For several months, some publishers have observed a new category in their audience reports: sessions from ChatGPT, Perplexity or Gemini. Should we see the emergence of a new acquisition channel, the “referral LLM”, or an illusion of attribution linked to analytics tools?
A still marginal but measurable phenomenon
The appearance of clickable links in responses generated by ChatGPT (via integrated UTMs) or Perplexity created a first visible flow. In some cases, ChatGPT appears in Web Analytics tools as 3rd source of referral trafficbehind LinkedIn and Twitter as we can see on FRENCHWEB.FR. These signals feed the idea that a new channel is emerging.
But these figures must be qualified:
- Not all LLMs offer outgoing links (Claude, Gemini, Mistral do not systematically do so).
- Clicking is not the dominant mode of use: most users are satisfied with the textual response.
- Some of the sessions attributed to ChatGPT may actually come from internal or indirect sharing tests (copy and paste from a response, click on a link relayed elsewhere).
Attribution biases
The identification of “referral LLM” traffic faces several technical limits:
- Partial UTMs : Only publishers who voluntarily embed tracking parameters in their responses (like ChatGPT in beta) allow clear attribution.
- Absence of referrer : in many cases, the traffic arrives classified as “direct” due to lack of information transmitted by the browser.
- Interface partitioning : when a response is generated in a mobile app, referral data is not always transmitted.
Result: the importance of the channel is often underestimatedbut its reality remains difficult to prove exhaustively.
An invisible redistribution of value
Even without clicks, LLMs change the traffic value chain. Each response generated obtains a search intention which no longer passes through Google, and reduces the probability of a click to the source sites. The visible referral is only the tip of the iceberg:
- Notoriety gains (the brand is cited in the response without an associated click).
- SEO traffic losses (the user finds his answer directly in the AI and does not open a site).
Towards a new measurement model
To analyze this emerging channel, several avenues are explored:
- Segmentation in GA4 : create a specific “referral LLM” category bringing together ChatGPT, Perplexity, etc.
- Citation monitoring : use monitoring tools to know when and how a brand is cited in the responses generated, even without a click.
- Qualitative studies : user surveys to understand how many actually use the links suggested by AI.
What implications for brands?
- Short term : the volume of LLM referral traffic remains low compared to SEO and social networks.
- Medium term : visibility measurement can no longer be limited to sessions and clicks; it will be necessary to integrate the presence in AI responses as an indicator of authority.
- Long term : the switch to “answer engines” could redefine the acquisition model: less direct traffic, but more value associated with brand mention in the corpora used by AIs.
How to measure LLM referral traffic
| Method | Principle | Benefits | Boundaries | Recommended use |
|---|---|---|---|---|
| UTM integrated by LLMs (e.g. ChatGPT, Perplexity) | The models automatically insert a link with tracking parameters (utm_source, utm_medium) | Clear attribution, visible in GA4 as “referral” | Limited to AIs that display clickable links; not standardized between models | Short-term monitoring of clicks actually generated by AI responses |
| GA4 / Analytics segmentation | Creation of a personalized “LLM referral” channel bringing together ChatGPT, Perplexity, Gemini (if identifiable) | Centralization of data, allows comparison with SEO and social | Many “unallocated” cases (traffic classified live), depends on the settings | Good for monthly reporting and trend monitoring |
| Citation monitoring (brand tracking LLM) | Tools that regularly query models to see if a brand or link is cited | Measuring visibility even without a click, gives an authority score | Does not reflect real traffic, depends on the tool and its prompts | To be combined with analytics to manage awareness “excluding traffic” |
| Qualitative analysis (surveys, user tests) | Study the use of AI by users and their propensity to click on links | Provides behavioral insights (e.g. 20% of users click on a proposed link) | Limited sample, no automated tracking | Useful for validating quantitative data and understanding the user journey |
| Indirect monitoring (social mentions, backlinks) | Identify repeats of content copied from AI responses to other platforms | Side effect detection (quotes on LinkedIn, forums, blogs) | Fuzzy attribution, does not distinguish AI origin | To capture the “derived value” of GEO beyond direct traffic |
The 3 GEO KPIs to follow today
-
Traffic referral LLM
- Measure the volume of visits coming directly from LLMs (via UTM or GA4).
- Allows you to identify real flows generated by ChatGPT, Perplexity or Gemini.
-
Citation rate in AI responses
- Percentage of key questions in the sector where the brand, the site or its content are cited.
- Informational authority indicator, even without an associated click.
-
GEO Authority Index
- Score combining presence in reference corpora (Wikipedia, specialized media, technical reports) and regularity of citations.
- Long-term measurement to monitor the progress of visibility “beyond SEO”.
