Tracking traffic from ChatGPT, Perplexity, Gemini and Copilot

People now find websites through AI assistants as well as search engines: a link in a ChatGPT answer, a cited source in Perplexity, a reference in Gemini or Copilot. That traffic is usually small next to organic search, but it is growing for many sites, and it often converts well because the visitor arrives after an assistant has already narrowed their options. Measuring it takes a little setup, because most analytics tools lump it into "referral" or "direct" by default.

How AI visits identify themselves

There are three signals, of varying reliability:

Signal	Example	Reliability
Referrer domain	chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai	Good when present; missing for many app-based visits
UTM parameters added by the assistant	utm_source=chatgpt.com	Good when present; not all assistants add them
Landing page patterns	Deep pages with long, specific queries behind them	Weak; only a hint

Vendors change domains and behaviour from time to time. Older referrers (for example earlier domains used by the same products) may still appear in historical data. Keep a list and review it every few months against your actual referrer report.

Setting up an AI channel

Most analytics tools let you define custom channels or segments. The idea is the same everywhere: match the source or referrer against a list of AI domains and group them into one channel, placed before the generic "Referral" rule so it takes priority.

A regular expression covering common sources might look like this (adjust to what you see in your own data):

(chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|claude\.ai|you\.com|phind\.com)

In Google Analytics 4, this is done with a custom channel group under the admin settings for data display, using a condition on the session source matching the regex. Custom channel groups apply to historical data as well, which lets you see a trend straight away. In other tools, a saved segment or filter on referrer does the same job.

Test the rule against your referrer report before relying on it. Make sure it does not accidentally capture unrelated domains, and that Google's own organic traffic (including visits from AI Overviews, which arrive as normal Google search) stays in the organic channel.

What goes missing

Not every AI visit carries a referrer or tag. Common gaps:

• Mobile and desktop apps that open links in a way that strips the referrer.
• Copied links pasted into a browser.
• Privacy settings in browsers and extensions that remove referrers.
• Visits that happen later: someone reads an assistant's recommendation, then searches your brand on Google or types your URL. These appear as branded organic or direct visits.

So treat the AI channel as a lower bound. A useful companion metric is branded search impressions in Search Console: if assistants recommend you more often, more people search your name.

What server logs add

Analytics only sees visitors who load the page in a browser with scripts running. Server logs show the requests AI systems make on your site directly:

• User-triggered fetchers such as ChatGPT-User, Perplexity-User and Claude-User fetch a page when an assistant user asks about it or the assistant decides it needs to read it. Each request is evidence that your page was used in an answer, even if nobody clicked.
• Search crawlers such as OAI-SearchBot, PerplexityBot and Claude-SearchBot show which pages are being indexed for AI search, and how often.
• Training crawlers such as GPTBot and ClaudeBot show what is being collected for model training.

Counting user-triggered fetches per page over time is one of the more direct measures of AI usage available to site owners. Verify the IP addresses against the ranges vendors publish, since user-agent strings are easy to fake. The log analysis guide covers the method, and the AI crawlers guide lists the agents.

Reading the numbers

• Landing pages: which pages do AI visitors land on? They are often deep pages answering specific questions, product pages, or comparison pages. Those are the pages assistants find useful.
• Engagement and conversion: compare AI visitors with organic visitors on the same pages. Many sites see AI visitors spend less time browsing but convert at a similar or higher rate, because they arrive with a narrower intent. Measure your own rather than assuming.
• Trend: track the channel monthly. Short-term swings are common when vendors change how links are displayed.
• By assistant: break the channel down by source. Different assistants send very different volumes depending on your audience and market.

A worked example

A hypothetical online retailer of climbing equipment sets up an AI channel and looks at three months of data. The channel accounts for a small fraction of all sessions. Almost all of it lands on two kinds of pages: rope and harness comparison guides, and a page explaining how to retire old ropes safely. Conversion rate for the channel is close to organic search on the same guides, but the share of visitors who go straight to a product page is noticeably higher. In the server logs, the rope retirement page receives a steady stream of user-triggered fetches from two assistants, several times more fetches than clicks.

The conclusions are modest and useful. The comparison guides and the safety page are what assistants rely on, so they get priority for updates and fact-checking. The gap between fetches and clicks shows that the safety page is being used in answers far more often than analytics alone suggests. And the channel is small enough that a single month's change should not drive decisions; the retailer reviews it quarterly alongside its AI visibility checks.

Privacy-friendly and server-side analytics

Sites that use cookieless or server-side analytics see AI referrals in the same way, through the referrer and UTM parameters, and can build the same grouping. Where consent banners block analytics until a visitor agrees, a share of AI visitors, like all visitors, will not be measured at all. Server logs are unaffected by consent and scripts, which makes them a useful cross-check on the trend, as long as personal data in them is handled appropriately.

Reporting it

For managers and clients, show the AI channel as its own line next to organic search, with the caveat that it is a lower bound. Include the top landing pages and, if you track it, the share of brand-free assistant questions that mention you. That combination answers the question people actually ask, "are we showing up in AI, and does it bring business?", without overstating a small number.

Connecting traffic to visibility

Traffic tells you what happened after a citation; it does not tell you how often you are mentioned or recommended. For that, run a regular set of brand-free questions through the assistants your customers use and record mentions and citations, as described in measuring AI visibility. DeskSEO's AI visibility feature runs such a question set through one model, Anthropic's Claude, each time you start a run, and stores the answers. Put the two side by side: if mentions rise and AI referrals rise with them, the work is reaching people.

A setup checklist

1. Create an AI channel or segment using the referrer regex; place it before the generic referral rule.
2. Check your referrer report for AI sources not yet in the list.
3. Add a monthly view: sessions, conversions and top landing pages for the AI channel.
4. Track branded search impressions in Search Console as an indirect signal.
5. Count user-triggered AI fetches per page in server logs, with IP verification.
6. Review the domain list every quarter.

Common mistakes

• Counting Google AI Overview clicks as AI traffic. They are Google organic and cannot be separated in analytics.
• Assuming no referrer means no AI influence.
• Over-reading one month of data from a small channel.
• Forgetting to exclude internal testing: your own team asking assistants about your site and clicking the links inflates the numbers.