Two different events, two different causes
An AI answer can do two things with your brand, and they happen for different reasons. It can name you: your brand appears in the text. It can cite you: one of your URLs is attached as a source. A name comes from the model's memory or from a third-party page it read. A citation means your own page was in the retrieved set. You can have either without the other.
| State | What the reader sees | What produced it |
|---|---|---|
| Named and cited | Your brand in the sentence, your URL in the sources | Your page was retrieved and the model treated the brand as the answer |
| Cited, not named | Your URL in the sources, brand absent from the text | Your page was read as evidence for a claim; the brand itself was not the point |
| Named, not cited | Your brand in the text, no link to you | Memory, or a third-party page that mentions you |
| Neither | Nothing | Not retrieved, not remembered for this job |
The second row is the one that breaks most reporting. Semrush's June 2026 study of 3,981 domain appearances across four engines found that 61.7% of citations were ghosts: the URL was linked and the brand was never said. If your report counts "citations" as wins, most of those wins are a page doing background work for somebody else's answer.
Why one score hides the thing you need to know
Tools like to collapse named and cited into a single visibility number. That number goes up when your blog post is cited as background on a question about the category, and it goes up when your brand is recommended on a buying question. Those are not the same outcome, and leadership will make different decisions depending on which one moved. Keep two columns. If you must show one headline figure, show named on buying prompts, and show cited beside it.
An in-house lead at a payments company put it to us as a question during discovery in August: "citations or brand mentions, which one am I supposed to be counting?" The answer is both, on the same prompt set, without averaging them.
How to count it, by hand, this week
- Fix a prompt set. Eight to sixteen questions your buyers type, in their words. Freeze the wording.
- Run each prompt in ChatGPT with search on, and in Perplexity. Paste the answer and the source list into a sheet.
- For each prompt and engine, mark three cells: named (yes/no), cited (yes/no, and which URL), competitor named (which).
- A week later, run the identical prompts again. Same wording, same engines. Mark the same cells.
- Report the two counts side by side: named on N of M prompts, cited on N of M prompts, per engine, with both dates.
| Prompt | Engine | Named (wk 1 → wk 2) | Cited URL (wk 1 → wk 2) | Competitor named |
|---|---|---|---|---|
| tool to see why ChatGPT recommends a competitor | ChatGPT | no → yes | none → /guides/why-chatgpt-names-a-competitor | two others → one other |
| how to measure brand mentions in ChatGPT | Perplexity | no → no | none → none | same two |
Reading the result
- Cited went up, named did not. Your page entered the retrieved set as evidence. Good sign for retrieval, no change in the recommendation. Check what the page says in its first paragraph; it may be answering a neighbouring question.
- Named went up, cited did not. A third-party page or the model's memory is carrying you. Fragile. Find which page is cited on that prompt and ask whether you can be on it.
- Both went up on the same prompt. That is the outcome. Record the date and the exact prompt, and keep running it. A single week of both is an observation, not a trend.
- Neither moved. Normal for the first week after a change. Engines refresh on their own schedule. Give it a second and third run before concluding anything.