ChatGPT Referral Leads: The Playbook for Capturing Them

ChatGPT referral leads arrive on five referrer hostnames — chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com — plus a utm_source=chatgpt.com tag. Google Analytics 4 added an AI Assistant channel on 13 May 2026, but across 39 ChatGPT landings in our own server logs, 64% carried no referrer at all.

At a glance — capturing leads from AI assistant referrals:

• The hostnames to look for: chatgpt.com, chat.openai.com, perplexity.ai (often www.perplexity.ai), claude.ai, copilot.microsoft.com, gemini.google.com.
• The query tags: ChatGPT appends ?utm_source=chatgpt.com to links it hands out; we have also logged ?utm_source=copilot.com from Microsoft Copilot. These survive when the referrer header does not.
• GA4 files them three different ways: AI Assistant channel, Referral, or Organic Search — Google’s own documentation puts AI Overviews and AI Mode clicks in Organic Search. App-based sessions send no referrer at all and land in Direct.
• Intent is different: the engine already answered the question and produced a shortlist. The visitor wants proof and a booking link, not a definition or a PDF download.
• Volumes are small relative to the crawl: in the 30 days to 4 September 2026 our origin logs recorded 27,822 automated fetches of leadsnow.ai, 17,593 of them from AI companies’ agents, and only 5,048 of those — 29% — were live, user-triggered browses.

Step one: find the referrer hostnames in your own logs

Before you touch Google Analytics, look at raw traffic. Every AI assistant that sends a human click sends it from a hostname you can grep for. Ours, over the 15 days from 22 August to 5 September 2026, were chatgpt.com, claude.ai, www.perplexity.ai, copilot.microsoft.com and gemini.google.com. That is the whole set worth checking for most sites; DeepSeek, Grok, Meta AI and Poe exist but did not appear anywhere in that window of our logs.

Two things make this harder than it looks. First, ChatGPT rewrites the links it hands out, appending ?utm_source=chatgpt.com to the destination URL — a gift, because that tag lives in your access log whether or not a referrer header arrives. Second, a large share of those clicks arrive with no referrer at all: from the desktop and mobile apps, from copy-pasted links, and from sessions where the referrer is stripped. In our 15-day window, 39 landings carried the ChatGPT query tag and only 14 also carried a chatgpt.com referrer header. Twenty-five of 39 — 64% — would have been unattributed had we relied on the referrer alone.

So the practical rule: match on the query tag first and the referrer second. A landing-page report filtered to URLs containing utm_source=chatgpt.com will find traffic that no channel report shows you.

Step two: understand why GA4 buckets these three different ways

Google Analytics 4 gained a dedicated AI Assistant channel on 13 May 2026. Google’s default channel group documentation defines it as “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”, and the mechanism is that GA4 sets the medium to ai-assistant and the campaign to (ai-assistant) when the referrer matches an internal list of AI assistants. That list is not published, and the definition page names neither Claude nor Perplexity.

The second oddity is Organic Search. The same page defines it as arrivals “via non-ad links in organic-search results, including Google’s AI Overviews and AI Mode”. Clicks out of an AI Overview are therefore not in your AI channel at all — they are mixed in with ten blue links. The third is what happens to everything else: the GA4 source-categories list Google publishes alongside that page contains 819 sources, of which 129 are classified SOURCE_CATEGORY_SEARCH. DuckDuckGo is on it. None of the AI assistant hostnames are. Anything not caught by the AI Assistant list falls through to Referral.

Step three: build the channel group and the exploration

In GA4, go to Admin → Data display → Channel groups and create a custom group. Add a channel named AI assistants and drag it above Referral and Organic Search, because Google’s custom channel group documentation states that “traffic is included in the first channel whose definition it matches given the current order of channels in the group”. Put the wrong order in and your Referral channel eats everything. Standard properties allow two custom groups of up to 50 channels each, and custom groups apply retroactively, so you get history back rather than starting a new clock.

The condition is Source matches regex, using a full-match pattern so partial hostnames cannot leak in:

^(chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|www\.perplexity\.ai|claude\.ai|copilot\.microsoft\.com|copilot\.com|gemini\.google\.com|aistudio\.google\.com|grok\.com|meta\.ai|deepseek\.com|chat\.mistral\.ai|poe\.com|you\.com|phind\.com)$

Add a second OR condition on Manual source with the same pattern, so the utm_source-tagged sessions are captured alongside the referrer-header ones. Then build a free-form exploration with Session source / medium and Landing page as rows, Sessions and your booking-page conversion as values, and a filter of Landing page contains utm_source=chatgpt.com as a second tab. The two tabs will disagree. The gap between them is your dark AI traffic, and on our numbers that gap is the majority of it.

Step four: verify in server logs, because the crawl is not the click

Analytics runs in the browser, so it misses anything that does not execute JavaScript and mis-files anything the browser does not report. We run an mu-plugin bot logger on leadsnow.ai that records every AI agent fetch at the origin with its user agent and the URL requested, which lets us put crawler behaviour and human behaviour side by side. The comparison is sobering. In the 30 days to 4 September 2026, the logger recorded 27,822 automated fetches of this one site, 17,593 of them from AI companies’ agents (OpenAI, Anthropic, Perplexity, Meta, ByteDance, Google-Extended, Common Crawl, Cohere and You.com), of which 5,048 — 29% — were live, user-triggered browses — ChatGPT-User and Perplexity-User fetching a page in real time because someone had asked a question. The rest were index and training crawls. In the shorter 15-day window where we still hold raw access logs, 1,896 live-browse fetches produced 50 identifiable human landings once datacentre crawlers and non-browser user agents are stripped out: roughly 38 machine fetches for every click we can name. Across the 30-day set, 10% of the live fetches hit the homepage and the rest were spread over 755 distinct URLs in all, which is the single most useful thing in the dataset — every page is a landing page.

Two operational consequences follow. First, do not block the agents that send you humans. OpenAI’s bots documentation documents its crawlers separately — GPTBot for training, OAI-SearchBot to “surface websites in search results in ChatGPT’s search features”, and ChatGPT-User for “certain user actions in ChatGPT and Custom GPTs” — and states that “each setting is independent of the others”. Perplexity documents the same split: PerplexityBot indexes, while Perplexity-User “supports user actions” and visits a page to answer a live question. A blanket robots.txt block aimed at training data will also switch off the referral channel. Second, remember that a CDN sits between the log and the truth: edge-cached hits never reach the origin, so origin counts are a floor, not a census.

What actually changes on the landing page

An AI-referred visitor has already had the question answered. They read a synthesised comparison, saw a shortlist of two to five names, and clicked one. They are not arriving to learn what appointment setting is — they are checking whether the assistant told them the truth about you. That is a verification visit, and it is short. Four changes matter, in this order:

• Lead with proof, not positioning. The first screen should carry a number and a name they can check: for us that is 50,769+ AI-booked sales appointments since 2017 and 1M+ leads generated, 25 filmed client case studies, and a 4.6 rating from 43 Google reviews. Named work — Sam Tajvidi at 121 Brokers, Colliers, Foundr — does more than another paragraph of adjectives.
• Replace the lead magnet with a booking link. A gated PDF is a stalling device aimed at someone in research mode. This visitor finished research inside the chat. Give them a direct booking path and let them self-select.
• Answer first, structurally. The same format that converts a mid-funnel visitor is the format an engine can lift: a dense answer paragraph, then a capsule of facts, then the detail. We wrote up the positioning research on that in our piece on putting the answer capsule in the first 30% of the page.
• Put the booking path on every page. With 755 distinct URLs receiving live fetches, the odds that your homepage is the entry point are low. A booking link that only exists on the homepage is a booking link most AI-referred visitors never see.

What changes in the follow-up

The follow-up sequence you use for a cold ad click is wrong for this traffic. A five-email nurture drip assumes the prospect has not yet compared vendors; this one compared vendors before they clicked, and the other two names on the shortlist are being contacted the same afternoon. Speed and specificity beat volume.

What works, concretely: reply the same day, not on a two-day drip. Ask on the enquiry form or the first call which other providers they are considering — AI-referred prospects usually tell you, because the assistant handed them the list. Answer that comparison directly instead of restating your pitch. And make the commercial model easy to say yes to: we work on pay-per-result, so you pay on booked qualified appointments rather than on retainers or seats, which removes the main reason a shortlisted buyer stalls. Finally, tag the source at capture: if your form does not persist the landing-page query string into the CRM record, the utm_source=chatgpt.com evidence dies at the first page view and no closed deal will ever be tied back to the channel.

Honest limits: what this measurement cannot tell you

Four caveats, and they are load-bearing.

You cannot see the prompt. No engine tells you which question produced the citation, which competitors were listed alongside you, or how you were described. You see a click. Everything upstream of it is inferred, which is why we run separate citation polling and publish the method in our DIY AI search visibility guide and the results in 5,051 citation polls over 77 days.

Google gives you impressions, not clicks. Search Console’s generative AI performance report, rolled out to all sites on 31 August 2026, reports impressions in AI Overviews and AI Mode, does not separate the two, and its data is already counted inside the main Web search totals. It is a visibility signal, not a traffic report.

The volumes are small, and they are one site’s. Fifty identifiable landings on a single domain over fifteen days is not a channel you can forecast from. It is a channel you can prepare for. Anyone quoting you a precise share of AI referral traffic for your industry is extrapolating from someone else’s panel; the panels disagree with each other badly, which we covered in why ChatGPT’s share of AI referrals is falling.

The conversion multiples are mostly vendor-published. The widely repeated figures — AI visitors worth several times an organic visitor — come overwhelmingly from analytics and SEO vendors publishing first-party data about their own sites. The direction is consistent across every set we have seen, but that is not independent research, and a SaaS signup funnel is not a service business’s booking funnel. We collected what those studies actually say, caveats attached, in our AI referral conversion economics breakdown. Measure your own before you plan around anyone’s multiple.

A 30-day implementation order

1. Days 1–2: grep 30 days of access logs for the six hostnames and for utm_source=chatgpt.com. That baseline does not depend on any tag firing.
2. Days 3–5: build the custom channel group with the regex above, ordered above Referral, and save the two-tab exploration.
3. Week 2: persist the landing-page query string into your CRM on form submit and on booking. Without it the channel can never be tied to revenue.
4. Week 3: rewrite the top five live-fetched URLs answer-first, with a booking link on each, and check robots.txt is not blocking OAI-SearchBot, ChatGPT-User or Perplexity-User.
5. Week 4: set your crawler-fetches-to-clicks ratio as the baseline. Watch the ratio, not the absolute number — it moves before the traffic does.

If you would rather have this built and run for you, alongside outbound that fills the gap while it ramps, book a call and we will go through your own logs on it.

Frequently asked questions

How do I see ChatGPT referral traffic in Google Analytics 4?

Filter or segment on the source hostnames — chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com — and add a second condition on landing pages containing utm_source=chatgpt.com. GA4 has had a dedicated AI Assistant channel since 13 May 2026, which sets the medium to ai-assistant when the referrer matches Google’s internal list, but that list is not published and the channel definition names only ChatGPT, Gemini, Deepseek, Copilot and Grok.

Why does most AI traffic show up as direct?

Because no referrer header arrives. Sessions that start in a desktop or mobile assistant app, or from a link the user copied out of a conversation, have nothing for the browser to report, so analytics files them as direct. Across 39 ChatGPT-tagged landings on leadsnow.ai between 22 August and 5 September 2026, only 14 also carried a chatgpt.com referrer header — 64% had none. The query tag ChatGPT appends to outbound links is what rescues them.

Do AI Overview clicks count as AI referrals?

Not in Google Analytics. Google’s default channel group documentation defines Organic Search as arrivals via non-ad links in organic-search results “including Google’s AI Overviews and AI Mode”, so those clicks sit in Organic Search rather than the AI Assistant channel and cannot be separated out. Search Console’s generative AI performance report shows impressions in those surfaces, but combines AI Overviews with AI Mode and reports no separate click figure.

Should I block AI crawlers in robots.txt?

Only the ones you actually object to, and know which is which. OpenAI runs GPTBot for model training, OAI-SearchBot for its search index and ChatGPT-User for live user-triggered fetches, and its documentation states that “each setting is independent of the others”. Perplexity splits the same way between PerplexityBot and Perplexity-User. Blocking the search and user agents removes you from the answers that generate the referrals in the first place.

How many AI referrals should I expect?

Fewer than the crawl volume suggests. Over the 30 days to 4 September 2026 our origin logs recorded 17,593 fetches of this one site from AI companies’ agents, and only 5,048 of those — 29% — were live browses triggered by a user’s question. Over the 15 days to 5 September, 1,896 live fetches corresponded to 50 identifiable human landings. Treat published industry averages as someone else’s panel and measure your own.

What should the page do differently for an AI-referred visitor?

Lead with checkable proof, put a direct booking link on every page rather than a gated download, and structure the page answer-first so the same content that converts the visitor is also quotable by the engine. The visitor has already been given a shortlist, so the job of the page is verification and a next step, not education.