How Small Service Businesses Get Found in AI Search

The short version

Over the 28 days ending September 17, 2026, Google Analytics attributed 24 sessions to ChatGPT on the website of a music-lesson business I run. Fifteen of those were engaged sessions. In the most recent seven days, eight sessions were attributed to ChatGPT and six were engaged. The 28 days before that recorded five.

In the last week, four booked students reported finding the business through ChatGPT.

What the data actually shows

These figures come from the site's Google Analytics property, with data through September 17, 2026, pulled on September 18. The windows and pull dates are recorded so they can be compared later.

Weekly sessions attributed to ChatGPT, oldest to newest across the last eight weeks: 0, 0, 2, 3, 7, 9, 0, 8. The final two days were still settling when the data was pulled, so the newest value can move.

Over the full 56-day window, the busiest landing pages were the homepage (10 sessions), a Los Angeles songwriting lesson page (3), a Los Angeles music production page (2), and an online piano lesson page (2). People arrived on pages that describe specific services, not only on the front page.

Two limits matter. Consent settings, privacy tools, and browser blockers suppress analytics, so the real number of AI-assisted visits is higher than what is recorded. And "engaged session" means the visit lasted past a few seconds, reached a second page, or triggered a tracked action. It does not mean they booked, paid, or became a customer.

What changed before the referrals

I cannot prove which of these changes produced the visits, and I would be skeptical of anyone who claims they can. What I can describe is the order of the work, because the referral trend followed it.

1. Made the pages reachable and indexable

Every service page loads, appears in internal navigation, sits in the sitemap, and does not return an accidental block to crawlers. This is unglamorous and it comes first. A page that cannot be fetched cannot be summarized by anything.

2. Gave every real service its own page

The business offers specific things in a specific area, so each one has a page that states what it is, who it suits, what it costs or how pricing works, and how to start. There are no dozens of near-identical pages with different city names in them. Those pages compete with each other and give a reader nothing to choose between.

3. Made the business identity identical everywhere

The business name, area served, phone number, email, website, and credentials read the same on the site, the map listings, the professional directories, and the social profiles. When those disagree, every system that reads them has to guess which version is real, and some will guess wrong.

4. Showed proof a stranger can check

Credentials, years of practice, specific instruments and software, published policies, and real examples. Reviews from real customers, linked back to the platform they came from. Nothing invented, and nothing phrased like a testimonial that a copywriter wrote on a customer's behalf.

5. Wrote for the questions customers actually ask

The pages answer the words people use when they are looking: what the service is, who it suits, what a first session looks like, what it costs, and whether they need to prepare anything. The direct answer sits near the top of the page in plain language, before the longer explanation.

6. Started measuring and asking

Referral sources are tracked in analytics, and every new customer is asked how they found the business. So far, the second method is the only one that has produced a firm answer, and in the last week it produced four.

The measurement most businesses skip

Analytics told me that visits came from ChatGPT. It never told me those visits mattered. The only evidence that this produces customers is a person saying so; in the last week, four booked students reported finding the business through ChatGPT.

So there are two separate records:

• Attributed sessions: visits where the referring source is an AI assistant. Useful for trend, direction, and knowing which pages attract interest.
• Named discovery: a customer who describes, in their own words, how they found the business. Small numbers, high value, and the only version that ties to revenue. Four of these reports arrived in a single week, which is a small sample and not a rate.

Keep them apart. They measure different things, and adding them together produces a number that means nothing.

There is a third, weaker signal: a short fixed list of questions a prospective customer would ask, checked periodically in a fresh chat session to see whether the business appears at all. That is a sample, not a rate. In one recent pass, the business was mentioned in four of the seven discovery questions, and three of the seven produced an answer that linked to the site. The sample was small enough that it should not be quoted as a percentage. It is a direction to watch.

What this does not prove

• It does not prove causation. Several changes landed in the same period, and answer systems change without notice or documentation.
• It is not a rate you can copy. One small business, one niche, one city, one reporting window.
• The student reports are self-reports. Four people said they found the business through ChatGPT. That is what they told me, not a tracked conversion, and I have no way to know how many of them would have found it another way. Four is also a floor rather than a total: not everyone who finds a business this way says so, and there is no count of the visitors who looked and left.
• It is not a promise. No search engine or assistant is obliged to recommend any business. Clear structure, accurate details, and real proof improve the odds of being understood correctly. They do not buy placement.
• It is not a position you hold. A mention today is not a ranking. It can change with the next model update.

How to apply this to a service business

A practical order of operations, roughly the order I would do it again:

1. Write down what is actually true. Real services, real service area, real credentials, real prices or price ranges. Everything below depends on this list.
2. Inventory the pages you already have. One page per real service. Delete or consolidate thin pages that only swap a city name.
3. Standardize the facts. One canonical set of details, then correct your own listings and profiles to match before creating anything new.
4. Rewrite the two or three pages closest to a buying decision. Put the answer in the first paragraph, then support it with details, limits, and what happens next.
5. Add proof. Credentials, examples, policies, and customer reviews hosted where they can be verified.
6. Check the boring technical things. Pages load, mobile works, crawlers are not blocked, the sitemap is current, and old URLs redirect instead of breaking.
7. Set up measurement and ask people. Track assistant referrals and record how every new customer says they found you.
8. Review on a schedule. Pages, listings, and answers all drift. The check is only useful if it repeats.

Three expensive shortcuts

• Location pages at scale. Forty pages with different city names and the same body text give a reader nothing and make the whole site look manufactured.
• Mass-generated articles. Volume without first-hand detail reads as filler to people and to the systems summarizing pages. One good page beats ten thin ones.
• Treating markup as the strategy. Structured data helps a page be interpreted correctly when it matches what is visible on the page. It does not substitute for accurate content, and adding it as a ritual does nothing.

Preballin's rule

Most of this is not a writing problem. It is a maintenance problem: a short list of checks that has to happen every week, and that everyone skips when the week gets busy. That is the kind of repeated work a small workflow system can prepare for review without letting anything publish itself. The checks watch your pages, your unanswered customer questions, and your business details for drift, then hand you a short list of what changed and what to fix. A person still approves every change.

If you want the cost and scope side of that, read how much AI automation costs or how to turn repeated tasks into safe automation rules. You can also see the systems I build or browse the portfolio.

FAQ

Does AI search actually send customers to local service businesses?

It can send visits, and in this case it did, at a small scale that grew over several weeks. Whether those visits become customers is a separate question that analytics cannot answer. You find out by asking the people who book.

How do I know if it is working?

Two signals, kept separate: assistant-attributed sessions in your analytics, and new customers telling you they found you through an assistant. The second one is the only one that proves the channel produced business.

Do I need special markup or a special file for AI search?

No. There is no tag, file, or formatting trick that gets a business recommended. Google's own guidance is that its normal search practices apply to AI features. The work is standard, careful site and content work: reachable pages, accurate details, consistent identity, real proof, and clear answers.

How long does it take?

Unknown, and anyone quoting a timeline is guessing. In this case the referrals appeared gradually across a two-month window, after the page and identity work, and they cannot be tied to a specific change. The reliable part is that pages which are clear and accurate stay useful for every system that reads them.

Can any of this be automated?

The checking can. Finding unanswered questions, spotting pages that drift out of date, detecting broken links, and flagging business details that disagree across profiles are all repetitive work. Publishing, price changes, and anything customer-facing should stay behind a human approval step.