How to track leads from ChatGPT and Perplexity in your CRM

Where these leads go today

Someone asks ChatGPT which tools solve their problem, clicks one of the links in the answer, reads two pages and books a demo. In web analytics that is a referral session from chatgpt.com, filed beside a partner’s blog. In the CRM it is usually nothing: the form had no UTM to capture, so the contact is created with an empty source or with whatever the last tagged campaign was.

Gemini is worse. Its referrer is a google.com subdomain, and most referrer rules see "google" and call it organic search. The visit is counted, under the wrong heading, and nobody goes looking.

The result is a channel that is growing and cannot be seen. You cannot argue for investing in it, and you cannot tell whether the work you are doing to be named in AI answers produces a single lead.

Two signals, and you need both

The first is the referrer. When an assistant runs in a browser tab, a click carries the assistant’s host, and that host can be matched exactly. The match has to be on the full host. A rule that matches "google.com" will swallow Gemini, and one that matches "microsoft.com" would mislabel every visit from Microsoft’s own site.

The second is the tag ChatGPT adds for you. It appends utm_source=chatgpt.com to the links it cites and sends no medium. That matters more than it looks, because the desktop and mobile apps often send no referrer at all, so the tag is the only evidence. It also sets a trap. Most tracking gives UTMs priority over the referrer, reads a source with no medium, and stores a blank channel. The visit is tagged and still unreportable.

• Match the assistant’s exact host, and check it before the search-engine rules.
• When a link has an assistant as its utm_source and no medium, fill the medium in. When a medium is present, leave it alone: that is a link someone tagged on purpose.
• Normalise the source, so "chatgpt.com" from the tag and "chatgpt" from the referrer are one row in the report.

Getting it onto the contact

Classification in the browser is half of it. The value has to reach the record, in fields you own, as both first touch and most recent. First touch answers "did an assistant introduce us". Most recent answers "did one bring them back". For AI traffic the first question is the interesting one, since an assistant is where a buyer builds the shortlist.

Once it is there, the reporting is ordinary CRM work. In HubSpot, a contact or deal report filtered on your first-touch medium property equal to "ai-assistant", broken down by first-touch source. In Salesforce, the same on the lead or contact field, joined to opportunities through their contact roles. No new object and no new tool.

What this will not see

Be straight with your leadership about the floor. Copilot answers given inside Bing arrive as bing.com, and Grok inside X arrives as x.com; neither can be told apart from search or social. Google’s AI answers on the results page are google.com like any other search click. An assistant that sends no referrer and adds no tag lands as direct. And the buyer who reads the answer, remembers your name and types it in next week is invisible to every method.

So the number you report is a minimum. Expect it to be small today. Judge it on pipeline value and on direction month to month, not on lead count.

How MarqIQ handles it

The tracking script recognises ChatGPT, Perplexity, Gemini, Copilot, Claude, You.com, Phind, Meta AI, Mistral, Poe, Duck.ai, Grok and DeepSeek by exact host, and handles ChatGPT’s utm_source when no medium comes with it. It stores the medium as "ai-assistant" and names the assistant as the source, in the same first-touch and most-recent fields as every other channel, and writes them to the contact. Existing channels classify exactly as before; Google search is still search.

Priority follows the same order as the rest of the script. UTMs in the link come first, which includes ChatGPT’s own tag. UTM values already stored for the visitor come second, and are kept for seven days, so an untagged click from an assistant does not overwrite the paid campaign that brought the person in on Monday. The referrer comes last. First touch is never overwritten by any of it.

It pairs with AI Answer Visibility, which measures the other half: whether the assistants name you when buyers ask. One tells you that you are being recommended. The other tells you that somebody acted on it.

There is nothing to reinstall. The tag already on your site loads the current script, so the channel appears on its own.