ChatGPT is recommending my business. How did that happen?

Surely this is just SEO

“Surely this is just SEO” is a fair first reaction. Your business already appears in search, other sites link to it, and ChatGPT can use public web pages when it answers buyer questions. The recommendation can look like a familiar ranking result with a new interface.

That is the first useful objection.

If you are asking, “chatgpt is recommending my business how did that happen”, the practical answer is usually more specific. A buyer asked a question, the system selected sources for that question, and one of those sources mentioned your business in a way the answer could use. ChatGPT did not necessarily discover your homepage and form an independent opinion about it.

The path runs through a page.

The page may be yours. Often, it is not. In the CiteGraph research data, listicles and roundups account for 67.5% of recorded citations, while competitor product pages account for 24.4%. Together, those two page types make up 91.9% of citations. The data covers 45,309 citations across 7,260 AI answers, so this is a large sample, though it is still a sample.

That changes where you look first.

Find the exact page

Work backwards from the answer your customer saw. Find the wording of the buyer question if you can, then inspect the source links attached to ChatGPT's response. You are looking for the exact page, not merely the domain. A recommendation from a comparison article tells you something different from a recommendation drawn from your product page.

Start with the URL.

Read the surrounding passage. Does it name your business in a shortlist, compare it with alternatives, describe a use case, or repeat a customer view? The answer might have compressed the source, so the page will not necessarily contain the same sentence. It should still reveal the role your business played in the source.

The cited page gives you something to inspect.

A brand mention and a citation are not the same thing. ChatGPT may know the name of a business from several sources, then cite one page because that page supports the particular claim in the answer. A brand mention does not tell you which page earned the link.

(CiteGraph's visibility checker logs the answer, cited page and domain, if screenshots get old.)

Why that page got the job

A page is useful to an answer engine when it gives the system something specific to extract. A roundup may already organise products around the buyer's decision. A competitor page may state who a product is for, what it does and how it differs from nearby options. Those formats reduce the amount of interpretation required.

Specific claims travel further.

That does not mean the page is universally better, or that it has won a permanent place in some hidden catalogue. It means the page fitted the question and the available evidence at that moment. A buyer asking for tools for a particular team has created a different sourcing task from a buyer asking which products integrate with a named platform.

The query sets the shape of the evidence.

Consider a simple before and after. Before: “We help teams work better together.” One paragraph, one demo button. After: “Built for support teams moving off spreadsheets. No Salesforce integration yet.” The second version gives a buyer, a use case and a limitation in words an answer system can reuse.

The page became easier to use as evidence.

Nothing magical happened to the domain. A human can infer plenty from the first version. An answer system has less reason to infer when the second version puts the relevant facts in plain view. ChatGPT may recommend you because another page gave the system a clear, usable fact about your business for that particular question.

The page may be yours, or someone else's

Your own site is only one possible source. In the recorded sample, the scanned product's own site accounted for 1.5% of citations. Community threads accounted for 3.4%, documentation for 1.1%, review platforms for 0.8% and social posts for 0.8%. The percentages are small beside roundups and competitor pages, but they describe different routes into an answer.

Your site is one route among several, and the route that produced the recommendation may sit elsewhere.

A customer saying “ChatGPT recommended us” may therefore be reporting a result that your site did not directly produce. A comparison article could have included you. A forum discussion could have supplied context. A product page on another domain could have made the recommendation credible enough to cite.

The cited page may sit elsewhere.

This is why chasing a single mention can waste time. If the cited page describes your product accurately, the useful work may be to understand the buyer question and the evidence around it. If the page is thin, outdated or wrong, the problem is less about persuading ChatGPT and more about making the facts available on pages that other sites can check.

The methodology explains how we compare these scans. The selection still happens per question.

The domain can stay while the page changes

This is the part that surprises founders who hear about the recommendation from a customer. You may revisit the result later, find the same business still present, and assume the system has a stable relationship with your domain. The exact page may already have changed.

A familiar domain does not settle the matter. The page behind it can already be gone.

In our own project, we compared scans 17 days apart. Of 40 domains cited three or more times in the first scan, 34 were still cited later. But only 41 of the 179 exact pages cited in the first scan appeared again. The domain often survived. The URL usually did not.

The figure below compares those three forms of persistence.

In CiteGraph's scan, 85% of heavily cited domains, 68% of all cited domains and 23% of exact pages persisted after 17 days. That is why a recommendation can feel real and still be difficult to reproduce. The business remains in the wider source set while the page supporting a particular answer rotates.

A later scan can change the result without changing the business.

Counting every cited domain, 69 of 102 came back in the later scan, and our methodology says the long tail is mostly sampling rather than change. A page disappearing from one answer does not prove that ChatGPT rejected your business. It may reflect a different question, retrieval result or sample.

The caveat that matters most

The evidence can show what was cited, not the private reasoning that selected it. We can inspect the question, answer, domain and URL in a measured sample. We cannot know every page the system considered, or whether a recommendation will appear for a different buyer, account or date.

That limit changes how you interpret the result.

Treat the customer's report as a lead to investigate, not as proof that your site has secured a lasting position. Trace the cited page. Record the claim it supported. Check whether your own pages state the same facts clearly enough for the next question.

The most useful explanation is usually concrete: this page was cited, it supplied this part of the answer, and another page may replace it later.

We can show what was cited. We cannot show why it was chosen over everything else.