Local businesses get recommended when an assistant, while searching for an answer, finds them on the pages it trusts for that question: directory profiles with reviews, and the business's own service pages that answer exactly what a customer asked. Those pages must be reachable by crawlers and indexed, and even then assistants name businesses mostly on "who should I choose" questions, rarely on questions about costs, risks or treatments.
The rest of this article shows what that looks like in a real measurement, which levers you control, and the mistakes that quietly keep a business out of answers.
What happens when people ask an assistant about a local business?
On 20 October 2026 we put 20 patient questions about an orthopaedic practice in Bonn to ChatGPT, Claude, Gemini and Perplexity. The four assistants named the practice in 12 of 80 answers: Claude in 5 of 20, Gemini in 4, ChatGPT in 2 and Perplexity in 1.
The sources behind those answers were more telling than the score. The doctor directory jameda.de was cited in 12 answers. The most cited sources overall were the own websites of competing practices. And the type of question decided everything: "who should I go to" questions named practices, while questions about costs, risks and treatments named none.
That gives a local business two routes into an answer:
- Through a directory profile that the assistant cites when it looks for options.
- Through its own website, where service pages answer the question better than a competitor's page does.
The numbers also show that each assistant behaves differently. A practice that shows up in Claude may be nearly invisible in Perplexity, so testing one assistant tells you little about the others.
Which questions actually lead to a recommendation?
Most questions customers ask never produce a business name. In a 23 September 2026 test of 20 German business questions, the model answered 10 without searching at all and cited nobody. Where it did search, it cited primary sources such as laws, ministries, the OECD and large accounting firms on 9 of 10. Firms were named only on "who should I hire" and "how do I choose" questions.
For a local business this means two kinds of work:
- Selection questions ("best orthopaedist in Bonn for knee pain", "which plumber in Cologne does emergency call-outs on Sundays"). Here the assistant names businesses, and you want to be one of them.
- Informational questions ("how long does recovery from knee arthroscopy take"). Here the assistant usually names nobody, but it may cite a page as its source. When an assistant searches the web, OpenAI's web search documentation describes inline citations tied to specific text plus a separate list of every URL consulted, which is typically larger than what gets cited. Being in that consulted set is the precondition for being cited at all.
If you want the reasoning behind which businesses get picked, we cover it in how ChatGPT decides which businesses to recommend.
Why do directory profiles matter so much?
In the Bonn example, a directory was cited in as many answers as the practice itself was named. For selection questions, an assistant looks for pages that list several options with details and opinions, and directories are built exactly that way.
What to do with your profiles:
- Claim and complete them on the directories that matter in your field and region. For doctors in Germany, that clearly includes jameda.de; for a restaurant or a tradesperson it will be other platforms.
- Use the same name, address and phone number everywhere. If one profile says "Orthopädie am Markt" and another "Dr. Weber Orthopaedics", a model has to guess whether they are the same business.
- List services in the words customers use, not internal labels. "Sports injuries of the knee" matches a question; "conservative orthopaedics" often does not.
- Keep opening hours, languages spoken and insurance or payment details current. These are the details people add to their questions.
On your own site, mark up your location with structured data. Google's LocalBusiness documentation says to use the most specific subtype available and lists name and address as the required properties. Treat markup as hygiene, not as the lever: we have seen a site with complete markup but only 75 URLs stay absent from answers on its own article's exact topic. More in does schema markup help with AI search.
Do reviews help an assistant recommend you?
We do not have a measurement that isolates the effect of reviews, so we will not give you a number. The mechanism, though, is straightforward. The directory pages assistants cite for selection questions are the pages where reviews sit. Review text is written by customers, in customers' words, about specific services.
A review that says "got an appointment for my knee within a week and the doctor explained the MRI clearly" contains the same words as the question "orthopaedist in Bonn with short waiting times for knee problems". A star rating alone carries none of that.
Practical consequences:
- Ask satisfied customers for a review and suggest they mention what they came for, within each platform's rules.
- Reply to reviews, including critical ones, with specifics. Your reply is also text on the page.
- Never buy or write reviews yourself. Beyond the legal and platform risk, it does nothing for the specific, service-level wording that makes a review useful as a source.
What should your service pages look like?
The most cited sources in the Bonn measurement were competitors' own websites. That is the opening for any local business: a practice website can be a source, not just a brochure.
Build one page per question your customers really ask, not one page per department. Instead of a single "Knee" page, a practice could have pages such as "Knee arthroscopy in Bonn: who it is for and how long recovery takes", "What statutory insurance covers for knee MRI" and "Knee pain after running: when to see an orthopaedist". Each page answers its question in the first lines, then gives the details: who does it, where, how long it takes, what it costs or how billing works, and when it is not the right option.
What goes on the page matters as well. The research paper that coined the term GEO, Aggarwal et al. (KDD 2024), found that adding citations, quotations from relevant sources and statistics to content could raise a source's visibility in AI-generated answers by more than 40% across a 10,000-query benchmark. For a local business that translates into citing guidelines or insurers where relevant, stating your own real figures (appointment lead times, years in practice, number of procedures if you track them), and being specific about your town and neighbourhood.
Google points the same way. Its announcement of new guidance talks about valuable, non-commodity content for its generative AI features. A page that repeats what every other practice says gives an assistant no reason to pick it.
Volume of good pages counts too. In our data, a site with about 190,000 crawlable URLs was crawled constantly, while the 75-URL site above stayed invisible. You will not need hundreds of thousands of pages, but a website with five pages covers only a handful of questions.
Can the assistants reach your pages at all?
This is where many businesses fail without knowing it, and we did too. On itgrows.ai, a robots.txt rule meant to hide a single redirect page (/b) was written as a prefix. Because robots.txt rules match prefixes, from 11 August to 21 October 2026 it also closed /blog, /business and every article in our business blog to all crawlers. Bing Webmaster Tools listed our business page as "discovered but not crawled", and Google had none of those pages indexed. We changed the rule to /b$ on 21 October 2026. On that same day, measuring 40 questions across four assistants three times each, we were named in 0 of 160 answers.
Checks worth doing:
- robots.txt: make sure AI crawlers and search engines are allowed, and that no rule accidentally matches more than you intended. See is my website blocking GPTBot, ClaudeBot and PerplexityBot.
- Indexing: Google says its AI features such as AI Overviews surface content from its existing Search index, so a page Google has not indexed cannot appear there.
- Findability beyond Google: on 19 October 2026 we asked assistants for the exact titles of 18 articles older than three weeks on three client sites. ChatGPT found 1 of 18, Claude 7 of 18, while Google had indexed 29 of 30 articles of one of those sites. Being in Google's index did not mean an assistant could find the page.
Crawlers are active once the door is open. On one edtech platform we work with, ClaudeBot made 19,272,125 requests and GPTBot 2,744,985 between 28 July and 11 August 2026. In the same fortnight, live retrieval agents, which fetch a page while an assistant answers a real person, hit that platform's content pages 1,918 times; a client blog of a few dozen articles received 56 such fetches across 24 articles.
How do you know whether it is working?
Do not judge by website traffic. In about two weeks of logs from three course websites we run, human visitors arriving from AI assistants were: Copilot 1, every other assistant 0. Assistants mostly answer in place. For a local business the outcome is the phone call or booking that follows a recommendation, which rarely shows up as a referral in analytics.
Measure the answers instead, and do it consistently:
- Write down 15 to 20 questions your customers really ask, mostly selection questions with your town in them.
- Ask each assistant each question several times. The same question asked three times often names different businesses. We count a business as named if any of three asks names it, and compare weeks only on the same questions and the same number of asks.
- Record which sites are cited, not only whether you are named. If a directory keeps appearing, that profile is your priority.
This is cheap to run: with API prices of October 2026, a report of 20 questions across four assistants with three asks each costs us about $5.40. The method is described in how to measure AI visibility over time.
To see where you stand now, the free check at itgrows.ai generates five buyer questions from your homepage and shows within about 30 seconds whether ChatGPT and Perplexity name you. If you want the work done, ItGrows publishes question-answering articles on your own domain, fixes crawler access and indexing, and measures four assistants weekly for $499 a month (plus a one-time $499 setup, free with 3 months or a year). Nobody can guarantee that an assistant will name a business, and we do not either.
Frequently Asked Questions
Is a Google Business Profile enough to get recommended by ChatGPT?
We have not measured Google Business Profiles in isolation, so we cannot say how much weight they carry. In our orthopaedic-practice measurement, the cited sources were a directory and practices' own websites, which suggests a complete profile plus strong service pages is the safer combination.
Why does an assistant name my competitor but not me?
Usually the competitor has a page or profile that matches the question more closely, or your pages are not reachable or indexed. Check robots.txt and indexing first, then compare which pages the assistant cites against what you have published on the same topic.
How many service pages does a local business need?
There is no fixed number; the guide is the number of distinct questions your customers ask. Start with one page per common selection question and one per frequent informational question, each answering its question directly in the first lines.
Can I pay an AI assistant to recommend my business?
Assistants' regular answers are not something you can buy a place in; they come from the sources the model finds and trusts. Your budget is better spent on profiles, reviews, service pages and making sure crawlers can read your site.