How to Track the Traffic and Customers You Get from ChatGPT

Check the referrer (the site the visit came from) and the utm_source. ChatGPT tags its links with utm_source=chatgpt.com, but in our logs 8 in 10 visits from ChatGPT arrive with no referrer: if your analytics or your CRM only look at the referrer, they count them as direct or referral and your AI channel shows zero. This is how we measure it on kiwop.com: GA4, server logs, bots and CRM.

Review my tracking Tell us about your case
  • 8 in 10 visits from ChatGPT arrive with no referrer
Scroll

Definition

How do you track traffic from ChatGPT?

You classify every visit by its referrer (the site it came from) and its utm_source at the same time, without requiring utm_medium: ChatGPT tags its links with utm_source=chatgpt.com and no medium, and on kiwop.com about 8 in 10 of those visits arrive with no referrer. Then you separate bots from people by verified IP, not by user agent, and carry the channel through to the form or the sale.

We explain it with our own data: the kiwop.com nginx logs between 11 and 25 September 2026. And it happened to us: until 25 September 2026 our CRM stored ChatGPT leads as “utm:chatgpt.com” instead of “chatgpt”, because the channel rule only looked for chatgpt.com in the referrer. Volume is small (about 7 visits a day in our case), so every misclassified visit matters. Where we could not verify something, such as exactly how GA4 treats a visit with utm_source and no referrer, we say so.

In numbers

What we saw in the kiwop.com logs

15 days of first-party data, 11 to 25 September 2026.

  • 8 in 10 Arrive with no referrer Of 112 landing requests tagged utm_source=chatgpt.com, 93 had no referrer and 19 had chatgpt.com as referrer.
  • 106 Human visits from ChatGPT In 15 days, counted as IP × day: about 7 a day. 2 ended in a contact form.
  • 91% On desktop On a B2B services website. An online store may see a different split.
  • ~1,100 Bots with a fake user agent Requests with an OpenAI user agent from IPs that are not OpenAI’s: 394 ChatGPT-User, 382 OAI-SearchBot and 349 GPTBot.

Key points

How to Track ChatGPT Traffic in 6 Steps

The order in which we set it up on kiwop.com. The first three already tell you how many customers it brings.

  1. 01

    Classify by referrer AND utm_source

    A rule that only checks the referrer fails, and so does one that requires utm_medium. The one that works: if the referrer host or the utm_source contains chatgpt.com or chat.openai.com, the channel is ChatGPT; same for perplexity, claude.ai, copilot and gemini. And put it before your generic UTM rule, or “utm:chatgpt.com” wins.

  2. 02

    In GA4, build your own channel group

    In May 2026 Google added an AI Assistant channel to the default channel group, but according to its documentation it is assigned when the referrer matches Google’s list of assistants, which is not published. Since most ChatGPT visits arrive with no referrer, create a custom channel group with an “AI assistants” channel: Source matches regex chatgpt\.com|chat\.openai\.com|perplexity|claude\.ai|copilot|gemini, placed above Referral. It also applies to past data.

  3. 03

    Carry the channel into your CRM

    A visit is not a customer. Store the referrer and UTMs of the first page viewed, send them with the form and derive the channel on the server with the same rule as your analytics. Every lead then carries “chatgpt” and you can add up the revenue from that channel. That is what we do in marketing attribution.

  4. 04

    Separate bots from people by IP

    User agents are spoofed: in 15 days we saw about 1,100 requests with an OpenAI user agent from IPs that were not OpenAI’s. OpenAI publishes its ranges in three lists (searchbot.json, gptbot.json and chatgpt-user.json): count as an OpenAI bot only what comes from those IPs and treat the rest as suspicious.

  5. 05

    Count visits, not requests

    In server logs, a visit is one IP on one day; otherwise one person opening five pages counts five times. And clean your analytics: Plausible was counting bots that run JavaScript from data centres as visits. We have filtered them by network (ASN) since 25 September 2026.

  6. 06

    Also measure citations without clicks

    Many answers cite you and nobody clicks. Bing Webmaster Tools has an AI Performance report, in public preview since February 2026, with citations of your pages in Copilot and Bing’s AI summaries, without click data. To earn more citations, see how to show up in ChatGPT.

Who it is for

Is this level of detail worth it for you?

It depends on what each lead is worth.

Who it's for

  • B2B companies with few leads a month, where one misattributed contact changes where you invest.
  • Anyone who already sees “chatgpt.com” in their analytics and does not know how many customers it brings.
  • Teams investing in GEO who need to know whether it turns into leads and sales, not just mentions.
  • Websites with a CRM or forms that already store the visit’s referrer and UTMs.

Who it's not for

  • High-volume online stores: GA4 channel groups plus sales by channel give a good enough reading, and a few dozen misclassified visits will not change the decision.
  • Websites with no form and no online sales: there is no conversion to attribute, counting visits is enough.
  • Anyone who wants to know what each person asked ChatGPT: the link only carries utm_source, not the question.

FAQ

FAQs About Tracking ChatGPT Traffic

What people ask us once they see chatgpt.com in their analytics.

Why does ChatGPT traffic show up as direct or referral?

Because most visits arrive with no referrer. On kiwop.com, of 112 landing requests tagged utm_source=chatgpt.com, 93 had no referrer and 19 had chatgpt.com: about 8 in 10 with no referrer. If your tool classifies by referrer only, those visits are “direct”; if it requires utm_medium, they end up in “referral” or in a generic UTM bucket. It happened in our own CRM until 25 September 2026.

Does GA4 have a channel for ChatGPT and other AI assistants?

Yes. Since May 2026 the GA4 default channel group includes AI Assistant, with sources such as ChatGPT, Gemini, Deepseek, Copilot or Grok. According to the documentation, the medium is “ai-assistant” and GA4 sets it when the referrer matches its list of assistants, which it does not publish. The documentation does not say what happens to a visit that carries utm_source=chatgpt.com and no referrer, which is the usual case. Check it in Traffic acquisition by filtering the source for “chatgpt”.

Which regex should I use in a GA4 custom channel group?

Create an “AI assistants” channel with the condition Source matches regex chatgpt\.com|chat\.openai\.com|perplexity|claude\.ai|copilot|gemini and move it above Referral: traffic goes into the first channel whose rule it matches. A standard property allows 2 custom channel groups on top of the predefined one (5 on GA4 360), and they apply to past data too.

How do I know if a ChatGPT-User visit is real or a bot pretending to be it?

By IP. OpenAI publishes each bot’s ranges in chatgpt-user.json, searchbot.json and gptbot.json. In 15 days on kiwop.com, about 1,100 requests with an OpenAI user agent came from other IPs: 394 as ChatGPT-User, 382 as OAI-SearchBot and 349 as GPTBot. And remember that ChatGPT-User is ChatGPT reading your page to answer someone, not a person browsing your site.

How much ChatGPT traffic is normal?

We know of no reliable benchmark. Ours: 106 human visits in 15 days (about 7 a day), 91% on desktop, and 2 contact forms, on a B2B digital services website. Use it as an order of magnitude, not as a market average.

Can I see what someone asked ChatGPT before landing on my site?

No. The link only adds utm_source=chatgpt.com; the question does not travel with it. The closest you get is the landing page of each visit and, for Copilot, the “grounding queries” in the Bing Webmaster Tools AI Performance report: the phrases the AI used to retrieve the content it cited. At Kiwop Labs we publish the estimated assistant volume of 60 questions about agencies and AI in Spain.

Does my cookieless analytics count these visits correctly?

It overcounts if you do not filter bots. Plausible, which we use on kiwop.com, was counting bots that run JavaScript from data centres as visits. Since 25 September 2026 we filter them by network (ASN), so our figures before that date are inflated. Before trusting the total, check which networks are visiting you.

And how do I get ChatGPT to send me more visits?

That is a different question with its own page: how to show up in ChatGPT. Measuring comes first: without the channel properly classified you will not know whether what you do to show up is working.

Sources

Sources for the data on this page

First-party figures from kiwop.com and official documentation, checked on 25 September 2026.

  1. kiwop.com nginx logs, 11-25 Sep 2026 (first-party data) 112 landing requests tagged utm_source=chatgpt.com (93 with no referrer, 19 with chatgpt.com as referrer); 106 human visits counted as IP × day, 91% on desktop; about 1,100 requests with an OpenAI user agent from IPs outside its ranges (394 ChatGPT-User, 382 OAI-SearchBot, 349 GPTBot).
  2. Kiwop CRM and lead channel rule (first-party data) Until 25 Sep 2026 ChatGPT leads were stored as “utm:chatgpt.com”; since that day the rule checks both referrer and utm_source. 2 contact forms from ChatGPT in the period.
  3. kiwop.com Plausible (first-party data) Bot filter by network (ASN) active since 25 Sep 2026.
  4. OpenAI: documentation of its crawlers (OAI-SearchBot, GPTBot, ChatGPT-User) What each bot is for and a link to its IP list. Checked on 25 Sep 2026.
  5. OpenAI: ChatGPT-User IP ranges List dated 22 Sep 2026. Checked on 25 Sep 2026.
  6. OpenAI: OAI-SearchBot IP ranges List dated 2 Jan 2026. Checked on 25 Sep 2026.
  7. OpenAI: GPTBot IP ranges List dated 22 Sep 2026. Checked on 25 Sep 2026.
  8. Google Analytics: default channel group Definition of the AI Assistant channel (medium “ai-assistant”, assigned when the referrer matches Google’s list of assistants), Referral, Direct and Unassigned. Checked on 25 Sep 2026.
  9. Google Analytics: what’s new AI Assistant channel announcement, 13 May 2026. Checked on 25 Sep 2026.
  10. Google Analytics: custom channel groups 2 custom channel groups per standard property (5 on 360), 50 channels per group, retroactive, evaluated in order. Checked on 25 Sep 2026.
  11. Bing Webmaster Blog: AI Performance in Bing Webmaster Tools (public preview) Published 10 Feb 2026: citations in Copilot, Bing AI summaries and partners, grounding queries; no click data. Checked on 25 Sep 2026.

Next step

How Many Customers Does ChatGPT Really Bring You?

We review your analytics and your CRM, fix the channel rule and separate bots from people. If your volume does not justify it, we will tell you.

Review my tracking
  • No commitment
  • Response in 24h
  • Custom proposal
Last updated: September 2026

Let's talk.

Initial technical consultation

AI, security and performance. Diagnosis with phased proposal.

  • NDA available
  • Response <24h
  • Phased proposal

Your first meeting is with a Solutions Architect, not a salesperson.

Request diagnosis