A person seen from behind, chin in hand, studying a document on a large monitor with two laptops open on the desk.

Photograph by Olia Danilevich via Pexels. Illustrative image, not a customer or endorsement.

Blog · · 7 min read · NEXUM Team

Cited but Not Named: Why ChatGPT Can Use Your Pages and Still Not Mention Your Brand

ChatGPT can pull a page of yours, lean on it in the answer, and still never say your name. If you are tracking “AI visibility” as a single number, that gap is easy to miss — and easy to misread as a technical problem you can fix with markup.

It is usually a different problem. A citation is when an answer points to a URL as a source. A mention is when the answer names your brand, product or publication. STAT’s Tom Capper put it simply in a Search Engine Journal recap: citations have URLs, but mentions have entities. For most brands, being named and recommended is the outcome that matters. Citations help explain why.

This article is about that split, not about a trick to force a model to say your name. Nobody can guarantee a mention. You can measure the gap, and you can work on the things that actually correlate with being named.

Why the two metrics diverge

Retrieval and attribution are not one pipeline with a weak final step.

Retrieval is about whether a page looks useful for a question. Ahrefs’ study of 1.4 million ChatGPT 5.2 prompts describes a gate that sits before the model opens your page: title, snippet and URL do a lot of the work in deciding what gets fetched. Pages whose titles and retrieval data line up with the question (and with the sub-queries ChatGPT runs) are more likely to be cited. That is a good argument for clear titles and opening paragraphs. It is not, on its own, an argument that the answer will then credit your brand.

Attribution is closer to how the model writes. It will often name a source when the name already appears in material it trusts, when the claim is distinctive enough that a name adds credibility, or when the brand is already a known entity. A three-month-old site can supply useful pages and still be treated as anonymous source material. That pattern shows up often enough in public discussions that it should be in your measurement plan, not treated as an anomaly.

The two numbers will not move together. You can publish more citable pages and watch citations rise while named mentions stay flat. You can also be recommended in an answer that cites a review site, a retailer or a journalist — and not your own URL. Both happen.

A person looking at a laptop in a dim room, face lit by the screen, expression serious and unsmiling.

The answer is on the screen. The name may not be.

Photograph by Nicolás Langellotti via Pexels. Illustrative image, not a customer or endorsement.

What this is not

It is not proof that your schema is “broken.” Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared them with thousands of control pages. Adding schema did not produce a meaningful citation lift on Google AI Overviews, AI Mode or ChatGPT. Valid Organization or Article markup is still good hygiene. It is a weak explanation for a 100-to-1 gap between cited URLs and named mentions.

It is also not proof that you should chase freshness for its own sake. In the same 1.4 million-prompt analysis, the average cited page was about 500 days old. Some retrieval paths prefer recent documents. That is not a general rule that a new blog post will outrank a clear, useful one.

And it is not something you can close with an llms.txt file. Google’s guidance on generative AI features has been explicit about not treating files like that, or extra schema, as the lever. The more useful split in that guidance is commodity versus non-commodity content: if a model can write the same piece without you, it has little reason to cite you — and even less reason to treat you as a named authority.

We have written in more detail about what actually helps AI crawlers in 2026 and about how GEO relates to SEO. The short version is the same: technical access is a baseline; it does not buy you a byline in the answer.

A simple way to measure the gap

If you already measure AI share of voice, add two columns. Do not average them.

For a fixed set of prompts, engines and dates, log:

  1. Mentioned: did the answer name your brand or product?
  2. Cited: did it point to one of your URLs?
  3. Named someone else: which competitors or publishers appeared?
  4. Cited someone else: which third-party URLs appeared when you were (or were not) named?

Run each prompt more than once. Answers move. A single screenshot is not a baseline.

Then look at the shape of the results, not just the totals:

  • High citations, low mentions. Your pages are useful enough to retrieve. The model is not treating you as the entity to recommend or credit. Off-site name use and distinctive claims usually matter more here than another pass at the About page.
  • High mentions, low citations. You are being recommended, but the evidence in the answer often comes from other people’s pages. That is not a failure. It tells you which sources to earn or to keep accurate.
  • Low both. You have a retrieval and an entity problem. Start with crawl access, indexation, and whether you have a page that actually answers the question in the title and the first paragraph. See how to get cited by ChatGPT, Perplexity and Gemini for that baseline.

If almost all of your citations come from a handful of URLs, that concentration is useful. It tells you what these systems already find quotable. It is a better brief than “we need more content” in the abstract.

What is worth doing first

Work in this order. It is slower than a checklist of GEO tactics, and more honest.

Make sure the engines can reach a clear page. If important URLs are blocked, unindexable, or only exist after JavaScript runs, you are arguing about attribution before you have retrieval. This includes obvious robots mistakes and less obvious ones, like a canonical that points a product page back to the homepage.

Put the answer where retrieval can see it. Titles, opening paragraphs and URLs should sound like the question a buyer would ask. If the useful sentence is in paragraph seven, you are making it easy for another page to win the first cut. Ahrefs’ citation work is about that gate.

Put something on the page a model cannot already say. A number you measured, a limitation you will admit, a comparison a buyer can use, a method another writer cannot invent. That is the non-commodity test in Google’s AI-features guidance.

Look off-site for the name, not only on-site for the URL. If the only place your name appears with the claim is your own domain, the model has little reason to say “according to [you].” Reviews, reputable articles, and other people using the name in context are the signals most often associated with being mentioned as an entity. They are also slower to earn, which is why this gap frustrates teams that have already done a lot of on-site SEO.

Do not change five things at once and then declare a winner. If you want to know whether a change mattered, keep the prompt set fixed and give it time. Citation volume is volatile. Ahrefs has also documented how unstable AI citations can be.

A woman looking intently at a computer monitor in an office, with a concentrated, unsmiling expression.

Reading the answer closely is how the gap becomes visible.

Photograph by Mizuno K via Pexels. Illustrative image, not a customer or endorsement.

What we will not claim

We will not tell you that a particular ratio of citations to mentions is “normal.” We will not tell you that an About page, an author bio or a publisher schema will turn citations into names. And we will not tell you that a mention can be guaranteed if you ship the right file.

What you can do is stop treating “we were cited” as the same event as “we were recommended.” Once those are two metrics, the work gets clearer: some weeks you are building pages that can be retrieved; some weeks you are earning the name that makes a retrieval worth having.

If you want a starting point, request a free NEXUM audit. We will show you where you are mentioned or cited in AI answers and a first Fix Queue. We will not promise that the next answer will include your name.

FAQ

Is a citation worthless if the brand is not named?

No. A citation can still support the facts in the answer and send a click. It is simply not the same as being recommended or attributed. Track both.

Can I fix the gap with schema or llms.txt?

Those are baseline technical signals, not a naming strategy. Public tests and Google’s own guidance do not support treating them as the main lever.

Should I add my brand name next to every statistic?

Stating a distinctive claim with a clear source can help a reader — and sometimes a model — attribute it. Repeating the brand name on every line is not a substitute for being named off-site, and it can make the page worse for humans.

Do I need a tool to measure this?

No. A spreadsheet, a fixed prompt list and a regular review are enough to see the split. A tool helps when you want the same panel run across several engines, over time, without doing it by hand.