How to tell if your traffic comes from ChatGPT, Claude or Gemini

A customer mentions on a sales call that they found you by asking ChatGPT. You open GA4 to see how many more came in that way and there isn't much to look at: a new channel called AI Assistants with fourteen sessions this month, a referral from perplexity.ai with three, and a direct-traffic line that grew without you changing anything.

Those three things are connected, and each one is measured differently. This article pulls apart three questions that usually arrive tangled together: how to know which people landed on your site from an AI answer, how to know when the engines themselves are reading your pages, and whether that traffic buys more than the rest, as people keep saying it does.

Where the people who arrive from an answer come in

When someone clicks a link inside ChatGPT, Gemini, Perplexity or Claude, the browser usually sends a referrer, the address the visit came from. These are recognizable domains: chatgpt.com, gemini.google.com, perplexity.ai, claude.ai, copilot.microsoft.com. ChatGPT also appends a utm_source=chatgpt.com parameter to the links it outputs, though not always and not on every surface; that behavior has changed several times since 2024.

Until recently all of that landed in GA4's Referral channel, mixed in with the link a blog or a directory gave you. On May 13, 2026, Google added a new channel to the default channel group: AI Assistants. When the referrer matches an assistant Google recognizes, the session gets medium ai-assistant and campaign (ai-assistant), with nothing for you to configure. The rollout was staggered and finished reaching most properties in early June.

Three things worth knowing about that channel:

• The list of recognized assistants isn't published in full. Google names ChatGPT, Gemini, Copilot, DeepSeek and Grok as examples. Perplexity, as of this writing, still lands in Referral.
• It doesn't reclassify history. Anything before May 13 stays where it was.
• Visits from the same assistant can end up split across AI Assistants, Referral, Direct and Unassigned, depending on how the referrer traveled.

What GA4 can't see

Two big holes.

The first is Google itself. Clicks from an AI Overview or from AI Mode arrive with a google.com referrer, and GA4 counts them as Organic Search, same as a classic blue link. Search Console doesn't separate them either: AI Overview clicks are rolled into the Web search type. Today there is no way, with Google's own tools, to know what share of your organic traffic passed through a generated answer. All you can do is infer it: if a page holds its impressions while its CTR drops, it's likely showing up inside an answer that resolves the query without a click.

The second is referrer-less traffic. When the click comes from a native app (ChatGPT on a phone or on the desktop), from an in-app browser, or passes through a strict referrer policy, the referrer doesn't travel and the session lands in Direct. Loamly analyzed 446,000 AI-sourced visits in February 2026 and found 70.6% arrived with no referrer. Attrifast measured in May that ChatGPT's web interface passes a referrer on roughly 28% of clicks, and the desktop app on 6%.

Take those as orders of magnitude and what you see in the AI Assistants channel is about a third of what actually arrives. Then add the long path: someone reads a recommendation in ChatGPT, doesn't click, and later searches your brand on Google or types the URL. That visit is AI traffic in any reasonable sense, and in GA4 it's branded search or direct.

How to set it up, in order

1. Look at the AI Assistants channel, but don't stop there. Reports, Acquisition, Traffic acquisition, filtered by channel group. That number is the floor, not the total.

2. Build a custom channel group for what's missing. Under Admin, Channel groups, create a rule where source matches a regex along the lines of chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|grok\.com|deepseek\.com. Place it above Referral so it wins priority. Unlike Google's channel, custom groups do apply retroactively, so you get history.

3. Look by landing page, not just by source. AI links deep: to a product page, a comparison, an article. If you only look at the source you miss which pages are being cited and recommended, and that's the piece of information that tells you what content to produce next.

4. Watch direct and branded search. If the AI Assistants channel grows and in the same month Direct and queries containing your brand name in Search Console grow too, that's not a coincidence. You won't be able to attribute it precisely, but the joint movement is the signal.

5. Ask. A how did you hear about us field in the checkout or on the demo request form, with ChatGPT or another AI as an option, is still the most honest method there is. It has recall bias, but it's the only one that catches the long path.

When the visitor is the engine

There's a second kind of visit you will never see in GA4, because GA4 measures with JavaScript and bots don't run it. These are the AI companies' crawlers reading your pages. They show up in server logs, and it helps to know what to look for.

It isn't one bot, it's three

Each company runs several agents for different purposes, and each one identifies itself with its own name in the user-agent:

Perplexity states it doesn't train models on what it crawls. OpenAI's names are in its crawler documentation; Anthropic and Perplexity publish their own. Other names you'll see: Meta-ExternalAgent (Meta, training; in September it was the most active AI bot on Cloudflare's network), Bytespider (ByteDance) and Applebot (Apple).

The third row of the table is the one that matters most for marketing. When you see ChatGPT-User or Claude-User in your logs, it means that at that moment a person asked the assistant something and the assistant went to read your page to answer them. That's not background indexing: it's a real query on your topic, with your site as a source. These agents generally don't honor robots.txt, precisely because they act on behalf of a user. If you want to limit them, do it at the CDN or WAF, not in robots.txt.

Google is a separate case. It has no dedicated AI bot: the same Googlebot feeds classic results, AI Overviews and AI Mode. Google-Extended, which plenty of people search their logs for, isn't a crawler but a robots.txt token that controls whether your content is used to train Gemini. You won't find a single log line with that name, however hard you look.

Where to look and how to verify

• If you have server access: access logs filtered by user-agent. A grep -i "gptbot\|claudebot\|perplexitybot\|chatgpt-user\|claude-user" access.log already gives you a first picture of who comes by and which paths they hit.
• If you're on Cloudflare: the bots section (AI Crawl Control) lists AI agents by name, with request counts and the pages they fetched. For most teams it's the most accessible route, and you don't need to read raw logs.
• If you're on Shopify, Webflow or a similar platform without log access: either put Cloudflare in front of the domain, or accept that you won't see this layer.

How much they read and how much they send back

Cloudflare Radar publishes, for each AI company, the ratio between pages crawled and visits referred. September 2026:

AI bots were 5.3% of all requests Cloudflare classified that month. And the share dedicated to search and answering, the second and third rows of the first table, went from 2.6% of AI-bot requests in January to 10.6% in September. That's the part you care about: being read to train a model brings you nobody; being read to answer a question does.

The practical reading: seeing a lot of GPTBot in your logs doesn't mean ChatGPT is recommending you. It means ChatGPT is reading you. The ratio between how much they read and how much they send is the number worth tracking month over month. And if you're going to block something, block training and let search and user agents through: blocking OAI-SearchBot removes you from ChatGPT's answers, which is the opposite of what you want.

Does that traffic buy more?

The intuition says yes: if someone lands on your store after asking an AI, they've already researched, already compared, and they're here to close. The 2026 data says the intuition is right on average, with three caveats that change how you should use it.

What the numbers say

Adobe measures more than one trillion visits to US retail sites. In July 2026, AI-referred visits converted 60% higher than non-AI visits, generated 53% more revenue per visit, lasted 59% longer, bounced 33% less and added 28% more items to cart. It was the eleventh consecutive month in which AI traffic out-converted the rest.

Shopify, for Q1 2026, reported conversion close to 50% higher than organic, average order value 14% higher, and one figure that explains much of the rest: more than half of AI-referred sessions start on a product page, against 20% for organic.

Outside commerce, Semrush measured a 4.4x conversion premium across 500+ B2B topics (June 2025), and Ahrefs published that on its own site 0.5% of sessions came from AI and produced 12.1% of signups.

There is data pointing the other way, and it's worth keeping in view. In September 2025, BrightEdge analyzed Fortune 100 brands and found organic still converted better. And Adobe itself, in March 2025, was measuring AI traffic converting worse than everything else. The advantage appeared and widened over 2025 and 2026; it isn't a fixed property of the channel, it's a snapshot of today.

Why it happens

The research happened inside the chat. The person asked, compared options, narrowed by price or by feature, and the click to your site is the last step, not the first. Shopify's figure confirms it: they land on the product page, not the homepage. In that sense AI traffic looks far more like branded search than like a discovery visit, and that's the yardstick to measure it against.

The three caveats

1. There isn't much of it. Conductor estimated for 2026 that AI referrals are 1.08% of a site's total visits, growing 527% year over year. A high rate on a small base yields few sales in absolute terms. With 10,000 visits a month and a 2% organic conversion rate, organic leaves you about 200 sales; AI traffic, even converting at double the rate, about four. It's useful for understanding the channel's quality, not for moving budget yet.

2. Part of the edge is selection, not magic. The AI answers the question on screen. Across several 2025 and 2026 measurements, more than 80% of searches with an AI Overview end without a click, and in AI Mode the figure passes 90%. The people who do click are the ones for whom the answer wasn't enough: they want to see the product, the price, the stock. That's a self-selected high-intent group. The rate is high partly because the lukewarm stayed in the chat.

3. What you measure is a third, and a biased third. Around 70% of AI visits arrive without a referrer. The AI Assistants segment you're analyzing is the minority that did pass it, mostly from the desktop web. You don't know whether the mobile app behaves the same way. And the biggest influence of all, the person who read the recommendation and then searched your brand, is counted as branded organic or direct, and you're crediting another channel for it.

What to do with this

• Compare AI traffic's conversion rate against branded search, not against organic as a whole. If it performs about the same, that's consistent with what it is: people who had already decided most of it.
• Look at where they land. If the AI sends people to a product page with no visible price, unclear stock or no shipping information, the high intent is wasted in the first second.
• Track the channel's monthly trend alongside direct and branded. The absolute number is small; the joint movement is the data point.
• If at some point you blocked AI bots over traffic or server cost, review which ones. What they don't read, they don't recommend.

What you have today is three partial tools: a GA4 channel that sees a third, logs that see the bots but not the people, and a form that depends on the customer's memory. None of them is enough alone. All three together, reviewed every month, already tell you considerably more than most of your competitors know about theirs.

All of this measures traffic that already arrived. The question before it, whether your brand shows up at all when someone asks ChatGPT about your category, is answered from the other side: with an AI Visibility scan across your category's questions, or with the GEO Observer extension, which captures real sessions. And if you're new to the topic, what GEO and AEO are explains the three layers of the problem.

---

The GA4, Cloudflare and Adobe figures come from their publications between May and October 2026 and may change. Agent names and their behavior toward robots.txt are as documented by each company as of October 2026. The 10,000-visits, 2%-conversion example is illustrative, not a customer's data.