The Prompt Panel Method: How to Track AI Overviews Visibility and Share of Voice — OnyxRank

Sep 26, 2026 ·OnyxRank Team

Most businesses think they can check their AI Overviews visibility by typing a few queries into Google and seeing what shows up. They're wrong. One search gives you one sample from a system that changes its answer by session, location, and day, so a single check tells you almost nothing. The reliable way to measure AI visibility is a prompt panel: a fixed list of queries, run on a schedule, scored the same way every time. OnyxRank uses this method to report GEO optimization progress to clients, and this guide shows you how to build your own version in an afternoon.

If you have already read how to [measure GEO performance](/blog/geo-performance-metrics-2026) or how to [find AI Overview traffic in GA4](/blog/measuring-ai-overviews-traffic-ga4-2026), this is the missing piece between them. Those posts cover what to measure and where the traffic lands. This one covers how to collect the citation data itself without fooling yourself.

What a Prompt Panel Is and Why It Beats Spot Checks

A prompt panel is a fixed set of 30 to 60 queries that represent how your buyers search, checked on a regular schedule, with every result recorded in the same format. Three properties make it work:

1. **Fixed.** The same queries every month, so changes in your numbers reflect changes in visibility, not changes in what you asked.

2. **Representative.** Queries mapped to real buyer intent, not the ones you happen to rank for.

3. **Scored.** Every result gets a defined status, so two people checking the same query record the same thing.

Spot checks fail on all three. You tend to check queries where you expect to win, you forget what you saw last time, and you describe results loosely ("we showed up, I think"). A panel removes that drift.

Step 1: Build the Panel From Real Demand

Pull queries from four sources rather than brainstorming them:

**Search Console.** Export queries with impressions but low click through rates on pages that should be winning. These are often the queries where an AI Overview is absorbing the click.
**Sales and support conversations.** The exact phrases prospects use on calls. These skew toward long, conversational questions, which is what triggers AI answers.
**Competitor pages.** Headings and FAQ questions on the pages that outrank you.
**People Also Ask boxes.** Follow-up questions Google already associates with your topic.

Then sort every query into one of three groups, and aim for a rough split:

Group	Share of panel	Example (accounting software)

|---|---|---|

Category and how-to	50%	"how to automate invoice reminders"
Comparison and evaluation	30%	"best invoicing software for freelancers"
Brand and branded comparison	20%	"YourBrand vs competitor pricing"

The split matters because each group tells you something different. Category queries show whether you are part of the conversation. Comparison queries show whether you are on the shortlist. Brand queries show whether AI systems describe you accurately.

Keep the panel to queries you would genuinely want to be cited for. A panel of 200 vague queries produces noise. Forty specific ones produce a trend line.

Step 2: Record Three States, Not Two

The most common tracking mistake is a yes or no column: "Are we in the AI Overview?" That collapses two very different outcomes. Use three states instead:

**Cited.** Your domain appears as a linked source in the AI Overview.
**Mentioned.** Your brand is named in the answer text but your page is not linked.
**Absent.** An AI Overview appeared and you are neither cited nor mentioned.

Add a fourth code for queries that returned no AI Overview at all. Those rows do not count against you, and they teach you which of your queries are even eligible. If 35 percent of your panel never triggers an overview, that is a targeting insight, not a failure.

For every cited or mentioned result, also log three details: your position among the cited sources, which page was cited, and whether the surrounding claim is accurate. Inaccurate mentions are an underrated problem. If an AI answer states your pricing model wrong, that is a content fix, and only a panel will catch it.

Step 3: Control the Variance

AI answers are probabilistic. The same query can cite different sources an hour apart. Three rules keep your data honest:

**Run each query more than once.** Check each prompt on at least three separate days in the reporting month and record the most common state. If a query shows "Cited" once and "Absent" twice, its state is Absent with a note that it is unstable.

**Standardize the environment.** Use the same location setting, a logged out or clean browser profile, and the same language. Local businesses should run panels from within the service area, because AI Overviews vary by geography.

**Never compare a Tuesday check to a Friday check as though the gap were a trend.** Compare monthly aggregates, and treat a single week of movement as noise unless it persists.

Model and interface updates also move every number at once. When Google changes how overviews are displayed, your whole panel can shift with no action on your part. Keep a dated change log next to the panel so you can explain a step change instead of taking credit or blame for it.

Step 4: Calculate Share of Voice Correctly

Share of voice is where panels become useful for decisions. The formula is simple:

**Citation share = your citations divided by total citations across all brands, across the panel.**

Suppose your panel has 40 queries and 32 trigger an AI Overview. Across those 32 overviews, sources are cited 110 times in total. Your domain is cited 14 times. Your citation share is 14 divided by 110, or about 13 percent. A competitor cited 27 times holds 25 percent.

Two rules prevent misleading numbers:

**Always show the denominator.** "Share rose from 12 to 18 percent" means nothing without the panel size and the number of overviews triggered. A jump from 3 of 25 to 5 of 28 is a small sample and should be labeled that way.
**Report by group.** A blended 13 percent can hide a 30 percent share on category queries and a zero on comparison queries. The second number is the one costing you revenue.

To tally competitor citations you do not need any special tool. Every time you record an overview, list the cited domains in a column. After a month, a pivot table gives you the leaderboard.

Step 5: Turn the Data Into Decisions

A panel is worth building only if it changes what you do. Here is how to read the results:

**High Absent rate on category queries.** Your foundational content is not being selected. Audit for direct answer formatting, entity clarity, and freshness. Our [GEO priority matrix](/blog/geo-priority-matrix-ai-overviews-optimization-framework-2026) helps rank which pages to fix first.

**Mentioned but rarely Cited.** AI systems know your brand but do not trust a specific page enough to link it. Strengthen the page that should answer the query with original data, a named author, and clear structure.

**Cited but with inaccurate claims.** Correct the source page and add a plain, quotable statement of the fact that is wrong.

**A competitor dominating comparison queries.** Run a [citation gap audit](/blog/citation-gap-audit-ai-overviews-competitors-2026) on those exact queries to see which sources they are getting cited from.

**Unstable states.** Queries that flip between Cited and Absent are your best near term targets. They sit at the margin, and modest improvements often tip them.

A Monthly Reporting Template You Can Copy

Keep the report to one page:

1. Panel size and number of queries that triggered an AI Overview

2. Cited, Mentioned, and Absent counts, with change from last month

3. Citation share overall and by query group, with the denominator shown

4. Top three gains and top three losses, with the page involved

5. Change log of known platform updates during the month

6. Two actions for next month, tied to specific pages

Review it monthly and judge direction over a full quarter. If you want a baseline before you build your own, [try our free SEO audit](/free-audit) and we will show you which of your pages are currently eligible for citation.

Frequently Asked Questions

How many queries should an AI Overviews tracking panel include?

Start with 30 to 60 queries. Fewer than 30 gives you too few AI Overviews to calculate a stable share, and more than 100 becomes hard to check manually and dilutes focus. Make sure every query reflects real buyer intent.

How often should I check my AI Overviews visibility?

Check each query several times during the month, then report monthly. Weekly reporting exaggerates normal variation. Act on trends across at least a quarter, and keep a change log so you can separate your progress from platform updates.

What is AI share of voice?

AI share of voice is the percentage of citations in AI generated answers that go to your domain compared with all cited domains across a fixed set of queries. It works like search share of voice but is measured by citations rather than rankings.

Can I track AI Overviews without paid tools?

Yes. A spreadsheet, a consistent browser setup, and discipline are enough for a panel of 40 queries. Paid tools save time by automating checks across platforms, which becomes worthwhile when you need weekly data across many queries or several brands. See our comparison of [AI SEO tools versus an AI SEO service](/blog/ai-seo-tools-vs-ai-seo-service-2026) for when that upgrade makes sense.

Why do my AI Overview results change from day to day?

AI answers are generated probabilistically, and the sources selected can vary by location, session, and model update. That is why a panel repeats each query and records the most common result rather than treating any single check as fact.

Key Takeaways

A single search is a sample of one. A fixed panel of 30 to 60 queries is a measurement.
Score every result as Cited, Mentioned, Absent, or No Overview, and log accuracy for brand mentions.
Repeat checks, standardize location, and keep a change log to control variance.
Share of voice needs a denominator, and it should be reported by query group.
The point of the panel is action: each month should produce two concrete page level fixes.

OnyxRank builds and runs these panels as part of our GEO optimization work, alongside the fixes that move the numbers. If you would rather have the tracking and the optimization handled together, see our [pricing plans](/pricing), or start with the [free SEO audit](/free-audit) to get your baseline.