How to Measure AI Referral Traffic When UTMs Are Stripped
AI assistants strip UTM parameters, so referrer segmentation is the only reliable method. The setup, the domains to watch, and how to read the numbers.
Published 2026-08-18 · 4 min read
Traffic from AI assistants cannot be tracked the way campaign traffic can. Most assistants strip query parameters when a user clicks a cited source, which means the UTM scheme that works everywhere else returns nothing here. Referrer-based segmentation is the fallback, and it needs setting up deliberately rather than discovered later.
This is the setup, plus how to interpret numbers that will always undercount.
Why UTMs fail here
- The link is generated, not stored. A cited URL is often reconstructed from retrieved content rather than passed through from your page, and reconstruction drops query parameters.
- Privacy and cleanliness. Some assistants deliberately strip tracking parameters from links they present.
- There is no campaign to tag. You did not place the link, so there was no opportunity to add parameters. This is the fundamental difference from paid or email traffic.
The referrer domains to watch
| Referrer | What it represents |
|---|---|
| chatgpt.com | ChatGPT answers with cited sources |
| perplexity.ai | Perplexity, whose entire model is cited answers — usually the highest-volume AI referrer |
| claude.ai | Claude conversations |
| gemini.google.com | Gemini, and note that Google's AI Overviews may also arrive via google.com |
| copilot.microsoft.com | Microsoft Copilot |
| you.com | You.com |
| grok.com / x.com | Grok, which can arrive from either domain |
Setting it up
- 1
Create a saved segment for AI referrals
In whatever analytics you use, build a segment matching the referrer hostnames above. Save it — this is the only way to get a trend rather than a one-off number.
- 2
Build a second segment for 'no referrer'
A meaningful share of AI-originated traffic arrives with no referrer at all, and lands as direct. You cannot attribute it, but you can watch the direct bucket for unexplained rises that correlate with being cited.
- 3
Record a baseline before you change anything
If you permit retrieval crawlers or publish new content, you need the prior number to see any effect. Most people skip this and can never tell whether anything worked.
- 4
Log the landing pages, not just the sessions
Which pages AI traffic arrives on tells you which content is being retrieved. That is the actionable output — it shows you what to write more of.
- 5
Set a monthly cadence
Volumes are small early and vary run to run. A monthly check over several months shows a trend; a daily check shows noise.
What the numbers actually tell you
| Observation | Reading |
|---|---|
| AI referrals exist at all | Retrieval is reaching you. Permissions and crawlability are working |
| Zero AI referrals despite citations | Users are reading the answer and not clicking — common, and not a failure |
| Referrals concentrated on one or two pages | Those are your retrievable assets. Write more like them |
| Referrals rising but no signups | The traffic is real and the landing experience is the problem |
| A jump in direct traffic with no other change | Possibly unattributed AI traffic. Note it and correlate with any citation you know about |
| High AI referral volume, low conversion | Normal early — this traffic is exploratory. Judge it over months, not a week |
The undercounting problem
Every method here underestimates. Three separate leaks:
- Referrer stripped. Some traffic arrives with no referrer and lands in direct.
- Answer read, no click. Most assistant interactions end without a visit. Your product may be influencing decisions that never appear in analytics at all.
- Sampling and privacy modes. Private browsing and consent settings remove referrers by design.
The practical consequence is that AI referral counts should be treated as a trend indicator, never as a total. Doubling from four visits to eight is meaningful as direction and meaningless as volume. Do not build a business case on the absolute number.
Complementary signals
- Manual citation checks. A fixed set of category questions asked of each assistant monthly, recording whether you appear and how you are described. Slow, and the only direct measure of visibility.
- Crawler hits in server logs, filtered by user agent. Shows whether retrieval is reaching you at all — and if it is not, no writing improvement will help.
- llms.txt fetch counts in logs. Crude but a real indicator of machine interest.
- Mentions in communities. People saying "an assistant recommended X to me" is anecdotal and worth collecting.
You cannot measure AI visibility precisely, and any tool claiming to is measuring something narrower than it implies. What you can do is watch a trend and know which of your pages get retrieved — which is enough to decide what to write next.
More in SEO & GEO
- Entity Consistency: Why AI Models Fail to Attribute Your ProductWhen a model cannot tell your mentions describe one product, it attributes none of them. The signals that resolve an entity, and how to audit yours.
- The 10-Page SEO Starter Kit for a New AI ProductThe ten pages a new AI product should have before writing anything else, what each one is for, and the order to build them in.
- Backlinks for AI Startups: What Works When You Have No AuthorityEight link tactics ranked by effort and payoff for a domain with no authority, and the three that look productive and are not.