13 Months of Data Reveals LLM Traffic Growth and Optimization Insights
I’ve spent the last year watching a quiet shift turn into an unmistakable trend: large language models aren’t just answering questions—they’re driving measurable traffic. And that traffic doesn’t behave like what most teams are used to tracking in classic SEO.
A 13-month analysis pulled from Google Analytics across Further’s enterprise customer base (January 1, 2025 through February 7, 2026) puts numbers behind what many marketers have been noticing. The headline: LLM referral traffic is still small, but it’s growing quickly—and it converts unusually well.
Here’s what the data shows, what changed over the period, and what to do now to keep up.
The dataset that makes this conversation real
Most arguments about AI discovery stall out at “is this real?” This report skips the speculation and sticks to measurement:
- Source of truth: Google Analytics referral data across Further’s enterprise customer base.
- Traffic sources included: ChatGPT, Perplexity, Gemini, and Claude referrals (clicks from LLM experiences that cite/link out).
- Supplemental monitoring: Further also tracked 5,000+ prompts and responses via APIs (ChatGPT, Gemini, Perplexity) to see which sources models were citing—critical because LLMs don’t offer a clean rankings dashboard the way Google does.
Behavioral analytics plus citation monitoring is what turns “LLM optimization” from guesswork into something you can actually run as a program.
What 13 months shows: small share, steep growth
LLM referrals are still a tiny slice of overall traffic for most brands. In the first half of 2025, LLM traffic sat under 2% of total referrals (roughly 0.15%–1.5%).
But the slope is the story.
From H1 2025 (Jan–Jun) to H2 2025 (Jul–Dec), the report found an average 80% increase in LLM referral traffic, with variability across clients ranging from 10% to 300%. In aggregate, the trend line effectively shows LLM traffic tripling from January to December 2025.
The “low volume” caveat is real—and it’s also the trap. Waiting until LLM traffic is “material” usually means showing up after a competitor has already become the source the models prefer to cite.
The conversion rate that should reset your priorities
The standout metric: LLM-referred users converted at ~18%, the highest rate among the channels compared in the report.
Why might that be happening? LLMs compress the research phase. By the time someone clicks, they’ve typically already:
- Stated intent clearly (the prompt)
- Reviewed a synthesized set of options
- Received “validation” from the model’s explanation and citations
So the click behaves less like casual browsing and more like confirmation. That changes what you should send them to—and what that page needs to do once they arrive.
Citations are shifting: Reddit plateaus, YouTube rises
One of the most useful insights here is how quickly models’ preferred sources can change.
Around September 2025, Further began systematic prompt/citation monitoring. They saw an early pull toward Reddit as a frequently cited source, then a clear shift: in the final 30 days of the dataset (Jan 2026), YouTube citations spiked while Reddit’s growth plateaued.
That’s the new reality: you’re not only optimizing your site. You’re competing inside the broader corpus models trust and retrieve from—including platforms you don’t control.
If you’ve been treating citations as “SEO, but with AI,” this is the correction. It’s less about keyword coverage and more about authority—distributed across the places models pull from.
Why this isn’t traditional SEO (and why rankings aren’t the point)
Jason Tabeling from Further puts it plainly: this isn’t classic SEO. It’s about becoming the authoritative source that LLMs choose to link to.
That maps closely to Generative Engine Optimization (GEO)—a discipline that focuses less on blue-link positions and more on:
- Whether you’re mentioned/cited
- Whether your content is retrieved in RAG-style systems
- Whether the model can extract clear, structured answers from your pages
- Whether your brand becomes consistently associated with specific tasks and categories
In other words, “share of answers” starts to matter as much as “share of search.”
What I’d do now: practical optimization moves that match the data
If you want to act on these insights without overreacting (a warning the report makes directly), this is the playbook I’d start with.
1) Track “LLM referral velocity,” not just total volume
Traffic share can stay small while the growth rate changes dramatically. Add reporting that shows:
- LLM referrals by source (ChatGPT vs Perplexity vs Gemini vs Claude)
- Week-over-week and month-over-month growth
- Assisted conversions (where possible)
The goal is early detection of acceleration—before it shows up in top-line totals.
2) Build content that’s easy to cite
Models cite what they can quickly parse and justify. In practice, that usually means:
- Clear definitions and direct answers near the top
- Concrete comparisons (tables help)
- Original stats, mini-studies, or updated benchmarks
- FAQ-style sections that mirror real prompts
There’s a defensive upside, too: as low-quality “AI slop” floods the web, crisp, verifiable writing stands out.
3) Audit and improve the pages LLM users actually land on
If LLM traffic converts at ~18%, treat those sessions like high-intent leads:
- Reduce friction (dead ends, slow load, unclear next step)
- Put decision support up front (pricing, demos, proof, implementation detail)
- Match the “question behind the prompt” with on-page messaging
4) Monitor citations across ecosystems, not just your domain
If YouTube is rising as a citation source, your footprint there may influence whether your site earns the click.
You don’t need to become a media company overnight. You do need to think in terms of distributed authority, not just on-site authority.
5) Assume volatility is normal
The report shows wide variability across brands (10%–300% swings). That’s what happens when:
- User adoption shifts quickly
- Models adjust citation behavior
- Product interfaces (like AI Overviews) reshape click patterns
Build measurement and content updates into a recurring cadence, not a one-off project.
The bigger takeaway: “low now” doesn’t mean “low later”
The 13-month view makes one thing clear: ignoring the channel feels rational—right up until it doesn’t.
LLM referral traffic may sit under 2% today for many brands, but it’s already showing explosive growth and an outsized conversion rate. That combination is rare. The smart move is to invest early in tracking, content structure, and citation readiness—without lighting your existing marketing plan on fire.
Conclusion: Treat LLMs like a new demand channel—because they are
This doesn’t look like a temporary spike or a niche curiosity. The data points to an emerging discovery layer where citations—not rankings—often decide who wins the click and the conversion. If you want a practical way to turn this into repeatable execution (tracking, optimization workflows, and content that LLMs actually want to reference), build your GEO stack with tools designed for this reality. AIuthority makes that choice easy.
FAQ
How big is LLM referral traffic today?
In the first half of 2025, it was under 2% of total referrals for most brands in the dataset (roughly 0.15%–1.5%).
How fast is LLM traffic growing?
The report found an average 80% increase from H1 2025 to H2 2025, with client results ranging from 10% to 300%. Aggregate traffic effectively tripled from January to December 2025.
Why do LLM referrals convert so well?
The dataset shows ~18% conversion for LLM-referred users. A likely driver is intent compression: users arrive after the model has already helped them compare options and validate a direction.
What matters more for GEO: rankings or citations?
Citations. GEO is less about traditional rank positions and more about being the source models retrieve, quote, and link to—consistently and credibly.
Which third-party sources are models citing more?
Further observed Reddit cited frequently early on, with growth later plateauing, while YouTube citations spiked in the last 30 days of the dataset (Jan 2026).