ChatGPT Checks Companies’ Own Websites. Would It Find Yours?

When ChatGPT is asked which company to choose, it may go and check the companies themselves. In a study published this year, the search agency Nectiv found that ChatGPT often narrows an answer down to a handful of companies and then searches each one’s own website “to look for information directly from the source.” It does this with the “site:” operator, a search limited to one website, which appeared in 64% of the behind-the-scenes searches Nectiv recorded.¹

That detail has a consequence Nectiv doesn’t draw out. A “site:” search is still a search engine query. It can only find pages the search engine has already stored, and only when their wording matches what was searched. So a contractor’s page about a service can exist and still fail the check, if the search engine hasn’t picked it up or if the page calls the job something else. And unlike most of what happens inside an AI answer, it’s a check an owner can run for themselves.

Nectiv’s prompts covered verticals including software, e-commerce, travel, real estate, fashion and credit cards. Home services aren’t among the ones it reports, so applying this to contractors is our inference, not its finding.

How ChatGPT checks a company before naming it

Nectiv reran about 4,000 prompts from a study it had done a year earlier, this time on OpenAI’s newer GPT 5.6 Sol model, and extracted the searches ChatGPT ran behind each one. It averaged 7.61 searches per prompt, up from 2.17, and as many as 29 for a single prompt. The pattern it describes has two steps: broad searches to work out which companies are worth considering, then narrow searches on those companies’ sites.¹

The 64% isn’t only companies’ own sites. Nectiv notes that ChatGPT also ran “site:” searches on government sources and social platforms, and often added the word “official” to its searches. That was the second most common term it added.¹ But the narrowing to specific companies, and then to their websites, is the step that matters to a business.

For a homeowner’s question, our reading is that the first step would draw on wherever a contractor already appears: listings, reviews, directories. If the assistant then shortlists a company for “who fixes garage door springs in Hutto”, the second step would search that company’s site for garage door springs, and for Hutto.

What a “site:” search can miss

That second step can fail for reasons that have nothing to do with the quality of the work:

• The page isn’t stored. A search engine can only return pages it has indexed. A new page, or a site the search engine barely visits, may not come back at all.
• The words don’t match. If the homeowner says “garage door spring” and the site only says “torsion system repair”, a narrow search may not connect them.
• The town isn’t there. A business that serves Hutto but names it only in its Google profile gives a search of its own site nothing to find.

None of them is obvious from the outside. Each could leave a shortlisted contractor with nothing on its own site for the assistant to confirm.

How to see what ChatGPT would find on your site

The searches Nectiv recorded came through OpenAI’s API, and we don’t know which search engine ChatGPT used behind them. But the same kind of search in Google is a rough, free stand-in. Search for site:yourdomain.com followed by a service and a town you serve, in the words a customer would use. If nothing relevant comes back, the page is missing, isn’t indexed, or doesn’t use those words.

It’s an imperfect test, but it puts the question in the right place: not “is my business good?” but “can a search inside my own site find the answer?”

What the study doesn’t show

This is one agency’s analysis of prompts it chose, run through the API, which may not match the ChatGPT app exactly. It doesn’t cover local service questions, and it doesn’t measure whether a site’s content changed what ChatGPT recommended. Nectiv found the words ChatGPT added to its searches had “very little overlap” with a year earlier, which is a reminder that the next model update could change the pattern again. Nectiv sells search and AI-visibility services.¹

A Rule & Form site gives every service and every town its own page, and when the owner tells us something changed, the affected pages update once they approve it. Once a site is live, our AI checks also count when an assistant opens one of its pages while answering someone. That’s a related sign: an assistant reading your page, if not the search that led it there. It doesn’t tell anyone what the assistant concluded, and no page can guarantee a recommendation.

Sources

1. Chris Long, Nectiv, “What We Learned From Analyzing 28K+ ChatGPT And Gemini Fan-Out Queries”, 2026. Fan-out queries extracted through OpenAI’s API (GPT 5.6 Sol) for about 4,000 prompts in verticals including software, e-commerce, travel, real estate, fashion and credit cards, compared with the same prompts in 2025. Nectiv is a search marketing agency.