyer task. Also decide which market an answer sample can describe. That keeps the language project tied to its readers. Do not judge it by a test run somewhere else with no clear link to them.
Likewise, a location experiment is not evidence that adding city names will improve a page's citations. The study changed the test setup. It did not run a B2B page-edit campaign. Use its findings to check the report's scope. They do not promise a gain from city keywords.
If the answer set changes after a market setting changes, investigate the collection change first. A different source mix is not enough to call a content campaign a success or failure. Keep the existing sample's trend intact and label the new market as a separate view until there is a sound basis to compare them.
A Market Scope Sheet for a GEO Report
Use this sheet to make a proposed report reviewable. It does not require more permissions or a new tracking tool. Ask the supplier to fill each field from its actual collection records. “Unknown” is a useful answer when it prevents a false claim.
| Report field | Evidence to request | What the report may say | What needs a separate check |
|---|---|---|---|
| Buyer market | Approved countries or service areas | Which market the sample is meant to describe | Whether all served markets need sampling |
| Question type | Exact prompt and task | Nearby service, place-dependent need or general category | Whether sources imply vendor recommendations |
| Product and mode | App or API, model and search state | The tested entry point | Whether it matches the buyer's consumer product |
| Location signals | Declared field, network country, prompt wording; unknowns | What was controlled and recorded | Device, account or session factors the method cannot attest |
| Repeats and timing | Saved answers, dates and repeat rules | The observed sample and its variation | A longer trend or a market-wide rate |
| Change record | Any change in market or collection method | When a separate view began | Whether a before-and-after claim remains valid |
The owner of the report should approve its market claim before the team approves a global score. That makes a missing field an actionable question, rather than a caveat hidden at the end of a slide deck.
Worked Example: A Service-Area Question in a Global Report
Consider an illustrative firm that maintains industrial pumps in two countries. Its dashboard labels the results “global visibility.” The saved records show one API entry point and one declared city. The network country is not recorded. No real company or test result is implied here.
The team uses the scope sheet and finds that the report title exceeds the evidence. It changes the title to name the tested city and API conditions, labels the network field unknown and preserves the saved answers. It does not relabel them as observations from a second country.
The next brief defines the other market's real service question and the conditions that a supplier must record. That is a new sample, not an automatic continuation of the old trend. The check is complete when a reviewer can match each market label to the prompt, setting and saved answer. No new data is claimed until those runs take place.
Location and AI Visibility: Common Questions
Must we test every city where we have customers?
This research gives no rule that fits all firms. Choose markets that matter to your offer and question types. Explain which ones are covered and why. A limited pilot can be useful if its label stays within its evidence.
Does a country's name in a prompt establish the user's location?
It establishes what the prompt said. It does not attest the request's network, declared location field or consumer session. Keep prompt wording and collection settings in separate fields.
Can we compare two countries with different collection methods?
Describe each result. Do not say the country alone caused the gap. The model, product, prompt or timing can also matter. State the limits before using the comparison to choose content work.
Does this replace our standard visibility measurement plan?
No. It adds detail to the market field in that plan. Keep the event definitions, saved evidence and counting rules described in how to measure AI visibility. A well-labeled answer sample is still distinct from a buyer, a visit or revenue.
Define the market scope of a report
If a global GEO report cannot explain which markets its answers represent, bring the report, target markets and collection settings to SemShift's AI visibility audit. We can help define a clear reporting scope and the evidence needed for the next check.