The useful measure is not a single “AI rank.” It is how often an AI answer recommends the right location for a defined set of customer questions, under recorded conditions. You also need to know whether the answer describes that location correctly. This guide shows how to build a baseline, repeat it, and decide what to do with the results.
You can run a small audit yourself with a spreadsheet. Search Console and website analytics add useful evidence later, but neither can tell you every time an assistant recommends your business without sending a click.
This work is often called GEO (generative engine optimization) or AEO (answer engine optimization). Whatever name you use, the measurement question is concrete: does the correct location appear in relevant answers, and is the information right?
First, decide what you mean by visibility
An AI answer can interact with your business in several ways. Keep them separate:
| What happened | What to record | Why it matters |
|---|---|---|
| Your business does not appear | Absent | The answer gives you no exposure for this question. |
| Your name appears, but as background information | Mentioned | The name appeared; the answer has not necessarily suggested a visit. |
| The assistant presents your location as an option that fits the request | Recommended | This is the closest observable result to inclusion on a customer's shortlist. |
| It names the wrong branch or gives an incorrect fact | Inaccurate | A mention can still send someone to the wrong place or make them give up. |
| Your website is linked as a source | Cited | A page appeared in the source trail; this is distinct from a business recommendation. |
| A person calls, books, or visits | Customer outcome | This is the business result you ultimately care about, but it needs its own measurement. |
For a business with several branches, “our brand appeared” is too broad. If the customer asks for a haircut near JLT and the answer suggests your branch across Dubai, the brand was mentioned, but the requested location was not recommended. Your report should show both facts.
These definitions are an audit method, not a published scoring system used by ChatGPT, Gemini, or Google.
If you first need to diagnose why another business appeared, read Why Does ChatGPT Recommend Your Competitors — and Not You?. This guide focuses on how to measure the answers consistently.
Build a small set of real customer questions
Start with one location and one service you want to sell. Write six to ten questions that a customer might ask before they know your name. Vary the need, not just the wording:
- Category: “Where can I get a men's haircut near Dubai Marina?”
- Specific service: “Which barber near JLT does beard shaping?”
- Constraint: “Where can I book a haircut in Marina after work?”
- Fit: “Any quiet salon near JLT that is good with curly hair?”
Check that each question is relevant to the branch. If your business closes at 6 p.m., an “open late tonight” query tests whether the assistant handles a mismatch; it is not a fair measure of an opportunity you could serve.
Keep non-branded discovery questions separate from branded accuracy questions. “What are [business name]'s opening hours?” can reveal incorrect facts. It cannot tell you whether a customer who has never heard of you would discover your business.
Use the same wording at the next check. If you replace half the questions, you have changed the test as well as the date. Add new questions when your services or market change, but retain a stable core for comparison.
Record the conditions of every check
A useful answer record includes the question, platform, date, location context, language, account or personalization settings, and the complete answer. Save any links the answer displays. For “near me,” record how the platform determined the searcher's location; if you cannot control that reliably, put the neighborhood or city in the question instead.
Run each question in a fresh conversation where possible, with the same settings on later checks. ChatGPT can use memory and other context to personalize answers; its Temporary Chat settings let users choose an unpersonalized chat. This helps make checks more comparable, but it does not make them identical to every customer's experience.
Choose the platforms your customers plausibly use. ChatGPT Search, Gemini, Perplexity, Google AI Mode, and AI Overviews are different surfaces; do not merge their results into one unexplained number. AI Overviews also do not appear for every Google search, so record “feature not shown” separately from “shown, but our business absent.” For AI Overviews, show both how often the feature appeared and how often it recommended your branch when it appeared.
Repeat the same checks on different days. For a first manual baseline, two or three rounds per question are a manageable starting point; this is a practical sampling choice, not a statistically validated universal minimum. A single answer is a case to inspect, not a trend.
Count recommendations, then inspect the answers
For each valid answer, mark whether it mentions your brand, recommends the target branch, recommends a different branch, or does not mention you. Then check any facts the answer actually states: address, phone, website, hours, services, and booking route. Record named alternatives and cited links, too.
A compact scorecard can use these measures:
| Measure | Calculation | What it tells you |
|---|---|---|
| Recommendation rate | Answers recommending the target branch ÷ valid answers checked | How often this branch makes the tested shortlist. |
| Mention rate | Answers naming the brand ÷ valid answers checked | Whether the brand appears at all, even without a recommendation. |
| Branch accuracy | Answers naming the intended branch ÷ answers naming any branch of your brand | Whether the answer points to the right place. |
| Fact accuracy | Correct facts ÷ facts stated and verified | Whether actionable details in those answers are right. |
| Source record | Links cited with each saved answer | Which pages are visible in the observed source trail. |
Always show the counts and denominator alongside a percentage. “Recommended in 4 of 12 checks” is more useful than “33% AI visibility.” If an answer contains no phone number, it should not count as a correct or incorrect phone number. If an answer recommends several businesses, record all of them; the order can be kept as context, but it is not a stable search ranking.
Here is an illustrative example, not measured customer data. A salon checks one neighborhood question 12 times in ChatGPT over two weeks. Its JLT branch is recommended four times, another branch once, and neither branch seven times. The target branch's recommendation rate in this small sample is 4/12. The brand appears in 5/12 answers. Those are different findings, and both matter.
Save the actual answers behind these counts. An owner should be able to open a result and see what was asked, what appeared, and why the team classified it that way.
To make your first spreadsheet, use one row per answer and copy these column headings:
Date | Platform | Question | Target branch | Location and account context | Answer or screenshot | Mentioned? | Recommended branch | Facts checked | Cited links | Other businesses named
Read the source trail without guessing at the algorithm
When an answer cites a directory, article, business profile, or your own website, open the cited page. Does it describe the correct location and service? Is your business present? Are hours or contact details outdated? These are concrete questions your team can act on.
A cited link shows that the page was presented with that answer. It does not prove that the page caused your branch to be selected or omitted. Nor can you ask an assistant “why didn't you recommend us?” and treat its reply as an explanation of its internal selection process. Use the answer and sources as evidence of what happened, then frame possible reasons as hypotheses to check.
For example: “A cited neighborhood guide lists three nearby salons but not ours” is an observation. “Being added to that guide will make ChatGPT recommend us” is a prediction. The sensible next step is to assess whether the guide is relevant and accepts updates, then check the same customer questions again after any change.
Measure search exposure and customer outcomes separately
Your answer checks show whether a business location appears in the questions you tested. Other tools cover different parts of the customer journey:
- Google Search Console: Its Generative AI performance report shows impressions for your site's URLs in AI Overviews and AI Mode, with views such as page, country, device, and date. Google says it rolled out the report worldwide by August 2026, but it may still be unavailable to a property or absent when there are too few qualifying impressions. It measures URL exposure in Google's AI search features, not whether your physical business was recommended by name in ChatGPT or elsewhere. The report's documented dimensions do not provide a query-level breakdown.
- Website analytics: Check visits from AI assistants and what those visitors do next. OpenAI says links from ChatGPT Search include
utm_source=chatgpt.com, which can help identify visits from that source. A customer who reads an answer and later searches your name or calls without clicking will not appear as a ChatGPT referral. (OpenAI publisher FAQ, Google Analytics traffic acquisition) - Enquiries and bookings: Add a simple “How did you hear about us?” choice or ask during intake. Treat self-reported answers as directional evidence; customers may have used several sources before contacting you.
Do not add these numbers into a single “AI score.” A URL impression, a named recommendation, a site visit, and a booking have different denominators and answer different questions.
If your current report stops at website clicks and map performance, Local SEO vs. GEO explains how to add this location-level answer check to the work you already do.
Review the trend, then choose one action
Compare each branch against its own earlier results using the same core questions, platforms, and conditions. Look at the underlying answers as well as the counts. A move from three recommendations in 12 checks to six in 12 is worth investigating; with such a small sample, it is not proof that a particular edit caused the change.
Prioritize actions by the gap you actually found:
| Finding | Sensible first action |
|---|---|
| Wrong opening hours or phone number | Verify the correct details on your own site and relevant business profiles. |
| Wrong branch suggested for a neighborhood | Make each location's address, service range, and booking link clear on its page. |
| Relevant service is missing from accessible pages | Add a useful service description for customers, with the location and real constraints. |
| A cited source contains outdated information | Find its correction or update process. |
| Low recommendation rate, with no obvious factual error | Inspect the competing answers and sources before deciding what to change. |
Google says its AI features in Search rely on its core search systems, and that useful, crawlable pages and accurate Google Business Profile information can help local businesses be visible. It also says there is no special AI schema required for those features. OpenAI advises allowing OAI-SearchBot if you want your public pages to be discoverable for ChatGPT Search. These are sound checks; none guarantees that an assistant will recommend a particular business.
A simple audit you can start today
- Pick one branch and six relevant, non-branded questions.
- Check them in one AI platform, noting the city or neighborhood and saving every answer.
- Repeat on a different day under the same conditions.
- Count target-branch recommendations, brand mentions, wrong-branch answers, and factual errors. Keep the counts visible.
- Open the sources in the answers that matter most. Choose one verifiable correction or content improvement.
- Repeat the same set later and review the new answers alongside the old ones.
The first pass can be short; a useful baseline takes repeated checks, and a multi-platform, multi-location audit takes longer. The goal is to establish a measurement method you can repeat, not to manufacture a confident percentage from a handful of responses.
Whereon is building this repeatable workflow: customer questions by location and platform, saved answers and sources, gaps for the right branch, and actions to check again. The useful report is a clear record of where a location appeared, where it did not, and what your team should examine next.
Want to see how AI presents your location? Join the Whereon pilot.
Frequently asked questions
Can I check my AI visibility without a paid tool?
Yes. A small spreadsheet and saved answers are enough for an initial audit. A tool becomes useful when you need consistent checks across many questions, platforms, dates, and locations.
Is being cited the same as being recommended?
No. A link to your site can support an answer that does not name your business as an option. A business can also be named without its own website being cited. Record citations and recommendations separately.
Why does the same question give different answers?
The answer can change with context, available information, location, timing, and the platform's behavior. One result should therefore be saved and investigated, but repeated comparable checks give a better basis for decisions.
Can Search Console tell me whether ChatGPT recommends my location?
No. Search Console reports on Google Search. Its Generative AI performance report shows impressions for eligible site URLs in Google's AI features; it does not monitor ChatGPT answers or count recommendations of a physical location.
What is a good AI visibility score?
There is no universal score that works across every business, market, question, and platform. Start with the share of relevant checks that recommend the correct branch, show the sample size, and compare like with like over time.