The mechanism
How ChatGPT Decides Which Realtor to Recommend
A step-by-step account of what happens between a buyer's question and a named agent in the answer.
Part of the AI Search Visibility for Real Estate Agents guide · Last reviewed 2026-08-09
When someone asks ChatGPT to recommend a real estate agent, the assistant does not look up a ranking of agents. It runs web searches, retrieves a small set of pages, reads their text, and writes an answer using facts it can attribute to those pages. An agent gets named when a retrieved page states, in readable text, something that answers the specific question asked. That is the whole mechanism, and every practical step for getting recommended follows from it.
Key takeaways
- ChatGPT answers location questions by searching the web in the moment, not by consulting a stored list of agents.
- OpenAI documents OAI-SearchBot as the crawler behind search results and ChatGPT-User as the live fetch during a chat; both matter and they are separate.
- The assistant quotes passages, so the page that gets cited is the one with the matching answer, not necessarily the site's homepage.
- Pages whose content requires JavaScript to appear are effectively empty to a text retrieval pass.
- Nobody outside OpenAI can see the selection weights, so treat any specific ranking-factor claim about ChatGPT as unverified.
Step one: the question becomes several searches
A question like "who should I hire to sell my condo in Burnaby" is too broad to answer from one search, so the assistant decomposes it. It will look for agents in Burnaby, for condo-specific selling advice, possibly for local market conditions and typical commission structures. Google describes the same pattern in its own AI features documentation as query fan-out: many searches issued at once, then a single answer assembled from the results.
The practical implication is uncomfortable for most agent sites. You are not competing for one query. You are competing for inclusion on each of several sub-questions, and a site that only talks about itself matches almost none of them.
Step two: pages are fetched by a named crawler
OpenAI publishes which crawlers it operates and what each does. OAI-SearchBot builds the search index that ChatGPT's search results draw on. ChatGPT-User is the agent that fetches a page live when a conversation requires opening it. GPTBot is the training crawler and is a separate question with separate tradeoffs.
This distinction causes real, avoidable losses. An agent who read that they should block GPTBot to keep their content out of model training, and instead blocked all OpenAI user agents, has removed themselves from ChatGPT's search results while believing they took a privacy precaution. If you want to be findable but not used for training, the vendor documentation is the only reliable guide to which rule does which, and it changes.
Step three: the text is read, and most sites go quiet here
Retrieval reads the HTML the server returns. A modern agent site frequently delivers a nearly empty shell and then fills it in with JavaScript, particularly for anything coming from an IDX feed. Listings, neighbourhood pages, sometimes the agent's own bio arrive that way.
You can check this in under a minute without any tools. Open your page, view source, and search the source for a sentence you can see on screen. If the sentence is not there, an assistant reading your page does not have it either. This single check explains more failures than any content advice.
Step four: facts get attributed, or dropped
An assistant that names a business is making a claim it needs a source for. Text that states something checkable travels well: a licence jurisdiction, a city, a year, a property type, a language. Text that asserts quality without a source does not, because there is nothing to attribute.
This is why "trusted local expert with a passion for client service" is invisible in a way that has nothing to do with keywords. It contains no attributable fact. "Licensed in British Columbia since 2011, specialising in East Vancouver character homes, working in English and Cantonese" contains four.
What we cannot tell you, and why that matters
The weights ChatGPT applies when choosing among retrieved pages are not published. Nobody outside OpenAI knows whether domain age, review counts, or brand mentions influence selection, and any article giving you a ranked list of ChatGPT ranking factors is inferring it. We are not going to add to that.
What is documented, and therefore actionable, is the pipeline: crawler access, readable text, attributable facts. Those are necessary conditions. Fix them and you are eligible. Skip them and nothing else you do can matter.
| User agent | Purpose | If you block it |
|---|---|---|
| OAI-SearchBot | Indexes pages for ChatGPT search results | You are removed from ChatGPT search results |
| ChatGPT-User | Fetches a page live during a conversation | The assistant cannot open your page on request |
| GPTBot | Collects data for model training | Your content is excluded from training, search is unaffected |
From OpenAI's published crawler documentation. Confirm against that page before editing robots.txt, since vendors add and rename crawlers.
Frequently asked questions
Does ChatGPT have a list of real estate agents it recommends from?
No. For location-specific questions it searches the web during the conversation and builds an answer from the pages it retrieves. There is no standing directory of agents inside the model that you can be added to.
Should I block GPTBot?
That is a decision about model training, not about visibility, and the two are separate crawlers. Blocking GPTBot keeps your content out of training while leaving ChatGPT's search results intact. Blocking OAI-SearchBot is what removes you from those results. Check OpenAI's crawler documentation before writing either rule.
Why does ChatGPT cite my competitor's blog post instead of my homepage?
Because the assistant is answering a question, and the blog post contains a passage that answers it. Homepages describe a business in general terms and rarely answer anything specific. The fix is a page that answers the actual question, not a stronger homepage.
Do I need to be mentioned on other websites for ChatGPT to trust me?
Consistent descriptions across independent sites plausibly help, because agreement between sources is easier to rely on than a single self-description. But the specific weighting is not published, so treat this as sound general practice rather than a documented ranking factor.
Related pages in this guide
Related reading
Sources
Every claim on this page that could be checked against a primary source is linked below. Where something is not publicly documented by a vendor, the page says so rather than filling the gap with an estimate.
- OpenAI crawlers and user agents (OpenAI)
- AI features and your website (Google Search Central)