Do any AI assistants actually pass referrers — or are they all appearing as direct traffic?

Sixty-three percent of sites in a 3,000-site analytics panel received at least one visit from an AI chatbot in 2025. But those are the visits you could see. The more useful question is how much AI traffic shows up as Direct — attributed to nothing — and never gets counted against any particular AI platform at all. The answer depends entirely on which platform sent the traffic. And the differences are large.

How fast is AI referral traffic actually growing?

Before getting into what’s invisible, it helps to anchor on what is measurable. AI-sourced traffic is growing fast enough that errors in attribution have real financial consequences.

Adobe Digital Insights, working from over 1 trillion visits to U.S. retail sites, measured 393% year-over-year growth in AI-sourced traffic in Q1 2026. During the 2025 holiday season, the same sites saw 693% year-over-year growth. Similarweb measured a 357% increase in AI referral visits globally as of June 2025. The Ahrefs 3,000-site panel found visits from AI chatbots grew roughly 9.7× over the prior 12 months, which works out to approximately 870% year-over-year. The studies use different methodologies and different site populations, but the direction is consistent.

The caveat worth repeating: every one of these studies counts only traffic that passed a referrer header. The unattributed portion is structurally invisible.

Which platforms actually pass the referrer?

This is where the per-platform breakdown matters, and the gaps are significant.

Perplexity’s web interface generally passes perplexity.ai as the referrer — it shows up in most analytics setups in the Referral channel. Perplexity’s mobile app strips it. ChatGPT free-plan users on web see chatgpt.com appear in referral reports. Paid-plan users are a different story: citation links in ChatGPT responses carry a rel="noreferrer" attribute, meaning those visits land as Direct. Google’s AI search mode launched with a confirmed noreferrer bug — a Google spokesperson acknowledged it as unintended on LinkedIn — which was fixed around May 28, 2025. Post-fix, those clicks blend into organic search rather than appearing as a separate AI channel. Google’s conversational AI generally passes the referrer, with the exception of its deep research feature, which explicitly uses noreferrer.

Copilot is the outlier. In direct testing across multiple sessions, every visit originated by Copilot showed up as Direct in Google Analytics. Zero referrer headers. Not partial — zero.

How does crawler volume compare to referral volume?

One of the more clarifying data points here is the ratio between how aggressively AI platforms crawl content and how much human traffic they actually send back. A July 2025 study covering approximately 20% of global web traffic measured this ratio for three major platforms.

For each human visitor that OAI-SearchBot’s associated product sent to an external site, the crawler made 1,091 automated requests to that site. PerplexityBot ran at 194 crawler requests per human referral. For ClaudeBot’s associated platform, the ratio reached approximately 38,000 crawler requests per human visitor referred. That’s not a typo — nearly four orders of magnitude more indexing than referring.

The practical implication is that server log analysis tells you about indexing interest, not referral intent. Seeing GPTBot or PerplexityBot at high volume in your logs doesn’t mean those platforms are about to send proportional human traffic. Indexing and referring are largely decoupled.

What changed with ChatGPT’s UTM parameters?

In June 2025, ChatGPT started appending utm_source=chatgpt.com to citation links in responses. For paid-plan users — the cohort previously hit hardest by the noreferrer problem — this UTM parameter became the primary attribution signal. UTM parameters survive noreferrer because they’re embedded in the URL itself rather than transmitted as HTTP headers.

If you’re not filtering for utm_source=chatgpt.com in your analytics setup, you’re probably misattributing some ChatGPT referrals to Direct. No equivalent UTM scheme exists for Copilot.

Which signals can you actually rely on?

Ranked by reliability for a production analytics setup:

1. UTM parameters — utm_source=chatgpt.com where ChatGPT appends it. Survives noreferrer. Doesn’t exist for Copilot, Perplexity, or Google’s AI products.

2. Referrer domain matching — Works for Perplexity web, ChatGPT free web, and most of Google’s conversational AI traffic. Fails on mobile for all platforms. Doesn’t catch Copilot at all.

3. GA4 AI Assistant channel — Google Analytics launched an AI Assistant default channel group in May 2026, covering ChatGPT, Google’s conversational AI, Copilot, and several others. Perplexity isn’t included — it still routes to Referral. Better than manual matching alone, but not complete.

4. Crawler user-agent strings — GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, and ClaudeBot are all documented. Useful for understanding indexing patterns, not for counting human visitors arriving from AI interfaces.

5. Published IP allowlists — Platforms publish JSON files with their crawler IP ranges. Cross-referencing server logs catches high-confidence bot traffic, but Perplexity rotates IPs broadly in practice, which means the published files are incomplete within days of release.

What should you actually do about this?

Set up three signals in parallel: referrer matching for chatgpt.com and perplexity.ai, UTM filtering for utm_source=chatgpt.com, and crawler UA logging to track indexing separately from human referrals. These three together give you a reasonable floor on AI-driven activity. It’s not a complete picture, but it’s the most defensible one available without server-side enrichment.

If you’re running e-commerce, the attribution gap has direct revenue implications. Adobe’s March 2026 data shows AI-referred traffic converted 42% better than non-AI traffic — a complete reversal from March 2025, when the same cohort converted 38% worse. Getting the measurement right means knowing which content is being referred, not just which content is being crawled. Those are different optimization targets.

Copilot’s gap isn’t going to close on its own soon. If your audience skews enterprise, you’re probably already under-counting that segment. The fallback is server-side log analysis: cross-reference raw request logs against platform-published IP range files for a lower-bound estimate. Treat it as a floor, not a ceiling.