# Citation Rate in a Panel: What It Measures and What It Misses

> What citation rate is in a query panel, how to build a repeatable panel for ChatGPT or Gemini, and what that number cannot tell you.

URL: https://clione.ai/glossary/citation-rate/
Lang: en

When an agency or a tool shows you a percentage of "answers where you get cited," it is talking about a citation rate, and almost always a measure over a panel: a list of queries somebody chose, run on specific assistants, a specific number of times, on specific days. It has nothing to do with the citation index of scientific journals, or with the count of citations a search engine gives you for your own site. **It is the only way to measure your presence in assistants that publish no data about your site**, and that is why it is worth understanding: it is good for tracking a trend, as long as you know exactly what was counted.

## How it works

A citation rate in a panel is a proportion, and **every part of that proportion is a decision by whoever measures**. There is no standard that says which; what there is is a list of decisions that have to be written down before the first query is run:

- **The queries.** The literal text of each one, in what language, from what country and why those and not others. Once written, they are not touched: if they change, a new series starts.
- **The assistants and their modes.** ChatGPT with search, [Gemini](/glossary/gemini/), [Perplexity](/glossary/perplexity/) or Google's [AI Mode](/glossary/google-ai-mode/) are not added together: each one is counted separately.
- **The repetitions and the dates.** How many times each query is run and on which days.
- **What counts as a citation.** A link to your domain in the text, one in the sources panel, only the [product page](/glossary/product-page/) you care about, or any page on your site. An [AI citation](/glossary/ai-citation/) is not the same as a mention of your brand without a link, and a panel that mixes them measures something else.
- **What is divided by what.** Answers with at least one citation of you out of total answers, or queries where you appear at least once out of the panel's queries. Both are reasonable and give different figures on the same data.

**Repetitions are not a detail.** The paper that gave [GEO (Generative Engine Optimization)](/glossary/geo/) its name handled it this way in its own experiment: it generated five different answers for each query, at temperature 0.7, to reduce statistical deviations, and repeated the experiments with five random seeds to report the average.[^geo-paper] To do it they had to build their own query benchmark, GEO-bench, because no public set of queries designed for [generative engines](/glossary/generative-engine/) existed.[^geo-paper] A panel is the same thing at the scale of a store: a query set of your own, repeated.

**The session counts too.** OpenAI explains that ChatGPT can search the web automatically when the question calls for it, that it usually rewrites your query into one or several searches that it sends to its providers, that it uses an approximate location based on your IP address and that, **if memory is on, it may use saved memories when rewriting the search**.[^openai-search] To avoid that, a temporary chat can be opened as "not personalized," with no memories or custom instructions, and temporary chats do not create or update memories.[^openai-memory] In the Gemini API, the model decides whether a Google search can improve the answer, and each citation links a piece of the text to the source URL.[^gemini-grounding] So an answer without a search has no sources to count, and a good panel notes whether the assistant searched or not in each run.

**Against a panel, first-hand data.** Since February 10, 2026, [Bing Webmaster Tools](/glossary/bing-webmaster-tools/) has a public preview report, AI Performance, that counts your site's citations in [Copilot](/glossary/copilot/), in Bing's AI-generated summaries and in some partner integrations, with the citations for each page and the [grounding](/glossary/grounding/) queries that brought them in.[^bing-ai-performance] It is one engine's count on your site, not a sample. The panel covers what that report does not give: the other assistants and, if you design it that way, your competitors.

## What a panel cannot tell you

A well-measured citation rate answers one narrow question: on these queries, in these assistants, on these dates, how many times were you cited? Everything else is outside it.

- **How many people ask those questions.** The queries are chosen by whoever measures, and no assistant publishes how many times each thing is asked. Google says it about its own systems: **no third-party tool has access to its internal ranking or AI systems**.[^google-ai-guide]
- **How much you weigh in the answer.** Bing warns that its citation counts do not indicate the ranking, authority or role of a page within a particular answer.[^bing-ai-performance] A panel that only marks "cited" or "not cited" has the same limit.
- **Whether the citation was right.** OpenAI warns that the results and citations of its search may be incomplete, out of date or incorrect.[^openai-search] Counting a citation does not say whether the sentence it backs describes your product well.
- **Visits or sales.** A citation may bring a click or may not. What reaches the store is read in analytics, and what ends up as an order after passing through an assistant is [assisted conversion](/glossary/assisted-conversion/).
- **Why it changed.** If the rate goes up after you rewrite your product pages, it may be because of your pages or because the assistant changed. A panel shows correlation; cause needs something more, like a group of pages you did not touch.

## In a store

A backpack store freezes its panel before starting. It is a **fictional example**: the store, the queries and the URLs are invented to show what has to be written down.

| Panel sheet (fictional example) | What is written down before the first query                                           |
| ------------------------------- | ------------------------------------------------------------------------------------- |
| Queries                         | Ten purchase questions in English, without the store's brand, with their literal text |
| Assistants                      | ChatGPT with search and Gemini, each counted separately                               |
| Session                         | Temporary, non-personalized chat, signed out where possible, from the United States   |
| Repetitions and dates           | Three times each query, on three days of the same week                                |
| What counts as a citation       | A link to any example-store.com URL, in the text or in the sources, noted separately  |
| What else is noted              | Whether the assistant searched, all cited URLs in order and a screenshot              |


And this is what a single query from the panel, "waterproof backpack for a 15-inch laptop," looks like across its three runs in the same assistant:

| Run (fictional example) | Searched? | URLs cited, in order                                                       | Cites the store?    |
| ----------------------- | --------- | -------------------------------------------------------------------------- | ------------------- |
| Monday                  | Yes       | example-magazine.com, example-store.com/urban-backpack-20, other-store.com | Yes, in the sources |
| Wednesday               | Yes       | example-magazine.com, other-store.com                                      | No                  |
| Friday                  | No        | None                                                                       | No sources          |


Three runs, three results. **A Monday screenshot would have said "they cite us"; Wednesday's, "they don't cite us,"** and both would be true. The third row is the one most often forgotten: an answer without a search cannot cite anyone, and whoever counts it as a "no" and whoever takes it out of the total get different figures from the same data. That is why the panel sheet is published alongside the rate.


**Measuring from your own account and reading it as what your customers see.** If you run the panel signed in with memory on, ChatGPT may use what it remembers about you when rewriting the search and your approximate location to refine it.[^openai-search] Someone who has been asking about your store for months may see your site cited more often than a new shopper, and the rate then measures your conversations, not your market's. A temporary non-personalized chat takes memories and custom instructions out of the equation.[^openai-memory]


## How to check it

1. **Write the panel sheet before running anything**: literal queries, assistants, session, repetitions, dates and what counts as a citation. Save it with a date.
2. **Run each query in a temporary non-personalized chat**, from the country you care about, and note in each run whether the assistant searched.[^openai-memory][^openai-search]
3. **Write down all the cited URLs, not just yours**, in the order they appear and with a screenshot. Without everyone else's you will not be able to compare later.
4. **Calculate the rate with the written definition and publish it alongside the sheet.** If you change a query or an assistant, start a new series instead of continuing the old one.
5. **Cross-check against first-hand data**: the per-page citations in AI Performance in Bing Webmaster Tools.[^bing-ai-performance] If your panel goes up and that count does not move, review the panel before celebrating.

## Related concept: AI share of voice

[AI share of voice](/glossary/ai-share-of-voice/) starts from the same data and sets it against your competitors: of all the times a brand is cited or named in the answers, how much is yours. Citation rate in a panel looks only at you: in how many answers you appear as a source. The two can move in opposite directions: if answers start citing more sources, your rate can go up while your share of the total goes down. And both depend on the panel: **without the list of queries, neither figure can be checked**. [AI visibility](/glossary/ai-visibility/) is the term that groups them, impressions and mentions included.

To keep going, the [measurement and verification reading path](/learn/measurement-and-verification/) orders the rest of the terms, and the [catalog audit](/glossary/catalog-audit/) explains how to review the data from your store that those answers use.