How AI assistants decide which businesses to recommend

Two ways an assistant knows about a business

Every assistant draws on two kinds of knowledge, and it helps to keep them apart.

• What the model learned in training. The language model behind each platform was trained on a very large collection of text, much of it from the public web. If your business was written about widely and consistently before that collection was made, the model may recognise your name and associate it with your category and city. This knowledge stops at a cut-off date, can be out of date, and cannot be edited by anyone outside the platform.
• What the assistant finds when it searches. For many questions, especially ones about local businesses, prices, opening hours or “the best” of something, the assistant runs live web searches, reads a selection of the pages it finds, and writes its answer from them. Here, what matters is what the web says about you today, and whether the assistant can reach and read it.

Most recommendation questions lean on the second kind. That is good news: answers built from search respond to changes you make, usually faster than a model’s training data does.

What happens when the assistant searches

The details differ by platform and much of them is not public, but the general pattern, described in the platforms’ own documentation and visible in their answers, is similar:

1. The question is rewritten into searches. An assistant rarely searches for the customer’s exact words. It turns the question into one or more search queries, sometimes several at once. Google has described this for AI Mode as “query fan-out”: one question becomes a set of related searches.
2. Pages are retrieved and read. The assistant fetches a selection of results. Pages it cannot load, such as those blocked in robots.txt, behind a login, or built so that the text only appears after JavaScript runs, may be skipped or only partly read.
3. Candidates are compared with the question. The assistant looks for businesses that match what was asked: the service, the place, the price range, the language, any constraint the customer gave. A business that several independent sources describe in matching terms is an easier choice than one that appears once.
4. The answer is written, sometimes with citations. The assistant composes a short answer, often a list of options with a sentence on each. Perplexity shows its sources for every answer; ChatGPT, Gemini, Claude and Google AI Mode show links or source cards when they have searched.

Two consequences follow. First, a business can be recommended because of pages it does not own: a directory listing, a news article, a roundup of “the best places”. Second, a business can be missing from the answer even when its own website is excellent, if nothing the assistant read put it in front of the question.

The five platforms are not one system

It is tempting to talk about “AI” as if it were one engine. In practice the five platforms Pharos tracks behave differently enough that a business can be named on one and absent on another for the same question.

• ChatGPT decides for each question whether to search the web. When it does, it draws on third-party search providers and on OpenAI’s own search crawler, and its answers may take the user’s approximate location into account.
• Gemini can ground its answers in Google Search, so the signals that shape Google results, including Google Business Profile information for local questions, can reach its answers.
• Google AI Mode is a conversational mode inside Google Search. It is built on Google’s index and search systems, and its answers can differ from one country to another.
• Perplexity searches for almost every question and cites its sources inline, which makes it the most transparent of the five about where an answer came from.
• Claude can search the web when that feature is available and switched on, and otherwise answers from what it learned in training.

These descriptions are deliberately general. Each company changes its systems often, and none of them publishes the logic behind the businesses it recommends.

What tends to make a business easy to recommend

Nobody outside the platforms can list the exact factors. What can be said is which properties make a business easy for any system that works this way to find, understand and trust:

• Reachable pages. Your website loads without errors, its main text is in the HTML rather than added later by scripts, and robots.txt does not block the crawlers the platforms use for search.
• Clear, specific facts. What you do, where, for whom, at what price level and in which languages, stated in plain sentences on your own pages. An assistant cannot recommend you for teeth whitening in Brooklyn if no page says you offer it there.
• Consistency across sources. The same name, address, phone number and description on your site, your map listings and the directories in your category. Contradictions give the assistant a reason to prefer someone else.
• Independent mentions. Reviews, press coverage, directory listings and roundups written by others. They show that people other than you consider you a real option.
• Coverage of the actual question. Pages that answer what customers ask: comparisons, prices, how a service works, what to expect. These are the pages assistants read and cite.
• Presence in the customer’s language. A question asked in Spanish may be answered from Spanish-language sources, so a business whose customers ask in more than one language needs pages in each.

Why the answer changes from one ask to the next

Ask the same assistant the same question twice and you may get two different lists. Answers are generated, not looked up, so every response carries some randomness. The searches the assistant runs, the pages it happens to read, the user’s location, the conversation so far and, on some platforms, the user’s saved memory can all shift the result.

This is why a single screenshot proves little. One answer tells you what happened once. To know whether a business is usually named, you have to ask the same questions repeatedly, on each platform, and look at the rate. Measuring AI visibility explains how.

What you cannot control, and what you can

You cannot edit a platform’s training data, buy a guaranteed place in an organic answer, or see how any platform scores businesses internally. Be wary of anyone who promises otherwise.

You can control whether the platforms can read your site, whether your facts are clear and consistent, whether the sources that cover your category include you, and whether you answer the questions customers ask. Those are the levers. The other guides in this series take them one at a time, starting with a practical checklist for ChatGPT.

How Pharos helps

Pharos asks the questions your customers ask on ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, on the schedule your plan sets, and keeps every answer in full. It shows how often you are named, where you rank against competitors, how you are described and which sources the answers cite, so you can see which of the levers above is holding you back. The free audit runs a first set of questions for you.