Magic TeamsHow to Build a Practical AI Referral Dashboard
A practical AI referral dashboard tracks five things: whether AI answers name you (source visibility), whether what they say is correct (citation accuracy), how many people click through (visits), what those visitors do next (qualified actions), and which buyer questions get answered with someone else’s name (unanswered buyer questions). You can build it in a spreadsheet in an afternoon. The data comes from GA4, Search Console, Bing Webmaster Tools, your CRM, and a monthly prompt check you run by hand.
Here’s the usual situation. You run a 10 to 30 person agency or a small advisory practice. A prospect mentions they “asked ChatGPT” before booking. Your marketing lead wants a number for it, and someone suggests buying a visibility tool that reports a score nobody can explain.
This post covers what each measure is for, where each one breaks, and a worked example you can copy. I’m Satya Phanindra Reddy, founder of Magic Teams AI. We build reviewable, human-in-the-loop AI workflows for founders, and measurement is the part most teams skip. For the attribution mechanics in more depth, read how to measure AI search referrals. For the scoring routine, see how to measure AI search visibility without an expensive tool. This one is about putting it all on one screen.
Why can’t one analytics tool give you this?
No single platform sees the whole path from AI answer to booked call. Each one covers a slice.
GA4 sees some clicks. Google Analytics added an AI Assistant channel to the default channel group in its May 13, 2026 release. The channel definition names sources “like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” The same page says Google’s own AI Overviews and AI Mode are excluded and count as Organic Search.
Search Console shows Google AI impressions, not clicks. Google’s AI features guidance says AI Overviews and AI Mode traffic is “included in the overall search traffic in Search Console.” Since June 2026 there’s also a generative AI performance report. Its help page describes impressions only, broken down by page, country, device, and date. It rolled out in stages, so check your Performance menu.
Bing shows citations, not clicks. Bing Webmaster Tools launched AI Performance in public preview in February 2026. It reports total citations, cited pages, and grounding queries across Copilot and Bing’s AI summaries. Search Engine Journal’s summary notes Microsoft calls the data a sample, and Adobe’s walkthrough says it “does not yet include other AI assistants such as ChatGPT.”
None of them check accuracy. That gap matters. The EBU and BBC’s News Integrity in AI Assistants study reviewed more than 3,000 answers from ChatGPT, Copilot, Gemini, and Perplexity. It found that 45% had at least one significant issue.
That study looked at news, not service businesses. Still, it’s a good reason to read what assistants say about you instead of just counting mentions.
So the dashboard is five small measures, each with a named source and a named owner.
What are the five measures, and where does each come from?
We call this the Five-Tile Referral Board. It came from a simple exercise. We listed the questions founders actually ask about AI search: Are we showing up? Is it right? Is anyone coming? Are they buying? What are we missing? Then we matched each one to the cheapest data source that answers it honestly. Each question became one tile.
| Tile | The question it answers | Where the data comes from | Cadence |
|---|---|---|---|
| 1. Source visibility | Do AI answers name or cite us for our buyer questions? | Monthly prompt panel, Bing AI Performance citations, Search Console generative AI impressions | Monthly |
| 2. Citation accuracy | When they name us, is it correct? | Same prompt panel, checked against your current offer | Monthly |
| 3. Visits | How many people click through? | GA4 AI Assistant channel plus a referrer filter, by landing page | Weekly |
| 4. Qualified actions | Do those visitors book, enquire, or buy? | GA4 key events by channel, CRM “how did you hear” field | Weekly |
| 5. Unanswered buyer questions | Which questions go to competitors or nobody? | Prompt panel misses, sales call notes, inbound emails | Monthly |
Here’s the board as a one-screen wireframe. The numbers are hypothetical and match the worked example below.
Notice what’s missing. There’s no “AI rank” and no blended visibility score. A generated answer doesn’t have a ranking position, and a score built on prompts you can’t see gives you nothing to act on.
How do you build it, step by step?
Build it in this order. Each step feeds the next.
1. Write 20 buyer questions. Pull them from discovery calls, proposal emails, and objections. Use the buyer’s wording, like “best HubSpot agency for B2B SaaS under 50 people,” not your keyword list.
2. Run the prompt panel. Ask each question in the assistants your buyers use. For most B2B buyers that means some mix of ChatGPT, Perplexity, Gemini, Copilot, and Google AI Mode. Pick your set and keep it fixed.
Log three things per answer: named or not, which URL was cited, and any factual error. Answers vary by user, location, and day, so treat this as a sample, not a census.
3. Set up GA4 views. Use the AI Assistant channel as your baseline. Then add an exploration filtered on session source with a regex like chatgpt\.com|openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com. That catches sources the channel may not classify.
Break it down by landing page. The page tells you which answer probably sent the visit.
4. Tag qualified actions. Mark your booking, contact, and quote events as key events in GA4. Then add “AI assistant (ChatGPT, Perplexity, etc.)” as an option in your intake form and CRM source field. Self-reports catch what referrer data drops.
5. Log unanswered questions. Every panel answer that names a competitor, or gives a generic answer, goes on a list. So does every sales call question your site doesn’t answer clearly. This is the tile that changes what you publish.
6. Review monthly for 30 minutes. Look at all five tiles together and make one decision. The rule below keeps that decision from turning into a debate.
How do you decide what to fix first?
Work down the tiles in a fixed order. Fix what’s wrong before you chase more of it.
The order matters. A wrong price in an AI answer sets the wrong expectation before the first call. A missing mention just means you weren’t in the conversation.
What does a worked example look like?
This example is hypothetical. It’s a composite built to show the math, not a Magic Teams client result.
A 14-person B2B design agency runs its first full month on the board. It checks four assistants: ChatGPT, Perplexity, Gemini, and Copilot.
- Source visibility. 20 questions across 4 assistants gives 80 answers. The agency is named in 11. That’s 14%.
- Citation accuracy. Of those 11 mentions, 3 contain errors. Two quote a 2024 retainer price. One lists a print design service they dropped. Accuracy is 8 of 11, or 73%.
- Visits. GA4 shows 62 AI Assistant sessions, and 38 land on one comparison article. Search Console’s generative AI report shows impressions on that same article but, by design, no clicks.
- Qualified actions. 4 contact form submissions come from AI Assistant sessions. 2 more prospects say “ChatGPT” on discovery calls but arrived as Direct.
- Unanswered buyer questions. 6 questions consistently cite competitors. Three are about Webflow migrations, which the agency does but never wrote about.
The decision. The first branch fires, so accuracy comes first. The team updates the pricing page, redirects the old print design page, and adds a dated “last updated” line.
Next month, if accuracy is clean, the second branch fires. They write one answer-first page on Webflow migrations. That’s one page, not three near-duplicates.
The money line. Tile 4 is where a dollar figure lives. Say the agency closes 1 in 4 qualified enquiries and the average first project is $12,000. Six AI-influenced enquiries (4 tracked, 2 self-reported) × 25% × $12,000 gives about $18,000 in pipeline. Report it as an estimate, never as attribution.
Add one column to your prompt log: “change this triggered.” Fill it in only after a person has read the AI answer. A row that never triggers a change is a question you can drop from the panel.
What should the dashboard not try to do?
Set these limits before anyone presents the numbers.
- It won’t give you full attribution. Mobile apps and copy-pasted links often arrive with no referrer. Your GA4 count is a floor, not a total.
- It can’t count clicks from Google AI Overviews or AI Mode. GA4 files them under Organic Search, and Search Console’s generative AI report shows impressions only. Don’t build a tile that pretends otherwise.
- The prompt panel is directional. Twenty questions run monthly will show movement. They won’t prove you “rank” anywhere.
- It doesn’t promise citations. Google describes its AI features as part of Search, built on the same foundations. Crawlable pages, clear answers, and accurate facts are the levers. The dashboard tells you whether they’re working.
If you work in a regulated field like law, accounting, or financial advice, add one check to the accuracy tile. Flag any AI answer that misstates what you’re licensed to do, or where. Have a qualified colleague review the fix before you publish it, and remember your bar association’s or professional body’s advertising rules still apply to the pages AI tools quote.
How does this connect to the rest of your AI search work?
The board is the measurement end of a system. The publishing end is in how to build an AI search content system for a business. The strategy behind it is in what answer engine optimization means for B2B. For the same pipeline math applied to operational workflows, see how to measure ROI on AI automation.
If keeping five tiles fed and reviewed sounds like one more thing landing on your desk, that’s the kind of connected, reviewable workflow we map in a fit call.
FAQ
Do I need a paid AI visibility tool?
Not to start. A spreadsheet, GA4, Search Console, and Bing Webmaster Tools cover the platform side, and your own prompt panel covers the rest. Paid tools can scale the panel. Only buy one if it shows you the exact prompts and answers behind its numbers.
How many buyer questions should the prompt panel include?
Twenty is enough for a small firm to see patterns without losing a day. Keep the list stable for at least three months so trends mean something. Then swap out questions that no longer match what buyers ask.
Why isn’t Perplexity showing in my AI Assistant channel?
Google’s channel definition lists examples, and Perplexity isn’t one of them. Check the session source dimension directly and keep the referrer regex as a backup.
Who should own the dashboard?
One person updates it, and the founder reviews it monthly. In a small agency that’s usually the marketing lead or an ops coordinator. Corrections to pricing or service claims go to whoever owns the offer.