# How AI assistants decide which businesses to recommend _By Nakul Kelkar, September 29, 2026_ An AI assistant recommends businesses in the context of a particular request. The customer’s question sets the need; the service may retrieve outside information; the model produces an answer from the information available to it. You can inspect the question, the response and any citations. You cannot recover a complete selection formula from those observations. For a business owner, the useful work is to trace what can be checked. Which need did the question describe? Which firms appeared? What evidence did the answer provide? Is the information about your business accurate? ## The question changes the comparison “Find a personal injury lawyer in Austin” and “Find a Spanish-speaking lawyer for a car accident in Austin” describe different needs. The second adds a language requirement and a type of case. A practice may serve both needs, but the public information supporting each could be different. Write down the exact question when evaluating a recommendation. Include the city, service, constraints and any earlier conversation that could affect the response. If you ask “Which is best?” after discussing three firms, you have supplied a shortlist. That answer measures a different situation from a fresh request with no firms named. Use customer questions that fit your business. A firm that handles estate planning should not treat absence from a personal-injury answer as a visibility problem. ## Retrieval can bring outside sources into the answer When a service uses web search, it can retrieve pages relevant to the request. Those pages may include official business websites, directories, comparison articles and local coverage. [OpenAI's web-search guide](https://learn.chatgpt.com/docs/web-search) explains how ChatGPT provides citations when it searches, giving you links to inspect alongside the answer. Some responses use other available information, including the conversation and the model's existing knowledge. A cited answer gives you more material to inspect, but its citations still do not expose every input. An uncited answer provides less evidence about where a particular business description came from. Keep the product and mode in your record. A search-enabled answer and an answer produced without web access should not silently become interchangeable rows in a report. ## Citations give you research leads Open the pages attached to a recommendation and check their contents. Does the page mention the named business? Does it support the service, location or qualification described in the answer? Is the page current enough for the fact you are checking? In the [Austin personal-injury study](/blog/ai-personal-injury-lawyer-austin), the September 27, 2026 report lists cited-source counts of nine for Super Lawyers, eight for Avvo, six for Expertise.com and five for Justia. Those counts make the pages worth examining for that sample. They do not establish how much weight an assistant gives each directory. A directory can also be relevant to a response without every listed firm being named. Buying a listing or adding more words to a profile does not guarantee inclusion. Treat a source as a place to verify facts and investigate coverage. ## Different services can produce different shortlists The Austin study requested ten questions across three services. Twenty-seven answers returned, naming 57 distinct firms; seven firms appeared across all three services. FVF Law Firm appeared eight times in ChatGPT answers, three times in Perplexity answers and zero times in Google's AI answers in that sample. The counts describe returned responses on one day. They do not rank the firm's legal work or show how many clients contacted it. The visible variation is a reason to record each service separately. Combining everything into one score would hide which service named the firm and which did not. The [interactive sample report](https://vaaya.ai/aeo-monitor/law-firms#sample) lets you inspect questions, names and citations together. This study cannot tell us whether one profile change would alter a future answer. It also cannot establish which information was in a model's training data. Those claims need evidence beyond a set of recommendations. ## Turn the observations into a specific task Suppose a cited profile gives an old office address while your website shows the current one. You have found a factual correction to make. Suppose an answer describes a service you no longer offer. Check the cited page and your own service pages before deciding where the outdated information persists. Record the correction and its date, then repeat the same questions under similar conditions. If a later answer changes, preserve that result without automatically attributing it to your edit. Sources, settings and responses can change for other reasons. Use the [guide to getting named in AI answers](/blog/get-named-in-ai-answers) to organize that work. [AEO Monitor](https://vaaya.ai/aeo-monitor) can provide a first set of questions and responses to inspect; start with an answer whose business description you can verify against its sources. ## Questions **Why do AI assistants recommend different businesses?** The question, conversation, service, settings and available sources can differ. Compare responses under recorded conditions before drawing conclusions about a business's visibility. **Does a citation explain why an AI recommended a business?** It identifies a source attached to the response. It does not expose every model input, a source's weight, or the cause of a business's inclusion. **Will a directory listing get my business recommended?** A relevant, accurate listing gives readers another way to verify your business. It does not guarantee an AI mention.