Your clients are asking a new question in every SEO report: “How are we doing in AI search?” And most agencies answer with a screenshot and a shrug — a ChatGPT answer here, a Perplexity citation there, no baseline, no trend, no method anyone could repeat.
That stops being acceptable the moment AI visibility becomes a line item in the retainer. Clients do not need you to know the exact number — they need a dashboard: defined metrics, a documented method, and a trend line that moves when the work moves it.
This guide builds one. You will learn which metrics actually describe AI visibility, which data sources feed them, how to assemble a working dashboard in a spreadsheet before spending a dollar on tooling, and how to present it to clients without overselling what the numbers mean.
Key Takeaways
- An AI visibility dashboard answers one question — “are we being seen and cited in AI answers?” — with five metrics: citation share of voice, AI Overview presence rate, engine-by-engine coverage, cited-page distribution, and AI referral traffic.
- The data comes from four places: GA4 (AI referral traffic), Search Console (query discovery), manual prompt testing (the spreadsheet baseline), and dedicated trackers (automation at scale).
- Build the spreadsheet version first. It costs nothing, teaches you what the tools actually measure, and gives you a baseline to judge any paid tool against.
- Report trends, not snapshots: week-over-week direction on a consistent prompt set is the signal; any single number is noise.
- Never present AI visibility metrics as exact. Volatility is structural — say what the method can and cannot prove.
What Should an AI Visibility Dashboard Measure?
Five metrics, each answering a different part of the question. Resist the urge to add more — a dashboard with twenty metrics is a report nobody reads.
1. Citation share of voice. The percentage of AI answers (for questions you care about) that cite your brand. Formula: (answers citing you ÷ AI answers triggered) × 100. This is the headline KPI — the closest thing AI search has to “rankings.”
2. AI Overview presence rate. For your target question set, what percentage trigger a Google AI Overview at all? This tells you how much of your market’s search behavior now flows through AI answers. It is a market metric, not a performance metric — it goes up when Google expands coverage, not when you do good work.
3. Engine-by-engine coverage. Your citation share of voice broken out per AI surface: ChatGPT search, Perplexity, Google AI Overviews, Gemini, Copilot. Engines differ wildly — a brand can be cited in Perplexity and invisible in ChatGPT. One blended number hides that.
4. Cited-page distribution. Which of your URLs actually get cited? This is the diagnostic metric: it tells you what content the engines consider authoritative, and it is usually not the pages you would guess. Shifts here reveal which content investments are paying off.
5. AI referral traffic. The clicks you can actually attribute: sessions from AI referrers (chatgpt.com, perplexity.ai, and friends) in GA4. This undercounts your citations — many citations earn zero clicks — so treat it as a floor on value, not the total. The setup method is covered in my guide to measuring AI traffic in GA4.
Five metrics, one page. If a number does not change a decision, it does not belong on the dashboard.
Where Does the Data Come From?
Four sources, each with a job. The dashboard’s credibility comes from combining them honestly, not from any single one.
GA4 — the click layer. Segment sessions by AI referrers to measure the traffic AI citations actually send you. Free, already in your stack, but it only sees clicks — it knows nothing about the citations that earned no visit.
Search Console — the question layer. Your Performance report is where the target question set comes from: real queries, real impressions, real buyers. It cannot show you AI answers, but it tells you which questions are worth testing. For the full picture of what GSC can and cannot reveal about AI search, see the AI search analytics guide.
Manual prompt testing — the truth layer. A spreadsheet, a clean browser session, and a weekly run of 20–50 buyer questions per engine: does an AI answer appear, are you cited, which domains are cited. This is the method the AI Overview citation tracking guide documents in full. It is slow and it is the only number you fully control.
Dedicated trackers — the scale layer. When the prompt set outgrows manual testing, tools like Otterly.AI, Ahrefs Brand Radar, Semrush, SE Ranking, Peec AI, and Profound automate prompt testing and citation extraction across engines. Capabilities and pricing shift constantly — verify before buying, and run any tool alongside your spreadsheet for a month before trusting its numbers.
The rule: no metric on the dashboard may come from a source you do not understand. If you cannot explain how a number was produced, it does not go in the client report.
How Do You Build the Spreadsheet Version?
Start here even if you plan to buy tooling later. The spreadsheet is the method; everything after it is convenience.
Step 1: Define the question set. Pull 20–50 questions from Search Console, your sales inbox, and People Also Ask — the queries buyers actually ask, favoring specific, high-intent phrasing. Freeze the list. Changing questions mid-quarter destroys comparability.
Step 2: Define the engines. Pick the surfaces that matter to your market: at minimum ChatGPT search, Perplexity, and Google AI Overviews. Add Gemini or Copilot if your audience lives there. Same engines every week.
Step 3: Set the scoring rubric. For each question × engine, record: AI answer present (yes/no), your brand cited (yes/no), cited page URL, competitor brands cited. Four columns. Nothing fancier.
Step 4: Run it weekly, same conditions. Clean browser session, signed out, consistent location, same day each week. Log the date with every run — a number without a date is decoration.
Step 5: Build the five charts. One chart per metric: share-of-voice trend, presence-rate trend, per-engine coverage bars, cited-page table, AI referral sessions trend. Five visualizations, one page. If your spreadsheet tool supports it, put them on a single dashboard tab fed by the raw-data tab.
Step 6: Write the one-paragraph read. Every week, one paragraph: what moved, what you think caused it, what you are doing about it. The paragraph is the product — the charts are the evidence.
Total setup time: an afternoon. Total weekly run time: thirty to sixty minutes for a 30-question set. That is the entire cost of a working AI visibility practice before tooling enters the picture.
When Should You Upgrade to Paid Tooling?
Buy software when the spreadsheet costs more than the subscription — in time, in missed engines, or in competitor coverage you cannot do by hand. The triggers:
The question set outgrows manual runs. Past 50–100 questions across multiple engines, weekly manual testing becomes a part-time job. Trackers run it nightly.
You need competitor tracking at scale. Manual checks of five competitors across thirty questions is doable; across three engines and a hundred questions it is not. Tools log competitor citations automatically.
You need the audit trail. Client-facing dashboards benefit from a tool’s cached result copies and exportable records — proof of what the AI said on a given date, which a spreadsheet can claim but not prove.
You need daily cadence. Weekly sampling is enough for strategy; if a client wants daily monitoring (launches, PR crises), automation is the only sane route.
Before buying, run the vendor’s trial against your spreadsheet for two to four weeks on the same question set. You want to see the tool disagree with your manual checks at least once — that disagreement teaches you its methodology, and a tool whose methodology you understand is worth ten tools you trust blindly. Pricing and plan structures move often; confirm current terms before committing.
How Do You Present This to Clients?
The dashboard is only half the deliverable. The other half is the narrative discipline around it.
Lead with the trend, not the number. “Citation share of voice rose from 18% to 24% this quarter across ChatGPT and Perplexity” beats “your share of voice is 24%.” The first sentence contains a story; the second is trivia.
Separate market movement from your movement. If AI Overview presence rate jumped because Google expanded coverage, say so — do not let a rising tide look like your rowing. Clients who learn this distinction trust you more, not less.
Show the diagnostic, not just the score. The cited-page distribution is where the insight lives: “your pricing page is cited in 40% of answers while your homepage appears nowhere — here is what that tells us about what the engines consider authoritative.” Metrics describe; diagnostics prescribe.
State the uncertainty out loud. “These numbers come from weekly sampling of 30 questions across 3 engines. They are directionally reliable; they are not exact.” This sentence has never lost a client and has saved many.
Connect to the work. Every report ends with: here is what we changed, here is what we are changing next, and here is the metric we expect it to move. A dashboard without a work plan is a weather report.
The temptation to oversell is strongest exactly when the numbers look good. Resist it. The clients who stay are the ones who were never misled.
Final Thoughts
An AI visibility dashboard is not a tool purchase — it is a discipline with a spreadsheet at its center. Define the questions, run them honestly, track the trends, and narrate what they mean. The tools can scale the method later, but they cannot replace it: every vendor number you cannot reproduce by hand is a number you do not truly understand. Start this week with twenty questions and one afternoon. In a quarter, you will have the one thing most agencies still lack — an honest, repeatable answer to “how are we doing in AI search?”
People Also Ask
How many questions should I track?
Twenty to fifty is the practical range. Below twenty, the sample is too thin for a stable trend; above fifty, manual weekly runs become unsustainable and it is time for tooling. Start at the low end and grow the set as the practice matures.
Can I build this entirely in Google Sheets?
Yes. Raw-data tab for weekly runs, a dashboard tab with charts fed by formulas, and a notes column for the weekly read. No add-ons required. The only paid input you might add later is a tracker subscription for scale.
How is this different from rank tracking?
Rank tracking measures your position for queries on one engine. AI visibility measures your presence inside answers across several engines — a citation is not a position, and an AI answer is not a results page. The two practices complement each other; neither replaces the other.
Should clients see the raw data or just the dashboard?
The dashboard plus the weekly one-paragraph read. Raw data stays available on request — transparency builds trust — but leading with raw exports invites misreading of volatile single data points.
How often should the dashboard update?
Weekly sampling is the standard for strategy work: frequent enough to catch trends, sparse enough to stay sane. Daily only when a specific event (launch, PR crisis) justifies it.
