ChatGPT citations are not the same signal as Google rankings

TL;DR: If you are tracking AI search visibility, treat Google rankings and OpenAI API outputs as directional signals, not a clean proxy for the pages ChatGPT actually cites.

Can Google results predict ChatGPT citations?

Matt G. Southern reported in Search Engine Journal, in “OpenAI’s API Only Partly Matched ChatGPT’s Citations, Paper Finds,” that a paper tested whether Google results and OpenAI’s API outputs can stand in for ChatGPT’s cited sources. The short version: they matched cited sites more often than they matched exact cited pages.

That distinction matters.

A domain-level match says, “ChatGPT and Google both seem to trust this publisher.” A page-level match says, “This exact URL is the answer source.” Those are very different signals if you run SEO, content, or reputation programs.

If ChatGPT cites example.com/specific-guide, and Google ranks example.com/category-page, the brand may still be visible, but the measurement is off. You cannot use that as proof that your optimized page is the one AI systems prefer. You can only say the domain is in the neighborhood.

I run Lucky Domains, which works on SEO and search visibility at and this is exactly the kind of measurement gap that gets teams in trouble: they compress “brand appears somewhere” into “our content strategy is working.”

Those are not the same thing.

Why does page-level mismatch matter?

SEO has always had proxy metrics. Rankings. Impressions. Backlinks. Referrers. Share of voice. None are perfect, but they have known failure modes.

AI citations add a newer failure mode: the assistant can cite a source because it is useful for a generated answer, not because it maps cleanly to a search result page. The answer may be built around a query rewrite, a live retrieval step, a ranking layer, a safety layer, or some product-specific citation behavior we cannot fully inspect.

That means “we rank on Google” is useful context, not a final answer. Same for “the OpenAI API returned something similar.” Southern’s report says the paper found stronger overlap at the site level than at the exact page level. That is the practical takeaway.

For marketers, this weakens a lot of AI visibility dashboards if they imply precision they do not have. If the tool says your domain is likely to be cited, fine. If it says a specific URL will be cited by ChatGPT because it ranks well or appears in an API response, ask for receipts.

The better metric is probably a bundle: whether the domain appears, whether the exact page appears, whether the answer accurately describes the entity, and whether the cited page is the page you would actually want users to land on.

Where does the OpenAI API fit?

The API is still useful. Just not as a perfect replay of ChatGPT.

Product surfaces differ. ChatGPT has its own interface, citation behavior, retrieval stack, and user context. The API is a developer interface. It can help test prompts, compare phrasing, inspect answer tendencies, and run repeatable experiments. But if the question is “what will ChatGPT cite in the consumer product,” the API should be treated as a lab instrument, not the environment itself.

Same with Google. Google remains a huge discovery and authority signal. If a site consistently ranks for a topic, I would expect it to have a better shot at appearing in AI answers than a thin site with no visibility. But expectation is not attribution.

The catch is that teams want one number. “AI visibility score.” “Citation share.” “Answer engine rank.” Those can be helpful if they are labeled honestly. They become misleading when they hide the difference between domain-level and URL-level matching.

Practitioner’s take: run your own small audit before buying the dashboard story. Pick 25 high-intent queries, capture actual ChatGPT citations in the product, compare them against Google results and API outputs, then separate domain matches from exact URL matches. If your domain shows up but the wrong pages get cited, the work is not just “rank higher.” It is to build clearer, more citable pages around the exact questions assistants are answering.