Where the reading happens
When a buyer asks ChatGPT which tool to use, the answer is the product. It names two or three brands and explains why. The buyer reads that, forms a shortlist, and often opens nothing. If a link is opened, it is usually the cited source, and it is opened by the fraction of readers who want to check. Your analytics sees that fraction and nothing else.
So the session count from AI referrers measures the checking, not the reading. A brand that is named on every buying prompt and rarely clicked is winning the shortlist. A brand that is never named has nothing to be clicked. Both show near-zero AI referrals. Analytics cannot tell them apart.
What the referral number hides
| Situation | AI referrals in analytics | What is actually happening |
|---|---|---|
| Named and cited on buying prompts | A few sessions a week | You are the recommendation; most readers do not need to click |
| Cited, not named | A few sessions a week | Your page is background evidence for an answer that recommends someone else |
| Neither | Zero | You are not in the answer at all |
| Named on brand prompts only | Occasional branded sessions | People who already know you are checking; buyers on job prompts never see you |
The first and third rows are the same in your dashboard and opposite in the market. This is the reason the leading metric has to be measured in the answer, not on the site.
The leading metric: named and cited on a prompt set
Fix eight to sixteen prompts your buyers type. Run them weekly in ChatGPT with search on, and in whichever other engines matter. For each, record whether your brand is named and which URL is cited. Those two counts move before referrals do, and they move for a reason you can act on: a page entered the retrieved set, or it did not. The method is on how to measure whether ChatGPT mentions your brand.
The lagging metric: referral sessions, kept honest
Still track referrals. Just track them correctly and expect them to be small.
- Group the referrers: chatgpt.com, chat.openai.com, perplexity.ai, copilot.microsoft.com, gemini.google.com, and the assorted AI aggregators. Give the group its own channel.
- Expect Google AI Overviews and AI Mode to arrive as ordinary Google organic. They carry no distinct referrer. Search Console's generative AI report shows impressions for those features by page; it does not show clicks.
- Separate crawler hits from human sessions. Fetches by ChatGPT-User or PerplexityBot are not visits, and some analytics setups count them. The crawler reference on the research hub explains how to tell them apart.
- Read the trend over months, not weeks. A leading-metric change in week two may show up as referral change in week eight, if at all.
How to report it without confusing anyone
| Metric | Role | Where it comes from | Cadence |
|---|---|---|---|
| Named on buying prompts | Leading | The prompt set, run in the engine | Weekly |
| Cited URL on buying prompts | Leading | The prompt set, run in the engine | Weekly |
| Generative AI impressions by page (Google) | Leading, Google only | Search Console | Weekly |
| AI referral sessions | Lagging | Analytics, grouped channel | Monthly |
Put the leading rows first. If leadership only reads one line, it should be the one that changes when you ship a page, not the one that changes a quarter later for reasons nobody can trace.