Mini-test to assess the visibility of AI for your local business that you do yourself (just for yourself): pick 10-20 questions customers actually ask, run each at least 3x over the course of some days on ChatGPT/Perplexity/Gemini/Claude and see if you were mentioned/cited or not at all and how the answer itself looked, calculate a median with range per question (note: the same local queries you run repeatedly will have a spread between 20% and 33% run-to-run, BrightLocal Local AI Visibility Study Sep 2026 - vendor published study) - 1x run is a story like the single data point discussed in the demand post. Average over the runs. This is a figure that you can monitor monthly, discuss, use to hold any vendor (incl us) accountable. Why you shouldn’t do it 1x below, 5 types of questions, how to do the run, how to log it, how to get the baseline as monthly figure below…
Why one run is not a number
It’s all the same in the local studies that are getting published – you ask the same thing, and get a different list back. For example, BrightLocal Local AI Visibility Study (September 2026) shows 20%-33% variation when you ask the same exact localized question repeatedly. ~50% of your business showings are the same when you repeat (vendor publication link 2026-11-04 accessed). In other words, a “best plumber in [city]” answer that mentions you now, may say 4 other shops next week. Both are potentially true to your actual situation.
That’s not the kind of variation you can hope away with extra prompts. It’s a feature of the generated answers – there’s a new roll of the dice in the selection process. No engine says it’s committed to consistency, either. So if you take a screenshot of your look, where you seem invisible or like you “made it” – that’s not a measurement. It’s just one sample from a distribution, and you haven’t seen how wide that distribution is.
Another structural point is that there is no official dashboard on how chat assistants talk about you. The generative AI impressions report (Google) only tracks Google search AI stuff. The GA4 referral rows only track the long tail of clicks to your site. AI visibility is a different metric than google rank. Check why #1 on google doesn’t mean cited in ai answers for how that works. What actually works in local seo 2026 for the local basics that drive both of those surfaces. The numbers of “named”, “cited”, “mentioned”, “missing”, “per question”, “per engine” (on the citation side of things) don’t even exist until someone actually asks the questions. And that “someone” can be you once a month, following this protocol. The 5 metrics the results feed into are in ai visibility metrics basics post. This is the local, hands-on version of the measurement.
The question set: five archetypes, 10–20 questions
Make the set of archetypes as 5. So a customer could just type them. 10-20 questions in total, enough to be decision-useful. Few enough to be a monthly thing.
- Shortlist question. “best [category] in [city]”. It is the “head-to-head”. Each answer has 2-4 slots. (Brightlocal found that, on average, answers recommend only 2-4 businesses per prompt. Vendor published study, Sep 2026 link). So this archetype is your “headline number”.
- Intent question. “where to get [service] in [city]”. Or “who should I call for [problem] in [city]”. It’s the cash question. With urgency. Often what the owner picture in their head.
- Comparison question. “[your business] vs [competitor] [city]”. Or “[competitor] alternatives [city]”. It’s about where you stand relative to others. Often gets info from review platforms. Or comparison content.
- Location question. “[problem] near me”. It’s the same content as #2, but without the city in it. So it’s a test of how well each engine uses locations. But they vary. Some find city from IP. Some (at least Claude) have no location awareness at all. Read about how each engine finds local businesses.
- Accuracy question. “is [your business] open”. “how much [service] cost in [city]”. “what does [your business] do”. It’s about what AI says about you if it does say you. Hours, prices, services. Not just that it says you. If AI is wrong about these things, customer will check. (See the list of fixes for more about this).
Now, 3 rules, so you can use it honestly for months.
- you write them down. Questions are the instrument. If you change them during use… then you “improved” into “asked different questions”. Change the set if your business changes. (New service. New city. New competitor)
- how it’s phrased is important. no keywords. “best plumber in Tampa” is better than “plumbing services tampa fl 2026”. If you can’t imagine a customer asking it. It’s not a question.
- Include the named competitors. List the named competitors. You should have archetype 3 at a minimum with at least 2 or 3 real local competitors. You’re going to get names of alternatives anyway. When you ask for it, it lets you intentionally log it.
The run setup: surfaces, controls, cadence
Surfaces: Test entire set on each of the 4 chat engines. (ChatGPT: include search, [note which model is indicated by the answer if there is one in the UI], Perplexity [note whether or not pro search is on], Gemini [web tool on], Claude.) [Optional: log Google AI Mode / AI Overviews, and old school local 3-Pack for a reference (but consider this a separate system not a 5th run of the same test. Google’s AI has the same index as google. The other 3 are each using their own stack).]
Controls: Stick to the same account, same physical location as much as possible. Location is of course a variable in archetype 4 (IP based resolution), but should not be a variable inherited for the other tests. Date of test is important, as model versions may change. So you know more if you have dates of the runs on the log. Models are changing.
Cadence: Minimum 3 runs per question for different days. Month 1. Month 2 and beyond: monthly run of the entire set with same cadence. Less than 3 runs on a question? Not median. Sample size is 1, 2. If 20-33% variance is right (and it’s saying for that sample size you don’t know anything about the underlying position).
The log: what to record per answer
One per run per question. What do you need?
Fields:
| Field | What to record |
|---|---|
| Date, engine, model | When, which of the four, model/version shown (if any) |
| Question ID, variant | Which locked question; explicit city or “near me” |
| Named? | Yes / no — was your business mentioned at all |
| Cited? | Yes / no — was it mentioned with a source, and which (own site, GBP, reviews, directory, Reddit) |
| Position in answer | Where the name falls in the shortlist (1st, 2nd, 3rd, N/A) |
| Verbatim string | The exact sentence containing the name — or “absent” |
| Competitors named | Who else made the shortlist |
| Factual errors | Anything the answer said that is wrong against your known-correct NAP, hours, or pricing |
| Scored by | Whose initials ran it — two sets of eyes on month one is cheap insurance |
Scoring vocabulary will remain the same as in metrics post, namely cited / mentioned / absent by surface. For an example row, here’s one:
(note, EXAMPLE - this is a for-illustration purpose only, not an actual result for a business)
| Date | Engine | Q | Named | Cited | Position | Verbatim | Errors |
|---|---|---|---|---|---|---|---|
| 2026-11-02 | ChatGPT | Q4 (intent) | yes | yes — own site | 2 of 3 | “…or [Business], which has been open since 2011 and takes same-week jobs.” | Hours listed as 24/7; actual is 7am–7pm |
And what people will ignore and later lament is the verbatim column. Because months later when you’ll see a report saying you’re better, it’s the actual sentence you’ll look at. And you know, that’s where your list of fixes comes in handy - if there’s an answer for you but it gets your price wrong, it’s a problem you have. Because your AI has constructed that from surfaces that you control.
Why five archetypes, not one question
The minimum viable vanity metric is 1 question/month, maximum viable sales screenshot. An extra hour goes to 5 archetypes because failure groups by archetypes. Failure groups by archetypes, and each group needs a different fix.
Not existing anywhere on archetype 1 = it’s a shortlist issue. You’re not even in the set to start. Get on the list to crawl. Make your profile better.
2 and 3 but not 4 = location issues. The engines are not figuring out your city for “near me” searches. Some engine behavior, but also how prominent your profile/citations are for address.
Reviewed but not own site cited = not a visibility issue. It’s a content issue on your site.
Named everywhere with errors archetype 5 = Accuracy issue. You have some stale/inconsistent data you leave on the shelves and the AI is picking that up.
You can’t distinguish between those 4 failures with 1 question. Same reason why median is more important than any cell. Because in same archetype, 2/3 vs.
3/3 runs is often the same underlying position. And going from 3/3 to 1/3 in a month is what you should look into first before the phone stops ringing.
Making it a monthly number
The output for month zero is what we call the before. 5-archetypes x 4-engines matrix for median/spread/verbatim string. With competitor set named. So any later claims of improvement from you or an agency or tool, is relative to that matrix. With the same locked questions. Format of the report containing these numbers. And rules about what not to put in it. (No traffic promises. Samples being samples etc.). Explained in the monthly AI visibility report template.
Do realize this is testing for presence in answers. That’s the only place anyone can cleanly define a local AI metric. Presence in answers doesn’t mean how many customers you get from it. We’ve yet to see any engine put out click/visit volumes for local businesses. From AI answers. So any numbers you’re told for “AI leads”, isn’t a measurement. The honest story of measuring AI referral is in how to measure ai search traffic. And why you still have real demand for it even without that data is in Do customers find local businesses on chatGPT.
Once you’ve got the baseline, it’s usually next to go for a ranked list of what actually moves this stuff. By strength of evidence. In How to get your local business recommended by AI. Or if you want the baseline run/logged/re-run monthly. With a named competitor set. And exactly what question set/surfaces/run numbers/reportage is outlined before hand. See the methodology for that.
Sources double checked against linked originals 2026-11-04