Most vendor research on LLM referrals stops at traffic volume. This dataset goes further. It covers 13 months of Google Analytics data across an unnamed customer base, January 1, 2025 to February 7, 2026, tracking both referral volume and conversion events tied to real business outcomes.
Table of contents
- What this is about
- The design
- The numbers
- What could still be wrong
- What to do on Monday
- FAQ
- Key takeaways
Key takeaways
- Across the dataset (January 1, 2025 to February 7, 2026), LLM referrals averaged less than 2% of total referral traffic, with individual sites ranging from 0.15% to 1.5%. The number of sites behind that average is not disclosed.
- From H1 2025 to H2 2025, average LLM referral growth was 80%, though site-level results ran from 10% to 300%. Again, no site count is stated.
- The reported conversion rate for LLM-referred visitors was approximately 18%, described as higher than paid shopping, SEO and PPC channels. The base count and confidence interval are not disclosed.
- Citation sources inside LLM responses shifted noticeably during the period, with YouTube and Reddit both gaining share; the monitoring covered more than 5 000 prompts across Gemini, ChatGPT, and Perplexity APIs.
- Because the dataset is a vendor aggregate, segment-level conclusions cannot be drawn from it: not by industry, not by site size, not by conversion type.
What this is about
Jason Tabeling published an analysis of LLM prompt referral traffic drawn from Google Analytics data across his company’s customer base, covering January 1, 2025 to February 7, 2026. The headline finding is an approximate 18% conversion rate for LLM-referred visitors — higher than any other channel reported — against a traffic share of less than 2% of total referrals.
The design
What was measured: Referral sessions from four LLM sources (ChatGPT, Perplexity, Gemini, Claude) as tracked in Google Analytics. Alongside them, conversion events standing in for a business outcome: purchases for some clients, leads for others.
On what: An unspecified number of brand websites within the author’s customer base. Industry, site size, and geography are not characterized anywhere in the source.
For how long: 13 months, January 1, 2025 to February 7, 2026. The H1/H2 growth comparison covers calendar H1 2025 (January–June) against H2 2025 (July–December).
Citation monitoring: A separate subset of the analysis tracked more than 5 000 prompts and their responses via LLM APIs, including Gemini, ChatGPT, and Perplexity, to observe which sources those models cited. The date range for this subset is described as “since September of last year,” putting it at roughly September 2025 to February 2026.
Control: None. This is an observational cohort, not a controlled experiment, and there’s no holdout group, no randomisation, and no way to separate LLM referral behaviour from concurrent changes in site content, paid media, or organic search performance.
Method disclosure checklist. The same seven points decide which numbers survive in our review of three GEO claims. The seven points this site requires of any vendor research:
| Point | Disclosed? |
|---|---|
| Model or interface (API vs. chat UI) | Partial — API stated for citation monitoring; not stated for GA referral data |
| Date range | Yes |
| Region | No |
| Prompt set | Partial — “more than 5 000 prompts” for citation monitoring; not described for GA data |
| Repetitions | No |
| Stochastic variation | No |
| Sample n at site level | No |
Most of those points are absent or only partially addressed. The data isn’t useless because of that, but it does constrain which conclusions actually hold.
The numbers
Data from Jason Tabeling’s analysis of Google Analytics LLM referral data, January 1, 2025 to February 7, 2026, across an uncharacterised customer base.
| Metric | Value | n | Confidence |
|---|---|---|---|
| Average LLM share of referral traffic | <2% | Not disclosed | Not disclosed |
| Site-level range, LLM referral share | 0.15%–1.5% | Not disclosed | Not disclosed |
| Average H1→H2 2025 growth, LLM referrals | 80% | Not disclosed | Not disclosed |
| Site-level growth range | 10%–300% | Not disclosed | Not disclosed |
| Jan→Dec 2025 aggregate referral growth | ~3× | Not disclosed | Not disclosed |
| Approximate LLM conversion rate | 18% | Not disclosed | Not disclosed |
| Prompts monitored for citation shifts | >5 000 | Not disclosed | Not disclosed |
The 18% conversion rate is the number most likely to be quoted out of context. Per the source, that traffic is roughly 25 times smaller than SEO or direct. So the number describes a small, high-intent audience — not a channel with meaningful conversion volume for most sites today. Without a session count behind that 18% figure, the rate could reflect a few hundred conversions across dozens of sites, or tens of thousands — the source does not say.
What could still be wrong
Sampling and customer mix. The dataset covers the author’s customer base. That population is self-selected: companies that hired this vendor were presumably already attentive to digital marketing. Sites with more sophisticated content and stronger brand recognition tend to attract more qualified visitors from any channel. The 18% conversion rate may reflect the customer mix as much as anything specific to LLM referrals.
Conversion definition heterogeneity. Combining purchase events and lead-generation events into a single conversion rate produces a number that is difficult to interpret. A lead form submission and a completed checkout are not equivalent outcomes. If the customer base skews toward B2B lead-gen — where form fills are common — the aggregate rate will look high against an e-commerce benchmark.
GA referral attribution for LLM sources. The source does not describe how Google Analytics decides which sessions belong to LLM referrals. Nor does it say what share of LLM-originated visits goes untracked. Some LLM-referred sessions arrive without a recognisable referral signal. If that happens often, the dataset undercounts LLM referrals. Worse, the sessions it does capture may differ from the ones it misses, which would push the conversion rate up on its own.
Seasonality. The 80% H1-to-H2 growth figure spans a period in which, per the source, consumer adoption was growing and prompt algorithms kept changing. Separating platform-driven growth from content or optimisation changes on client sites is not possible from this dataset.
Citation monitoring scope. That 5 000-prompt figure covers API calls, not consumer chat sessions. Prompt distributions in API testing may not reflect the queries real users submit. The claim that YouTube and Reddit citations rose “over the last 30 days” rests on a window ending in early February 2026. Thirty days is short enough that a single model update could produce the whole effect.
What to do on Monday
The one actionable decision this dataset supports is straightforward: set up a dedicated LLM referral segment in your analytics before the channel grows further.
In your analytics platform, create a segment for sessions from the major LLM referral sources. The source identifies four: ChatGPT, Perplexity, Gemini and Claude. Apply this segment to your existing conversion goals. Record the baseline session count and conversion rate now, in writing, with the date.
That baseline is what will let you distinguish genuine growth from seasonal noise six months from now. Without it, any change you observe will be an uncontrolled before/after; the kind of measurement this site does not treat as evidence.
If the segment shows meaningful volume for your site, the next step is to look at landing pages, not conversion rates. Which pages are receiving LLM-referred sessions? Do those pages match the intent that would explain a high conversion rate, or are the conversions concentrated on a handful of pages that happen to rank well in LLM citations? Landing-page data gives you more to work with than the aggregate 18% figure.
FAQ
Is 18% a realistic LLM conversion rate, or is something wrong with the measurement?
It’s plausible but unverifiable from what’s published. The source reports that LLM referrals convert higher than paid shopping, SEO and PPC. One explanation is that a visitor arriving on an LLM’s recommendation has already had the choice validated for them, which raises intent before they land. The problem is that the base session count isn’t disclosed, the conversion types are mixed, and the customer base isn’t characterised. The rate could be accurate for this specific population and still not transfer to your site.
How do I track LLM referral traffic in Google Analytics 4 without a third-party tool?
In GA4, build a channel grouping or exploration segment for the referral domains you want to watch. Export a monthly snapshot to a spreadsheet. That way the growth rates are yours, calculated the same way each month.
Should the shift in LLM citations toward YouTube and Reddit change my content strategy now?
Not on the basis of a 30-day window from a single vendor’s prompt monitoring dataset. Citation patterns inside LLM responses shift with model updates, and a one-month observation is too short to distinguish a durable trend from a version-specific change. The more useful signal is whether your own GA4 segment shows referral sessions arriving from YouTube or Reddit as intermediate steps, that’s observable in your own data. If you see it consistently over two or three months, investigate what content on those platforms is being cited.
How much did LLM referral traffic grow across the full 13 months?
The source puts aggregate referral growth at roughly three times between January and December 2025, on top of an average 80% rise from the first half of the year to the second. Both are averages across an undisclosed number of sites, and the site-level range behind the second figure runs from 10% to 300%. Growth of that shape says more about a small base than about a channel arriving at scale.
