Is AI Referral Traffic Falling? #Newsletter 06

If you’re seeing fewer sessions from ChatGPT/Claude/Gemini, you’re not imagining it.

But the ‘why’ is not what most marketers think it is.

There’s a lot to unpack here, so let’s break it down.

The Confusing Part

AI usage is obviously up.

Similarweb says GenAI platforms are now getting 7B monthly web visits, and referral traffic from GenAI platforms climbed to 2B visits (reported as a big YoY jump).

Adobe reported a 10x+ increase in AI-driven referral traffic in the U.S. in 2025.

So why are so many sites still seeing AI referrals stall or drop?

Because growth in AI usage does not guarantee growth in outbound clicks per answer.

That’s the whole story.

What Changed: Fewer Links per Answer

LLM Scout’s 2026 analysis tracked 15,252 queries and 90,232 citations across ChatGPT, Claude, Gemini, and Perplexity (Sept 2025 > Jan 2026).

Their core finding is that answers are getting more compressed: fewer citations per response, across models that used to be link-heavy.

The cleanest example is Claude:

• October 2025: 17.19 citations per prompt
• January 2026: 6.09 citations per prompt

Gemini and Perplexity show a similar contraction.

ChatGPT barely moves because it was already link-light.

This is the part most people miss: If the model gives the user fewer places to click, total referral opportunity drops even if prompts increase.

‘But My Content Is Getting Mentioned More.’

That can be true at the same time.

In fact, BrightEdge has data that supports the consolidation angle: citations and mentions are mostly stable week-to-week, but when they change, it’s overwhelmingly domains dropping out (not being replaced by new winners).

That matters because it breaks the old assumption: You can maintain visibility (mentions/citations) and still lose clicks because there are fewer outbound slots overall.

So traffic becomes a messy KPI. It stops telling you whether you’re succeeding inside the answer.

Why Clicks Drop Even When Citations Exist

Google’s AI Overviews show the same pattern.

Pew found that when users encountered an AI summary, they clicked a traditional result in 8% of visits vs 15% when there was no AI summary. And clicks on links inside the AI summary were extremely rare (Pew describes it as a very small share; other coverage calls out 1%).

That’s the click behavior plain and simple: When the answer feels complete, users stop browsing.

So even if your page is cited, you are competing against a product designed to reduce the need for a click.

Why Other Reports Still Show Referrals Rising

LLM Scout is measuring citations per prompt (link density).

Similarweb/Adobe are often measuring total referral volume from AI platforms.

Both can be true:

• Total AI referrals can go up (more users, more AI baked into workflows)
• While referrals per answer go down (fewer links per response)

Also, a big percentage growth can still be a small slice of your overall traffic.

Conductor put AI referral traffic around 1.08% of total website traffic as of Nov 2025.

So yes, AI referrals can surge and still feel irrelevant or volatile elsewhere.

The Channel Is Also Changing

Even within AI referrals, the mix is changing.

Search Engine Land covered a dataset of 774,331 LLM sessions to SaaS sites (Nov 2024 > Dec 2025). ChatGPT drove most of it, but Copilot grew fast as it sits inside workflows.

Two points here for marketers:

1. The referrers that grow are often the ones embedded in productivity tools, not standalone chat experiences.
2. Where AI traffic lands is not what you expect: that same dataset found 41.4% of SaaS AI traffic lands on internal search pages first.

Most sites don’t build those pages to convert. They’re usually:

• a bare list of results
• weak context (no ‘best match’ explanation, no suggested next step)
• poor filtering/sorting
• sometimes blocked from indexing because they’re messy

So even when you get the AI click, you often send the visitor to a dead-end page that doesn’t help them take the next action.

If AI referrals now require your landing page to do more work than it used to.

The Cloudflare Factor (Real, but Usually Misdiagnosed)

Cloudflare has been pushing hard on crawler controls (robots.txt management, blocking AI bots on monetized areas, etc.).

They also publicly moved toward default blocking of known AI crawlers for protection/compensation reasons.

But here’s the mistake teams make:

They blame traffic loss on crawling restrictions.

Cloudflare’s own Radar write-up points out the bigger issue: AI bots generate a lot of activity, but much of it doesn’t turn into end users being referred back to source sites.

Blocking can increase problems, but it doesn’t explain the structural ‘fewer outbound links per answer’ change.

One is about access. The other is about product design.

What to Take From This

AI referral traffic is not a reliable growth channel the way Google used to be. The click is optional.

So if your entire AI strategy is get cited to get traffic, you will keep getting disappointed.

Here’s the key takeaway stack:

• Stop using sessions as the main KPI. Track presence inside answers: mentions, citations, and position.
• Assume link slots will keep shrinking. If your content is generic, it will be summarized and you will not get the click.
• Build reason-to-click assets. Original data, calculators, templates, interactive tools, product comparisons with constraints. Things the model can’t fully compress without losing utility.
• Treat AI clicks like messy mid-intent traffic. If 40%+ of AI traffic can land on internal search pages in SaaS, your landing matters: strong internal search UX, curated results, and next steps.
• Expect mixed signals in reporting. Total referrals from AI platforms can rise while clicks-per-answer fall. Don’t let a single chart drive strategy.
• Plan for a world where visibility ≠ visits. This is the reality: influence happens inside the LLM, whether you get the session or not.