More readers now reach articles and docs through an AI assistant: they ask ChatGPT, Perplexity or Gemini a question, and follow a link in the answer. Many site owners suspect this traffic is bigger than their analytics shows, because "direct" keeps growing on pages nobody would type into an address bar.
This post explains why AI referrals go missing, what analytics can still catch, and how to ask readers the question analytics can't answer.
The short answer
Use your analytics for the visits it can attribute, then ask readers about the rest. Add one multiple choice question, "How did you find this page?", with an option such as "AI assistant (ChatGPT, Perplexity, Gemini...)". It needs no cookies or tracking, and it catches visits that arrive without a referrer. The answers are self-reported and approximate: use them to see a rough share and how it changes, not to count visits.
Why AI assistant visits end up in "direct"
Analytics tools work out where a visit came from mostly from the Referer header, which the browser sends with the address of the previous page. When that header is missing and the link has no campaign tags, the visit has no source. Google Analytics files it under Direct, which its documentation describes as users who arrive through a saved link or by typing your URL.
An AI assistant visit can lose its referrer in several ordinary ways:
- The reader taps a link in the assistant's desktop or phone app, which opens the browser without a referring web page.
- The link carries
rel="noreferrer", which tells the browser to leave the header out. - The reader copies the link from the answer and pastes it into a new tab.
- The reader reads the answer, then types your site name or searches for the page, so the visit is credited to direct or to search.
None of these show up as an error. The visits are simply counted under the wrong source, or under none.
What your analytics can still catch
Before you ask readers anything, use what your tools already record. Each signal catches some AI visits and misses others:
| Signal | What it catches | What it misses |
|---|---|---|
| Referrer | Visits where the browser passed on the assistant's address. Since May 2026, Google Analytics groups these in an **AI Assistant** channel. | Visits with no referrer, and Google's own AI Overviews and AI Mode, which that channel leaves out. |
| Campaign tags | Links the assistant tagged itself, such as `utm_source=chatgpt.com`. | Untagged links. Assistants decide whether to tag, and can change it without notice. |
| Direct visits to deep pages | A hint: few people type the address of a long article, so a rise there needs explaining. | Where the visits came from. It could be AI, email apps, chat apps or bookmarks. |
| Asking readers | How the reader says they found the page, whatever the browser sent. | Readers who skip the question, and answers people misremember. |
Use the first three for the visits they can attribute. Asking is for the gap they leave.
Ask "How did you find this page?"
Add one multiple choice question to the survey on the pages you care about. Keep the options short and in words readers use:
- Search engine (Google, Bing...)
- AI assistant (ChatGPT, Perplexity, Gemini...)
- Social media
- A link on another website
- Newsletter or email
- Someone sent me the link
- I don't remember
Put it second, after a one-tap Yes / No question such as "Did this page answer your question?". The first question gets the most answers, so keep it for whether the page helped. Readers who have tapped once are more likely to answer a second question.
Keep the wording and options the same for as long as you want to compare results. If you add "AI assistant" halfway through, the weeks before and after are no longer comparable.
Tip: Readers who clicked a link in Google's AI Overview will usually pick "Search engine". That is fine: this question is about AI assistants they chose to ask, and Search Console covers Google.
How far to trust self-reported answers
Readers answer from memory, and memory is rough. Treat the results as an estimate with known weak spots:
- Routes blur. A reader who asked ChatGPT, then searched for your site by name, may pick either option.
- Not every reader answers, and the ones who do are not a random sample. The share of "AI assistant" among those who answered is not the share among all your visitors.
- "I don't remember" is a real answer. Keep it, so readers who don't know don't guess.
- Small numbers swing. Wait for a few dozen answers on a page or section before you read anything into the split.
The answers are most useful compared with themselves: the same question, on the same pages, month after month. If "AI assistant" goes from a few answers to a large share while direct traffic climbs in analytics, the two together make a fair case.
Where to ask, and for how long
You don't need the question on every page forever. Pick where it will tell you something:
-
Start with the pages where direct is growing
Usually long guides, reference pages and how-tos. Ask on that section, for example every page under
/guides/. -
Run it for a fixed period
Six to eight weeks is often enough to see the split on a busy section. A quiet site needs longer.
-
Repeat it later with the same wording
Ask again in three or six months and compare. That shows the trend, which matters more than any single figure.
Example
Example: a tutorial site with growing direct traffic
Say you run a site of programming tutorials. Over a year, direct traffic to your troubleshooting guides has roughly doubled, while the AI Assistant channel in your analytics shows only a trickle. Nobody types the address of a guide called "Fix a failed database migration".
You add "How did you find this page?" as the second question on every page under /troubleshooting/ for eight weeks. It collects 160 answers: 70 pick "Search engine", 48 pick "AI assistant (ChatGPT, Perplexity, Gemini...)", 22 pick "I don't remember" and the rest are spread across the other options.
You can't say how many visits came from AI assistants. You can say that AI assistants are a regular route to these guides, much larger than the referral channel suggests. You then check whether the guides answer the follow-up questions a reader brings from an assistant's short answer, and you plan to ask again in six months.
Where Asklet fits
Asklet shows a short survey on your pages from one snippet, with multiple choice questions of 2 to 12 options. Results show how often each option was picked, per website, per page and per path, for the dates you choose. The Free plan allows 3 questions per survey, enough for a Yes / No question, "How did you find this page?" and an optional comment.
It sets no cookies and uses no tracking identifiers. It never sees how a reader reached your page: it gets the path of the page the widget is on, without the query string, so it doesn't read campaign tags either. The reader's answer is the only source. Results are per question, so Asklet can't split the Yes / No answers by how readers found the page.
Your next steps
-
Check the AI Assistant channel and campaign tags first
Note what share of visits your analytics already attributes to AI assistants.
-
Find the pages where direct traffic is growing
Deep pages with rising direct visits are the best place to ask.
-
Add "How did you find this page?" as the second question
Seven short options, including "AI assistant" and "I don't remember".
-
Run it for six to eight weeks without changing the options
Then read the split per section, not per page.
-
Compare it with your analytics for the same dates
A large "AI assistant" share alongside rising direct traffic is a fair signal.
-
Ask again in a few months
Same wording, same pages, so you see the trend.
Try these steps on your own site. Asklet adds the questions to every page with one snippet. Free plan, no card needed. Compare the plans.
Start freeCommon questions
Does Google Analytics show ChatGPT traffic?
Partly. Since May 2026 it groups visits whose referrer matches a known AI assistant in an AI Assistant channel. Visits that arrive without a referrer still count as Direct, and Google's AI Overviews and AI Mode are not in that channel.
Is asking readers accurate?
Roughly. Readers misremember, and the readers who answer are not a random sample. Use the answers to see whether AI assistants are a real route to your pages and whether the share is growing, not to count visits.
Do I need a cookie banner to ask this question?
Not for Asklet. The widget sets no cookies and uses no tracking identifiers. Its only browser storage is a short "already answered" note so the same reader isn't asked again on that page. Your own analytics may still need consent, depending on what it sets.
Should "AI assistant" include Google AI Overviews?
Keep them apart. Readers think of an AI Overview as part of a Google search, and Search Console already reports search clicks. "AI assistant" works best for tools readers chose to open and ask.
Related
Sources
More from the blog
- Privacy and law EU cookie consent changes: where the Digital Omnibus stands in October 2026 9 min read
- Privacy and law Session replay and pixel lawsuits: what website owners should know 9 min read
- Privacy and law How to make a feedback widget accessible, and what the European Accessibility Act asks 9 min read