AI Search Visibility for Local Businesses: What It Is and How to Measure It
A customer with a burst pipe used to open Google, scan a map pack, and call the third result. A growing share of them now open ChatGPT and type "who should I call for an emergency plumber in Austin right now." The assistant names two or three businesses and stops.
That is the whole shift, and it is unusually easy to state: in AI search there is no page two. A business is either named in the answer or it is not part of the conversation.
AI search visibility is the measure of how often a business gets named — cited — in the answers generative engines give to buying-intent questions in its category and city. It is not a ranking. It is closer to a share: out of every ten answers about your category in your market, how many include you?
Why this is not the same problem as SEO
There is real overlap in the fix work. Roughly speaking, the technical hygiene an experienced local SEO already does — consistent name, address and phone data; valid structured data; a real review corpus — is also the foundation of AI visibility. If someone tells you the two disciplines share nothing, they are selling something.
But three things genuinely differ.
The prize is binary. Ranking is a gradient: position four is worse than position two but still gets clicks. Citation is a switch. Being "almost cited" is worth nothing.
The selection mechanism is entity confidence, not relevance scoring. A generative engine names a business when its internal representation of that business is consistent enough, and corroborated by enough independent sources, that naming it is a low-risk answer. Contradictory hours across three directories is not a small hygiene issue in that model — it is a reason to name someone else.
The measurement is unstable by design. Ask the same question three times and you can get three slightly different answers. Any audit that runs a prompt once is producing noise, not data.
The nine factors that actually move citation
Across audits of local service businesses, the factors that correlate with getting named cluster into nine areas, and they are not equally weighted:
- Entity consistency — identical name, address, phone and hours across every source the engines ingest. The single heaviest factor.
- Structured data coverage — LocalBusiness schema with the specific subtype, service area, hours, and price range actually filled in.
- Review corpus depth and recency — volume matters less than recency and textual specificity. Reviews that name the service performed are quotable; five-star reviews with no text are not.
- Third-party corroboration — mentions on sources the engine already trusts: local press, chambers, trade associations, well-maintained directories.
- Quotable on-site content — pages that answer the question directly, in a sentence an engine can lift without rewriting.
- Service and location specificity — a page per service per area beats one page listing everything.
- Freshness signals — recently updated content and recent reviews both act as confidence inputs.
- Aggregator dependency — if the engines cite Thumbtack or Yelp instead of the business, the aggregator is capturing the lead and reselling it back. This one is measured as a negative.
- Answer-surface presence — whether the business appears in the sources engines pull from most heavily in that specific vertical.
How to measure it properly
The protocol matters more than the tooling, because bad session hygiene produces personalized answers that mean nothing.
Control the session. Fresh browser profile, no account logged in, no prior history. Set the location explicitly rather than letting it infer one. Personalization is the enemy of a comparable measurement.
Use a fixed prompt set. Sixty prompts across six intent categories works well: emergency, comparison, price, trust, specific-need, and local-general. Using the same set every month is what makes the trend line real.
Run each prompt three times. Generative answers vary run to run. One run tells you almost nothing; three gives you a usable frequency.
Cover four engines. ChatGPT with web search, Perplexity, Gemini, and Copilot. They disagree, and where they disagree is diagnostic — a business cited by Perplexity but not ChatGPT usually has a corroboration problem rather than a data problem.
Log three things per run: was the business named, who was named instead, and what source did the engine credit. The third one is the most useful and the most commonly skipped. Cited sources are a map of what the engine trusts in that market, and therefore a to-do list.
Turning the measurement into a number the owner cares about
"You appear in 2 of 60 answers" is a fact. It is not yet a reason to spend money.
Convert it: monthly search volume for the category in the market, multiplied by the share of that demand now resolving inside an AI answer, multiplied by the business's close rate, multiplied by average job value. The output is a monthly revenue figure attached to being absent.
Every input is an estimate and should be labelled as one. That is fine — the point of the number is not precision, it is scale. An owner who learns the exposure is roughly $400 a month behaves differently from one who learns it is roughly $14,000 a month, and both of those are honest conclusions from the same method.
What to do first
The fix work has a dependency order, and doing it out of order means doing it twice:
- Entity consistency. Fix the identity data everywhere before anything else. Publishing new content on top of contradictory data just adds another contradictory source.
- Structured data. Once the facts are stable, make them machine-legible.
- Quotable corpus. Then write the material that answers real questions in liftable sentences.
- Third-party citations. Finally, get independent sources to corroborate all of it.
First movement typically shows up in the second monthly re-run. Entity fixes propagate faster than content does.
Start with the measurement
You cannot argue about a fix you have not measured, and the measurement is the part that sells itself — nobody defends being absent from their own category.
The full 60-prompt test library is free: run it against one business tonight in a clean session and you will know within an hour whether this is a problem worth solving. The scored 9-point audit workbook, with the weighting and the revenue-at-risk model built in, is the Audit Kit at $27.