The Short Answer
To track traffic from ChatGPT, count landings whose address carries the tag ChatGPT adds to its links, utm_source=chatgpt.com, and stop relying on the chatgpt.com referrer, which most visits no longer send. Then add two measurements analytics cannot give you: the pages ChatGPT reads on behalf of users, from your server log, and the people who use your product through ChatGPT without ever visiting. Clicks are the smallest of the three. We learned all of this the hard way in October 2026, when a tracking bug made ChatGPT traffic on three of our products drop by 60 to 100 percent in a single morning.
The one sentence version
Count the tag, verify the agent in your logs, and measure the people who never click, because ChatGPT sends far more attention than it sends visits.
The Three Things People Mean by ChatGPT Traffic
When a team asks how much traffic ChatGPT sends, they usually mean three different things at once, and each lives in a different place. Mixing them is why reports disagree.
| What it is | Where you see it | What it tells you |
|---|---|---|
| A person clicks a link in a ChatGPT answer | Analytics, by the utm_source tag | Visits and conversions from answers |
| ChatGPT reads your page to answer someone | Your server log, the ChatGPT-User agent | Which pages ChatGPT uses for live answers |
| ChatGPT indexes your page for later search | Your server log, the OAI-SearchBot agent | Whether you are in its search index |
| A person uses your product inside ChatGPT | Your own app data, by connector or key | Usage that never becomes a visit |
Only the first row is traffic in the analytics sense. The second and third are evidence that you are being considered and used. The fourth is the fastest growing part on some products, and it is invisible to every web analytics tool. Our guide to tracking brand citations in ChatGPT covers a fifth thing, whether answers mention you at all, which needs a different method again.
Step 1: Count the Tag, Not the Referrer
Links in ChatGPT answers carry utm_source=chatgpt.com in the address. Analytics tools read that tag and attribute the session to ChatGPT, regardless of what the browser sends as the referrer. The referrer is the weaker signal. In one week on one of our sites, 244 requests arrived with a chatgpt.com referrer and not one of them was a page load; they were small background files a browser fetches after it has landed. The real landings that week, identified by the tag, numbered three.
| Signal | Reliability | Use it for |
|---|---|---|
| utm_source=chatgpt.com in the landing address | Good, when present | Counting visits and conversions |
| chatgpt.com in the referrer header | Poor, often missing | Nothing on its own |
| ChatGPT-User in the user agent | Good, once verified by address | Pages ChatGPT reads live |
| Direct traffic spikes | Very poor | A hint that tags went missing |
Bots copy the tag
One of the three tagged landings in that week was a bot announcing itself in its user agent. Check the user agent of tagged landings in your raw logs before treating every one as a person.
Step 2: Set Up a Channel for AI Assistants
In Google Analytics 4 you can define a custom channel group so assistant traffic stops hiding inside referral or direct. Create a channel for AI assistants with a rule on the session source, matching chatgpt.com and the other assistants you care about, and place it above the referral channel so it wins. The exact menu names move between releases, so follow the current help page for custom channel groups. What matters is the rule: source, not referrer, and the tag value as the source.
Create the channel group
Copy the default channel group so the original stays intact, then add a channel named AI assistants.
Match on session source
Use a condition that the session source matches chatgpt.com, and add other assistants by their tag values.
Order it above Referral
Channels are evaluated in order, so put AI assistants before Referral and Direct.
Build a landing page report
Landing page by channel shows which pages assistants send people to, which is often not your homepage.
Check it monthly against the logs
Compare tagged landings in analytics with tagged landings in the server log, to catch tracking that silently breaks.
What Happened When the Tag Disappeared
From 9 to 15 October 2026, links clicked in ChatGPT answers stopped carrying the tag. A newer model had reached free users the day before, and its links went out without it. A fix rolled out from 15 October, as Search Engine Journal reported. On our side it looked like this:
| Product | Before 9 Oct | 9 to 15 Oct | 16 to 21 Oct |
|---|---|---|---|
| Product A, sign ups from ChatGPT | about 78 a day | 32 a day | 120 to 126 a day |
| Product B, visits from ChatGPT | about 83 a day | 21 a day | about 63 a day |
| Product C, ChatGPT arrivals | 4.9 a day | 0 | about 1 a day |
Three unrelated products losing the same channel on the same morning was the clue that the measurement had broken, not the market. The second clue was the server log: on the site where we count ChatGPT live reads, the reads held steady that week while tagged visits fell. When an independent signal holds and the tagged one drops, the visitors are still arriving and being filed as direct. Mark that week in every chart that includes ChatGPT traffic, and never compare a week inside it with a week outside it.
Step 3: Read the ChatGPT Agents in Your Logs
OpenAI documents several agents, each with its own user agent and a published list of the addresses it uses. Two matter for traffic questions. ChatGPT-User fetches a page live because a person asked something that needed it. OAI-SearchBot crawls pages for ChatGPT search. GPTBot collects training data and says nothing about today's answers. Our list of AI crawlers covers each in detail.
| Agent | Why it visits | What a rise means |
|---|---|---|
| ChatGPT-User | A person asked a question that needed your page | Your page is being used in live answers |
| OAI-SearchBot | Building the ChatGPT search index | You are being considered for future answers |
| GPTBot | Collecting training data | Little about current answers |
Count only successful fetches of real pages, and verify the addresses against the published lists, because scanners wear these names too. On one of our sites in a single week, ChatGPT-User made 214 requests and 135 were successful page fetches; the rest were redirects and errors, some of them scanners probing for configuration files. Successful live fetches by page are the closest thing to a list of pages ChatGPT actually uses, and they rarely match the pages Google shows most. If ChatGPT never reads your pages at all, our guide on why ChatGPT is not showing your site goes through the causes in order.
Behind a CDN, restore the real address
If your site sits behind a content delivery network, your origin log may record the network edge instead of the visitor. Verification against published address lists needs the real visitor address restored at the origin first.
Step 4: Measure the People Who Never Click
If your product can be used inside ChatGPT, through an app or a connector, a growing share of its users may never visit your site. As of 13 October 2026, 129 people used one of our products through ChatGPT and only 14 of them had ever opened its website. Analytics would have counted the 14. The product's own data counted all 129 and thousands of tool calls.
Count by key or connection
Every person who connects your product creates a key or session. That is your user count for this channel, not page views.
Log calls by client
Record which client made each call, so assistant usage is separate from automations and your own website.
Keep the in chat version honest
OpenAI rules forbid upsells and checkout links inside ChatGPT, so the free experience there is the whole impression most of these users get.
Our guide to getting your product into ChatGPT explains the submission side. Once you are in, this is the measurement that tells you whether it is working.
How Big Is ChatGPT Traffic, Honestly?
Smaller than its influence, on most sites. On our own agency site, in the week of 6 to 12 October, ChatGPT's live agent made 135 successful page fetches on behalf of people asking questions. In the same week three landings carried the ChatGPT tag, and one of those was a bot. That is roughly 45 reads for every tagged visit. Most people read the answer and never click, and an answer built partly from your page still shapes what they think of you.
So judge ChatGPT by all three layers together. A page with many live reads and few clicks is doing work for you that analytics will never credit. A page with clicks but no live reads is probably being reached through links in answers that came from somewhere else. And a product with heavy usage inside ChatGPT can be growing while its referral line looks flat.
Other Assistants Work the Same Way, Mostly
Every assistant labels its links differently, and some send no tag at all, so the same method applies with different values. Before you add an assistant to your channel rule, check your own landing log for what its visits actually carry: a tag, a referrer, a recognisable in app browser name, or nothing. Write that down, because it changes. ChatGPT's October bug was not announced in advance, and the next change from any assistant will not be either. A five minute monthly look at raw landings is the cheapest insurance against reporting a collapse that never happened. Keep the notes next to the report, so whoever reads it next month knows which numbers changed because the world changed and which changed because a label did.
A Simple Monthly Report
| Line | Source | Why it is on the report |
|---|---|---|
| Tagged ChatGPT landings and conversions | Analytics, AI assistants channel | The visits answers send |
| Tagged landings checked in the raw log | Server log | Catches broken tags and bots |
| ChatGPT-User successful fetches by page | Server log, verified | What ChatGPT reads for live answers |
| OAI-SearchBot fetches | Server log, verified | Whether you are being indexed |
| People using the product through ChatGPT | Product data, by key | Usage that never becomes a visit |
Five lines, three sources, one page. It takes about an hour a month once the queries exist, and it answers the question executives actually ask, what is ChatGPT doing for us, without pretending the referral line is the whole story.
Common Mistakes
Counting the chatgpt.com referrer
Most visits no longer send it, and background requests that do send it are not visits.
Comparing across a tracking break
A week with missing tags compared with a normal week shows a fall that never happened.
Trusting agent names
Scanners wear ChatGPT agent names. Count successful fetches from verified addresses only.
Reporting clicks as the whole value
Live reads and in chat usage are larger than clicks, and invisible in analytics.
The Bottom Line
ChatGPT traffic is real, small and badly labelled. Count it by the tag, check it against your server log, and add the two numbers analytics will never show: the pages ChatGPT reads for people, and the people who use you without visiting. When the tagged line drops on every product at once, suspect the measurement before the market, and look at the log before you change a single page. The October bug proved how fragile a single tag is, and how much a second, independent signal is worth when it breaks.