How much traffic comes from ChatGPT, and how to measure it
“Is ChatGPT sending us any traffic, or is it all talk?” There is a short answer and an honest one. The short answer: ChatGPT referrals carry utm_source=chatgpt.com, so your analytics already sees them. The honest one: traffic is the smallest part of what AI engines do to your business, and relying on it alone means measuring the buyers who click and ignoring the ones who do not.
The three instruments you already have
1. The utm parameter. OpenAI’s publisher documentation states that ChatGPT referrals carry utm_source=chatgpt.com (https://help.openai.com/en/articles/12627856). Filter your analytics for it and you have referral traffic from ChatGPT without installing anything. Segment it by landing page and conversion, and you know which questions bring visitors who act.
2. Search Console, generative AI reports. Google ships a generative AI performance report covering AI Overviews and AI Mode, with one requirement: you must opt in to the generative AI features in Search Console for eligibility and reporting. If the opt-in is missing, you are not in the population being reported. Check it before reading any trend (https://developers.google.com/search/docs/appearance/ai-features).
3. Branded search lift. When assistants mention a brand, some people google it afterwards. If AI answers mention you more than before and branded searches rise with them, that correlation is weak evidence individually and useful evidence in aggregate. Watch it alongside the other two, not instead of them.
Why traffic undercounts what is happening
Here is the uncomfortable part. The purchase walk for many categories now looks like this: ask an assistant, open a retailer app, read two reviews, buy at the shelf. Times the buyer visited the brand’s website: zero. Every one of those buyers was influenced by an AI answer, and none of them appears in any traffic report.
Traffic measures the click. Influence happens before the click, and increasingly without one. A brand can be losing an entire category conversation inside the answers and its analytics dashboard stays green, because nothing dropped. That is the gap a traffic-only view cannot close: it counts the visits you won, not the recommendations you lost.
Traffic also cannot answer the competitive questions: who gets named instead of you, in which sources the assistants place their trust, whether the answer calls you “the best” or buries you in a list of eight. Those are measurement questions, and they are answered by a study, not by analytics.
How the two layers fit together
Use the analytics layer for what it is good for: confirming that ChatGPT referrals exist, tracking their growth, and catching the moment they start converting. Use a measurement layer for the part analytics cannot see: your share of the answers, your competitors’ share, the sources behind both.
The monitoring layer makes the second one a habit: the same frozen prompt bank, re-run on a cadence, rates comparable month over month. When a traffic number moves, the study tells you why; when nothing moves in traffic but your share of recommendations climbs, you caught the effect the dashboard structurally misses.
Next step: Visibility monitoring