How ChatGPT, Claude and Perplexity actually pick which local business to name

Someone types "plumber near me who can do emergency callouts" into ChatGPT, Claude or Perplexity, and a name comes back. What actually happened in between is not a mystery model guessing off the top of its head. It's a retrieval step: the assistant issues a search query, a crawler-built index returns candidate pages, and the model reads those pages to compose an answer with citations.

What actually happens when someone asks

Each of the three main assistants documents a live web search step, not pure recall from training data. Anthropic's web search tool runs a query, gets back search results with page content, and lets Claude cite the source it read to build its answer, with the tool priced per 1,000 searches on top of standard token costs for whatever content it retrieves (Anthropic, web search tool docs). OpenAI runs a dedicated crawler, OAI-SearchBot, whose stated job is to "surface websites in search results in ChatGPT's search features" (OpenAI, GPTBot and other bots). Perplexity runs PerplexityBot for the same purpose: to "surface and link websites in search results on Perplexity" (Perplexity, crawlers documentation).

The index comes from crawling, same as search engines

None of these bots discover a plumber by magic. They crawl the web the same way Googlebot and Bingbot do, following links and reading page content into an index the assistant then queries. Google describes Googlebot as the crawler that "indexes web content" for Google Search, fetching pages and their resources on a recurring schedule (Google Search Central, What is Googlebot). Bing documents Bingbot performing the equivalent role for its own index, with a Webmaster Tools console for site owners to check crawl status and IP verification (Bing Webmaster Tools, crawler help). If a business's site was never crawled, or was crawled and found thin or broken, it has nothing in the index for an AI assistant's search step to retrieve.

This is also why the three providers separate their crawlers by function. OpenAI's documentation is explicit that GPTBot (model training) and OAI-SearchBot (search results) are controlled by different robots.txt tags, and a site can allow one while disallowing the other. Sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers" (OpenAI, bots documentation). A business that blindly blocks "all AI bots" in robots.txt, thinking it's only stopping model training, may be quietly removing itself from the one crawler that actually feeds recommendations.

Structured data helps the crawler read the page correctly

Once a crawler reaches a business's page, how cleanly it can extract the name, address, phone number, category and opening hours matters. Google's guidance on LocalBusiness structured data says it "enhances Google Search and Maps results" when a page includes the required schema.org properties: @context, @type, name and address (Google Search Central, Local Business structured data). That data is machine-readable by design, which is precisely the format a retrieval step benefits from over an assistant trying to infer a suburb from a paragraph of marketing copy. None of the AI providers document ingesting schema.org markup directly as a separate signal from crawling, but the same principle that helps Google's structured-result eligibility applies to any tool reading a page programmatically: explicit, consistent data is easier to extract correctly than prose.

Directories and aggregation pages do double duty

A business rarely has just its own site in the index. Directory listings, review aggregators and industry bodies get crawled too, and they often carry consistent name, address and category data across many businesses in one place, which makes them efficient sources for a retrieval step to pull from. This is also why consistency across sources matters more than volume: if a business's phone number or suburb differs between its own site, its directory listings and its Google Business Profile, a crawler-fed answer has conflicting inputs to reconcile, and an assistant citing the wrong one erodes trust in the answer, not just the business.

What this means in practice

None of the three providers publish a ranking formula for which business gets named, and this article won't guess at one. What their own documentation supports is narrower and more mechanical: these tools retrieve from crawled, indexed web content, not from thin air. A business that is unreachable by relevant crawlers, whether through a robots.txt block, a broken site, or being crawled but poorly structured, is not in the pool these tools draw from. Being crawlable, having accurate structured data where relevant, and keeping name, address and category consistent everywhere the business is listed puts it in the pool. None of it is a guarantee of a specific mention, but it is the documented precondition for one.

For businesses assessing their own visibility across verticals such as trade services or home builders, the starting checks are the same as for traditional search: is the site actually crawlable, and is the core business data consistent across every place it appears. Related background on getting the basics of a hiring decision right sits in the hiring licensed trades guide.