How do you track leads that come from ChatGPT?

Tracking leads from ChatGPT takes three layers, because no single one of them sees the whole picture. The first is analytics: GA4 now has a native AI Assistant channel that separates chatbot referrals from the rest of your referral traffic. The second is the link tag: ChatGPT appends utm_source=chatgpt.com to many of the links it shows, which survives even when the referrer does not. The third is the buyer: a free-text "how did you hear about us" field on your form, which catches the person who read an answer, closed the tab, and typed your name into a browser two days later.

Run one layer and you will underreport. Analytics misses the sessions that arrive with no referrer at all, which is most of the mobile app traffic. Link tags only exist where the engine chose to add them, and Perplexity, Claude and Gemini do not tag consistently. The self-reported answer catches the invisible journeys and nothing else, and people fill it in inconsistently.

Set expectations before you build any of it. AI referral traffic is a small share of total sessions for most companies, and the leads it produces tend to arrive further along than a cold search visit, because the engine already answered the early questions. Judging the channel on volume will make you switch it off; judging it on what happens after the click is the honest test.

Worth separating this from the other measurement question. Whether you are cited at all is a different exercise, run against prompts rather than against your analytics, and our guide to tracking AI visibility covers that side. This page is about what happens once a person clicks through and becomes a lead you can count.

Why does ChatGPT traffic show up as direct in Google Analytics?

ChatGPT traffic lands in direct because a large share of it arrives with no referrer header for GA4 to read. A referrer is set by the browser when a link is followed from one page to another, and an answer delivered inside a native mobile app, a desktop app or a privacy-restricted context often does not send one. GA4 sees a session with no origin, which is the definition of direct traffic.

The effect is a channel that looks broken in both directions. Direct traffic inflates, so it appears that more people are typing your URL from memory than really are. The AI channel deflates, so the engine that actually produced the visit gets no credit and the report tells you AI is irrelevant. Both numbers are wrong, and the second is the one that gets a programme cancelled.

Two partial repairs exist, and neither closes the gap completely. Link tagging is one: OpenAI began attaching utm_source=chatgpt.com to citation links when ChatGPT Search launched on 31 October 2024, and a tagged link is attributable even with no referrer. Referrer-based grouping is the other, and it works for browser sessions on chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com.

What is left after both repairs is the honest residue: people who read an answer and came to you later, by a route no analytics package can see. That is not a tracking failure to be fixed with a better regex, it is how the channel behaves, and it is the reason the form field in the next section is not optional. Our page on whether AI search sends traffic deals with the related question of how much of this is worth chasing.

What does GA4's AI Assistant channel actually catch?

GA4's AI Assistant channel catches sessions whose referrer matches a recognised AI assistant, and nothing else. Google added it as a default channel group and confirmed it in the Analytics Help Center in May 2026, as Search Engine Journal reported when the channel appeared. Matching sessions get the medium value ai-assistant and are grouped under AI Assistant in the default channel reports, with no configuration needed from you.

Its limits decide what you still have to build. Google's documentation names a specific list of assistants, so a referrer outside that list falls back to ordinary referral. The change is not retroactive, because GA4 processes data daily and does not reclassify history, so your 2025 AI traffic stays where it was filed. Sessions with no referrer are untouched by any of this, since there is nothing to match against.

The practical setup is therefore both: keep the native channel and add a custom channel group of your own for anything it misses, including the utm_source=chatgpt.com tag and the engines Google has not listed. Run them side by side for a month and compare the totals, which tells you how much of your AI traffic the native channel is catching on your particular mix of desktop and mobile.

Be careful about one quiet consequence. Moving sessions out of referral and into a new channel changes your historical comparisons, so a dashboard that showed referral traffic rising will show it falling from the day the channel went live. Annotate the date in whatever reporting tool the business reads, or someone will spend a week investigating a drop that is a reclassification.

How do you set up AI lead tracking in an afternoon?

Setting up AI lead tracking is a short sequence, and the whole thing fits in an afternoon. Start in GA4 by confirming the AI Assistant channel is present in your default channel group report, then create a custom channel group that catches AI referrers and the ChatGPT UTM tag together, so that one definition covers both cases. Name it something the sales team will recognise in a report six months from now.

Next, make sure a lead is actually a countable event. Mark your form submission as a key event, so that AI sessions can be compared with other channels on conversions rather than on sessions. Without that step you can tell someone how many visits ChatGPT sent and nothing about whether any of them were worth having, which is the question you will be asked first.

Then add the field that catches everything else. A single free-text question on the form, asking how the person heard about you, will surface the ChatGPT, Claude and Perplexity mentions that no analytics tool can see, because the buyer remembers the conversation even when the browser forgot the referrer. Keep it free text rather than a dropdown: the wording people use is itself useful, and it feeds the vocabulary work that our methodology treats as the dominant scoring dimension.

Finally, connect the click back to the prompt. Record which of your pages AI sessions land on, then check which prompts cite those pages in your visibility tracking, and you have a rough path from question to page to lead. It is not deterministic attribution and it should not be presented as such, but it answers the question a founder actually asks, which is what people were asking when they found us. Our guide on automating AI visibility reports covers the scheduling side of that.

How much traffic should you expect AI engines to send?

Expect a small but fast-growing share, heavily concentrated in one engine. Previsible's third AI traffic study, covered by Search Engine Land in 2026, analysed 6.77 million AI-driven sessions across 166 websites from November 2024 to May 2026 and found that 92.4% of that traffic came from ChatGPT alone. Monthly ChatGPT sessions in the sample rose from 47,606 to 610,910 over those 19 months.

The same study found Claude growing from 133 monthly sessions to 8,528 and passing Perplexity in March 2026, which is a useful corrective to the assumption that Perplexity is the second engine that matters. For a company deciding where to put a limited amount of effort, those proportions are the argument for starting with ChatGPT, which is the order our page on which AI engine to optimise for first argues for on separate grounds.

Volume can also move on a product decision rather than on anything you did. SE Ranking's analysis of 101,574 websites across 250 countries found ChatGPT referral traffic jumping 36.7% in May 2026, after OpenAI began showing more prominent clickable links inside answers on 7 May 2026. Nothing about those sites changed. The interface did.

Hold both facts at once when you set a target. The channel is growing quickly and the growth is partly outside your control, so a month-on-month traffic goal for AI referrals is a goal you may hit or miss for reasons unrelated to your work. Citation share across a fixed prompt set is the metric you can actually influence, and traffic is the downstream consequence of it.

What should you report to the person paying for the work?

Report four numbers, and state what each one cannot tell you. A single figure in either direction invites the wrong conclusion: traffic alone makes the channel look trivial, citation share alone makes it look like a vanity metric, and neither answers the commercial question on its own. The table below is the reporting set we use, with the limits attached.

Metric	What it answers	Where it comes from	What it cannot tell you
Citation share on a fixed prompt set	Are we named when buyers ask	Repeated prompt runs across engines	Whether anyone clicked
AI Assistant sessions	How many people clicked through	GA4 default and custom channel groups	Sessions that arrived with no referrer
Key events from AI sessions	Did those visits turn into leads	GA4 key events, segmented by channel	Leads that converted on a later visit
Self-reported source on the form	Who says an AI engine sent them	Free-text field, read monthly	Anything, unless enough people answer

Report the volatility alongside the numbers, because AI citation counts swing for reasons that have nothing to do with your site. Trakkr Research tracked 108,650 citation URLs across 10,991 brands and eight models and found that a typical brand's week-over-week citation count swings by 51.8%, which means a single bad week is noise and a trend over six weeks is signal.

Set the cadence to match that. Monthly reporting on a fixed prompt set, with the same prompts every run, gives you a series you can read; ad hoc checks when someone gets nervous give you nothing but the variance. The mechanics of running that series on a schedule are in our guide to measuring AI citation decay.

What else do people ask about tracking leads from ChatGPT?