The question that sinks most GEO projects is not "how do I show up in ChatGPT?" but "how do I prove it worked?". Without a scorecard defined before you start, every outcome looks like a draw.
The attribution problem, head on
When an assistant recommends your brand, the user often does not click: they type your name into Google, go straight to your site, or simply remember you weeks later. Which means the conversion lands as direct or branded traffic, not as an AI referral.
That is why a GEO scorecard has two levels: visibility metrics (cause) and business metrics (effect). Never measure only the second.
Level 1: six visibility metrics
- Mention rate. Share of answers, across your fixed prompt set, where your brand appears. This is your headline metric.
- Share of Model. Your mentions divided by all brand mentions in those answers. It gives the first metric context.
- Position in the answer. Top pick, inside the list, or trailing mention. The difference in clicks is enormous.
- Sentiment and framing. How you are described: leader, budget alternative, limited option. Log the literal sentence.
- Source citations. Which domains back the answer, and how many of them include you. This is the most actionable lever.
- Coverage by prompt family. You can do well on "alternatives to X" and not exist in "best tools in the category".
Measure weekly with the same prompt set. If you change prompts, version them and record the change date.
Level 2: two business metrics
AI-assisted traffic. Segment sessions in your analytics whose referrer is chatgpt.com, perplexity.ai, gemini.google.com or copilot.microsoft.com. Typically low volume and high intent: look at conversion, not visits.
Branded search. In Search Console, track impressions and clicks for queries containing your name. A working GEO campaign pushes this curve even when the AI referrer stays invisible.
How to build the monthly report
A one-page structure that survives a leadership meeting:
- Headline: mention rate and Share of Model, with month-on-month change.
- Competitive movement: who rose and who fell in the same answers.
- Wins and losses: specific prompts you entered or dropped out of.
- Causes: content published, sources earned, technical fixes shipped.
- Business: assisted traffic, branded search and associated conversions.
- Next three actions, with an owner and a date.
What timeline to expect
- Weeks 1-4: baseline and technical fixes. Expect no movement in mentions.
- Weeks 4-12: the first long-tail prompts land, especially if you published answer-shaped content and earned third-party citations.
- Months 3-6: category prompts — the contested ones — start to move. This is where you learn whether the strategy works.
Anyone promising category-level answer leadership in three weeks is selling you smoke.
Mistakes that ruin measurement
- Changing prompts mid-quarter and losing comparability.
- Measuring a single assistant and extrapolating.
- Celebrating mentions without checking sentiment.
- Attributing every bump in direct traffic to GEO with no other evidence.
- Not logging what you published each week: with no causes, the series teaches nothing.
For the detail on the headline metric, read how to measure Share of Model and AEO KPIs for a board-level report.