A five-step ROI loop
- Baseline: pick 30 to 50 buyer-intent questions and run them on ChatGPT, Claude, Gemini and Perplexity.
- Record mentions and citations separately. A citation links to your page; a mention does not.
- Check access: can each AI agent fetch the cited page and read its main text?
- Ship one change and log the date.
- Re-run the same questions and re-probe the page. Compare like with like.
Metrics that hold up
| Metric | Why it matters |
|---|---|
| Citation rate per question set | Shows whether answers link to you |
| Agent fetch status and text length | Shows whether the model can read the page at all |
| AI referral sessions | chatgpt.com, perplexity.ai and similar referrers in analytics |
| Pipeline from AI referrals | Ties visibility to revenue in your CRM |
Common mistakes
Changing the question set between readings, counting mentions as wins, and running many fixes at once so none can be attributed.