How to Run a Prompt Panel for AI Citations

When a client asks whether they show up in AI answers, the honest reply is that nobody has a dashboard for it yet. The engines do not report citations back to site owners. Referral traffic tells you something, but only after the fact and only when someone clicks.

So we measure it ourselves. A prompt panel is a list of real buyer questions, run across the main answer engines on a schedule, with the sources each engine cites recorded in a spreadsheet. It is manual, it is a little tedious, and it is the only method we have found that shows who gets cited for a category and how that changes over time. This guide is the version we use in client work. You can run it in an afternoon and repeat it in an hour a month.

Why you need a panel at all

Rankings gave us a habit: type a query, see a position, report it. Answer engines broke the habit. There is no position to see. A question gets one composed answer, the sources behind it may or may not be shown, and two people asking the same thing can get different answers on the same day.

The tools that promise to track this for you are early and mostly opaque about how they sample. We use a couple of them for breadth, but we would not put their numbers in front of a board without something we ran ourselves. A panel is that something. It is small, it is repeatable, and every line in it can be shown to a client who asks where the number came from.

The other reason is more practical. The panel produces a list of the pages the engines already cite for your questions. That list is the starting point for the authority work described in How to Get Cited by AI Answer Engines. Without it you are guessing which publications and directories matter.

Step 1: build the question set

Aim for twenty to forty questions. Fewer and the sample is too noisy to read. More and nobody runs it a second time. Write them the way buyers actually type or say them, not the way your keyword tool phrases them. Include the awkward ones, the comparisons, and the questions where you are not the obvious answer yet.

Step 2: choose the surfaces

Run every question on the surfaces your buyers actually use. For most B2B companies in Europe and North America that means five: Google AI Overviews (and AI Mode where it is available), ChatGPT with search on, Perplexity, Microsoft Copilot and Gemini. Add others if your market uses them. Drop any where nobody in your customer base would look.

Each surface behaves a little differently, and it helps to know how before you read the results.

Step 3: run it the same way every time

The value of a panel is in the comparison from month to month, so the conditions have to stay fixed. Small changes in how you ask can change who gets cited.

• Use a clean browser profile or a logged-out session where the surface allows it, and the same location setting each time. Personal history changes answers.
• Ask each question with exactly the same wording. Keep the wording in the spreadsheet and copy it in.
• Ask each question two or three times and record every cited source across the runs. Answer engines are not deterministic; a single run will mislead you.
• Run the whole panel on the same day of the month, or the same week if it is large. Do not spread it over a fortnight.
• Record what you see, not what you expected. If an engine gives no sources, write ‘none shown’. If it gives a wrong answer that cites you, write that down too.

Step 4: record the results

A spreadsheet is enough. One row per question, per surface, per run. The columns below are the ones we have found we actually use; add more if you like, but do not drop any of these.

Keep the raw log forever. The summary numbers change, but the raw list of sources is what you will go back to when you plan the next quarter.

Step 5: the numbers that matter

Resist the urge to build a dashboard with fifteen charts. Three numbers carry almost all of the meaning, and they fit on one slide.

Report presence and share by stage as well as in total. It is common to be cited for evaluation questions about your own product and invisible for problem and category questions. That gap is where the content work goes.

Step 6: turn the list into work

The panel is only useful if something changes because of it. Three kinds of work usually come out of a run.

1. 01

   Fix the questions where you should be cited and are not

   You have a page that answers the question, and the engines cite someone else. Usually the page answers it late, vaguely, or without anything verifiable. Restructure it answer-first, add the specifics, and check the crawler access.

2. 02

   Write for the questions where nobody good is cited

   When the cited sources are thin forum threads or a competitor’s marketing page, the question is open. One clear, sourced page can take it.

3. 03

   Get onto the sources that dominate

   The top ten domains from source concentration are your authority list. Claim and correct your listings, contribute where contributions are welcome, and offer the publications something worth citing. This is slow, and it is the part that moves brand and category questions.

Then run the panel again next month, and compare. Give the work a quarter before judging it. Retrieval changes can show within weeks; authority changes take longer because they depend on other people.

Mistakes we have made so you do not have to

• Running the panel signed in on a personal account. The engine had learned what we wanted to see.
• Changing the wording of a question between months because the old wording felt clumsy. The comparison was lost.
• Counting a mention as a citation. If there is no link, the visitor cannot click and the engine did not verify you. Track both, but keep them apart.
• Only asking questions we already ranked for. The panel looked great and told us nothing.
• Reporting a change after one run. The next run reversed it.

The bottom line

• Build twenty to forty real buyer questions across the whole journey, including the ones you do not win yet.
• Run them on the surfaces your buyers use, in a clean session, with the same wording, two or three times each, on the same day every month.
• Log every cited source. The raw list is the asset.
• Report presence rate, citation share and source concentration, by stage.
• Turn the results into three kinds of work: fix pages that should be cited, write for open questions, and earn a place on the sources that dominate.

Frequently asked questions

How long does a prompt panel take to run?

Setting it up takes an afternoon: writing the questions, building the sheet, running the first pass. After that, a thirty-question panel across five surfaces with two runs each takes about an hour a month if you are disciplined about it.

Can I use a tool instead of doing it by hand?

You can add one, and some are useful for breadth across hundreds of prompts. Keep the manual panel as the reference. Tool samples are opaque, they change without notice, and you cannot show a client the exact answer a tool saw. Use official APIs where an engine offers them; do not scrape the chat products.

What is a good citation share?

There is no benchmark that means anything across categories. Compare yourself to the top cited domains in your own panel and to your own previous runs. Going from zero presence on problem questions to being cited on a few of them is a real result; a category-wide share figure without that context is not.

Why do the answers change every time I ask?

Because generation is probabilistic and retrieval picks slightly different shortlists. That is why the panel asks each question more than once and records the union of sources. Treat single-run differences as noise.

Should the panel be in more than one language?

If you sell in more than one language, yes, and treat each language as its own panel. Citations differ a lot between languages, and the sources that dominate in French or German are rarely the same as in English.

Who at Lucidens runs this for clients?

It is part of the Clarify stage of every engagement that touches answer engines, and then a monthly task inside the Answer Engine Visibility work. The client gets the raw log, not just the summary.

How this guide was put together

• This is the working method Lucidens uses in engagements as of September 2026. The example figures are illustrative and do not describe a named client.
• Surface behaviours are described from use; the providers change their interfaces often, so check the way each shows sources before a run.

Sources

1. Google Search Central: AI features and your website
2. OpenAI: Overview of OpenAI crawlers
3. Perplexity: Perplexity crawlers
4. Lucidens: How to Get Cited by AI Answer Engines
5. Lucidens: Answer-engine visibility, how generative systems choose their sources