Illustrative photograph. Photo: Solar calculator made in China.jpg by Suginami, via Wikimedia Commons — CC0. Downloaded Wikimedia version; no custom edits.
Your business appears more often in a prompt-tracking report. Website traffic barely changes. The sales team still reports poor-fit leads.
Those observations can coexist. Each measures a different part of discovery, and combining them into one “AI score” hides the decisions you need to make.
Start with a metric dictionary
Before building a dashboard, define what each number means and where it comes from.
Metric | What it can tell you | What it cannot establish |
|---|---|---|
Sampled mention rate | Presence in a defined set of tested answers | Total platform exposure |
Sampled citation rate | Whether tested answers link to selected sources | How many real customers saw those answers |
Referral sessions | Visits attributed by your analytics setup | All influence from an assistant |
Qualified inquiries | Leads meeting your stated fit criteria | Whether AI was the only influence |
Closed revenue | Recorded revenue linked through your attribution method | The incremental effect without further testing |
If a platform supplies impressions in the future, add its published definition. Do not assume it matches the denominator in your prompt tests.
Give every rate a denominator
Suppose a fictional company runs 30 fixed questions three times each. It is named in 18 of the 90 answers. Its sampled mention rate is 20%.
That denominator includes repeated trials, not 90 unique buyer intents. Report both the 30 questions and 90 answers so readers understand the design.
If the company later replaces half the questions with branded queries, its new rate is not directly comparable. Keep a fixed core set and track exploratory questions separately.
Build a referral view
OpenAI’s publisher guidance describes ChatGPT referral tracking using the source parameter utm_source=chatgpt.com. Use the documented signal as one input to your reporting, and validate what your own setup actually captures. See the publisher FAQ.
Inspect source values, referrers, landing pages, and conversions. Check that redirects preserve relevant parameters. Apply a consistent filtering policy for internal visits and obvious automated traffic.
Do not assume every assistant-influenced buyer arrives through a tracked link. Someone may research a company in one place and visit its website later through another route.
Connect traffic to fit
For a B2B company, track whether an inquiry matches the customer profile, whether sales accepts it, and whether it progresses. For a retailer, examine completed orders alongside cancellations and returns.
A fictional example makes the distinction clear: ten inquiries containing two strong opportunities may be more valuable than forty inquiries containing none. The actual answer depends on deal value, cost, and your qualification rules.
Add an optional discovery question to customer conversations, such as “Where did you first hear about us?” Treat self-reported answers as useful context, not precise tracking data. Avoid counting one sale twice when it appears in multiple attribution views.
Make the scorecard produce a decision
Pair each result with a next action. Inaccurate sampled answers suggest an information audit. Relevant referrals that fail to convert suggest a landing-page or offer problem. A rise in raw visits with no qualified activity calls for a traffic-quality check.
The useful report ends with a decision and an owner. More visibility becomes commercially meaningful when the business can explain which customers it helped reach and what happened next.
Next read: Test whether your changes made a difference.
