AI Knows Your Business. Does It Recommend You—and Get Your Details Right?

AI Search Visibility Is Three Different Outcomes

On September 29, 2026, Insites published its first Local AEO Benchmark, based on September audits of 8,016 local businesses across up to seven AI assistants. AI could find 99.1% of the businesses in at least one assistant, yet 40.3% were never recommended for any of their services in the questions tested. Being in the answer system is not the same as being in the buyer's shortlist.

The same study found a second gap: where businesses published a phone number, the assistants gave it correctly only 53.2% of the time. That is not merely a marketing measurement. A wrong number can turn a recommendation into a dead end, and a business that counts the mention as a win may never notice the customer could not reach it.

This is why an AI search visibility measurement program needs three separate questions: Does the assistant know who we are? Does it suggest us when a buyer describes a need? Does it communicate the facts needed to contact us? A branded prompt answers the first question. An unbranded, service-specific prompt answers the second. A fact check of the response answers the third. None can substitute for the others.

The benchmark is a vendor's sample of businesses its customers chose to audit, not a census of every town or a guarantee of what your customers see. Its value is the distinction it exposes. Treat the published percentages as a reason to test your own market, not as your own conversion rate.

Why One Screenshot Is a Bad Scorecard

A screenshot of your name in ChatGPT is easy to show a colleague. It is also an incomplete experiment. You may have asked the assistant to identify your company by name, supplied its location, and effectively handed it the answer. The buyer who has never heard of you asks a different question: who can fix this specific problem in this specific place?

A separate September 30 study of 28,488 Florida businesses tested nine cities, 18 industries, three question types, and four AI surfaces. Among businesses that appeared at least once, 43.2% appeared in only one of the available searches; 58.7% appeared on only one platform. That is a snapshot of a defined Florida sample, not a national estimate. It demonstrates why a single appearance cannot establish dependable visibility.

The study also found that 37.4% of local businesses appearing for a specific-need question had not appeared in either of its broader question types. A broad query such as “best accountant in Tampa” and a request for an accountant who knows interstate manufacturing tax issues are not interchangeable tests. If your business serves a particular need well, a generic leaderboard can hide its strongest opportunity.

Keep the denominator honest

Suppose a plumbing business tests five prompts on three assistants twice each month. That yields 30 answer opportunities for the period, not a universal ranking. Count how many responses recommended the business, and preserve which prompt and platform produced each appearance. If an assistant gives no AI-generated answer for a query, record that as a separate result rather than quietly scoring it as a failure to recommend.

Traditional organic clicks, map-pack placement, and AI recommendations are related but not identical. Keep Search Console and your local listing reports alongside the new scorecard. Do not add all their percentages together or infer that a mention created a sale. Where your intake team can ask how a customer found you, record that answer too, with the ordinary caveat that self-reported attribution is imperfect.

Repeat without pretending the answer is fixed

AI responses can differ with timing, account settings, location, and wording. Hold those conditions as steady as you can for a baseline, then repeat the same prompts on a predictable schedule. One changed answer is an observation; a sustained pattern across multiple prompts and runs is a reason to investigate. Save the date, exact prompt, surface, location, response, and cited sources so another person can reproduce the review.

Do not turn the scorecard into a contest to manufacture the highest percentage. It should help an owner decide what to fix: absent recommendations, an incorrect phone number, an outdated service area, or weak evidence for a specialty. A measure that cannot direct work is just a chart.

Build a Local AI Search Measurement Sheet That Leads to Action

A small business does not need a sprawling analytics program to start. A sheet with one row per prompt, platform, and test date is enough, provided the columns represent buyer outcomes rather than vanity mentions. Set the questions before you see the results so the team does not keep rewriting prompts until the company appears.

Start with a fixed question set. Choose three to five real buyer needs: a broad category query, a direct recommendation query, a specific problem, a service-area edge case, and a comparison or urgency question. Use wording that customers actually use in calls and emails. For a commercial refrigeration company, “emergency walk-in freezer repair near Columbus” is more useful than a prompt engineered around the company's exact name.

Record the surface and context. Keep ChatGPT, Gemini, Perplexity, and Google AI responses in separate rows. Note whether Google displayed an AI Overview or AI Mode response at all; do not silently treat a missing AI panel as an answer that rejected your business. The Florida research found that no Google AI Overview appeared in 37.4% of its tested Google opportunities. That figure belongs to that study, but it illustrates why presence of the answer format is part of the denominator.

Score three stages independently. For recognition, ask a branded question and check whether the answer identifies the right business. For recommendation, ask the preselected unbranded question and record whether the company is actually put forward as a suitable choice, not just cited as background. For accuracy, inspect the name, website, phone, location or service area, and relevant service description. A recommendation with a broken contact path needs its own error flag.

Capture the source trail. Note any citations, listing pages, reviews, or site pages the answer offers; never assume a cited page caused the recommendation. Insites reported that only 13.2% of sources cited about the audited businesses were their own websites. That makes third-party facts worth checking, but it does not prove that editing any one directory will cause a model to recommend you next week.

Assign the fix to an owner. The person who notices an error should be able to send it to whoever controls the listing, website, review response, or customer-service script. Record when the correction was made and retest after a reasonable interval. A weekly review of an untouched sheet is not an operating process.

What bad looks like—and what good looks like

Bad: A local heating contractor searches its own brand name in a single assistant, sees a flattering paragraph, and declares victory. Six weeks later a prospective customer asks for after-hours commercial boiler repair. The contractor is not recommended, while an old directory still lists a disconnected phone number. The team has a positive screenshot and no record of the lost contact path.

Good: The contractor preselects six unbranded questions covering commercial boiler work, emergency response, and its actual service towns. Each month it tests those prompts across the surfaces its customers use, logs answer availability and recommendations, and checks every phone number and service-area claim that appears. An AI coworker can prepare the comparison and flag changes; a person verifies the response and approves public corrections. The team learns that its strongest specialty page is unclear about emergency coverage and that one independent listing contains an old number. Those are fixable findings, not a vague demand for “more AI SEO.”

Run a Four-Week Baseline Before You Change the Plan

The first month is for establishing a trustworthy baseline, not proving that a new page worked overnight. The sequence below keeps your measurements comparable while allowing urgent factual errors to be corrected immediately.

1. Write down the commercial questions. Ask sales or the front desk for recurring requests, not just the phrase the marketing team wants to rank for. Divide the prompts into broad, direct recommendation, and specific-need groups. Note the geography and intended customer for each.

2. Inventory facts that must never be wrong. Keep one authoritative record for your business name, public phone, site URL, hours, service areas, and core services. Where hours or coverage vary by service, specify that rather than forcing everything into one flat listing.

3. Run the baseline and save evidence. Use the same query set, surfaces, and location context each week. Record the answer format, businesses named, rank within the answer if meaningful, facts stated, and available sources. Do not automate interactions in violation of a platform's terms; manual sampling is a legitimate start.

4. Separate errors from absences. An incorrect phone number needs urgent correction even if the business is recommended. A correct branded answer with no unbranded recommendations points to a different problem: buyers' use cases may not be clearly described or corroborated. A missing AI response on a Google query is neither of those.

5. Fix the most consequential gap. Correct wrong public facts at their source; make a service page explicit about the problem, buyer, and service area; request honest reviews that describe actual work without scripting customers' opinions. Prioritize the path from recommendation to contact before polishing a visibility chart.

6. Review changes against the original prompts. Compare four weekly runs, annotate site or listing changes, and look for a repeated shift rather than celebrating one favorable answer. If the questions themselves change because the business changes, start a new series and keep the old one for context.

This is a frontier-operations problem: the boundary between collection and judgment needs to be clear. A persistent AI coworker can keep the schedule, gather comparable evidence, and prepare a list of anomalies. A human decides whether a recommendation is commercially relevant, whether a statement is misleading, and which correction is authorized. That seam makes the measurement both repeatable and accountable.

For a broader explanation of why citation alone is not enough, read how AI search can use your content without naming your business. For the specific local discovery path inside ChatGPT, see how local recommendations differ from traditional search. The scorecard here is the missing measurement layer between those two problems and the work required to address them.

FAQ

Q: What is AI search visibility measurement for a local business? A: It is a repeatable check of whether assistants identify the business, recommend it for relevant unbranded buyer questions, and state its important details correctly. Record the exact questions, surfaces, context, date, and answers so results can be compared over time.

Q: If ChatGPT recognizes my name, am I visible to new customers? A: Not necessarily. Recognition on a branded prompt shows that an assistant has information about you. A buyer who does not know your name asks for a service and location. Test that unbranded question separately and verify the resulting contact details.

Q: How often should a small business test AI recommendations? A: A weekly run for four weeks establishes a useful baseline; a monthly operating review can then monitor the same fixed question set. Increase frequency if you change your services, discover an incorrect phone number, or see a material drop across several answers.

Q: Is there a single AI search ranking I can track? A: No dependable universal ranking exists across assistants, questions, locations, and time. Track the share of your predefined, commercially relevant answer opportunities in which you are recommended, alongside answer availability and factual accuracy. Explain the sample size whenever you report a percentage.

Q: Will adding schema or an llms.txt file guarantee recommendations? A: No. Clean technical signals can help systems interpret a site, but a recommendation also depends on the question and the broader evidence available. Prioritize accurate listings, useful service information, customer evidence, and a working path to contact; test results instead of promising a particular inclusion.

If your team needs a standing owner for this kind of discovery work, Associates AI Teammates are always-on AI coworkers that run on your own agent server, so recurring review work has an owner between check-ins. Explore the platform and choose a plan.