How Do I Track AI Search Traffic?

Track AI search traffic in three layers: referral traffic in GA4 by filtering session source for hostnames like chatgpt.com and perplexity.ai; crawler activity in server logs by filtering for user-agents like OAI-SearchBot and PerplexityBot; and citation appearance by manually running your key queries. No single tool covers all three.

Layer 1 — Referral traffic (who clicked through)

This is the only layer that shows real humans arriving on your site, and it's the easiest to set up.

When ChatGPT, Perplexity or Copilot names you as a source and someone clicks the citation, their browser sends a referrer. In GA4, that shows up under Reports → Acquisition → Traffic acquisition, with Session source / medium as the dimension. The hostnames to look for:

Platform	Source values you'll see
ChatGPT	`chatgpt.com`, `chat.openai.com`
Perplexity	`perplexity.ai`, `www.perplexity.ai`
Microsoft Copilot	`copilot.microsoft.com`, `bing.com` (mixed)
Google Gemini	`gemini.google.com`
Claude	`claude.ai`

Two practical notes. OpenAI appends ?utm_source=chatgpt.com to the links it renders, so ChatGPT clicks are also identifiable from the query string — useful if you're checking server logs or a tool that reads UTMs rather than referrers. And you should build a saved comparison or segment for these sources rather than eyeballing the list each month, because individually they're small enough to fall below the fold in a default report.

Set expectations honestly here: for most sites, AI referral volume is a small fraction of organic search. The reason to track it isn't the raw number — it's the trend line and the page-level detail. Knowing which pages get cited tells you what shape of content works, and that's worth more than the sessions themselves.

Layer 2 — Crawler activity (who's reading you)

Referrals only fire when someone clicks. Most AI answers get read without a click, so the second layer is server logs: which AI bots fetched which pages, and how often.

Filter your access logs by user-agent. The ones worth separating:

User-agent	Operator	What it's doing
`GPTBot`	OpenAI	Crawling for model training
`OAI-SearchBot`	OpenAI	Building the ChatGPT search index
`ChatGPT-User`	OpenAI	Fetching a page live, triggered by a user
`ClaudeBot`	Anthropic	Crawling for training
`Claude-User`	Anthropic	Live fetch during a conversation
`PerplexityBot`	Perplexity	Building its search index
`Perplexity-User`	Perplexity	Live fetch for an answer
`Google-Extended`	Google	Gemini and AI-features access control
`Bingbot`	Microsoft	Powers Copilot

The distinction between the index bots and the -User bots is the useful one. OAI-SearchBot and PerplexityBot visiting means you're eligible to be retrieved. ChatGPT-User and Perplexity-User visiting means a real person's question just caused an assistant to open your page — the closest thing to a live citation signal you can get from your own infrastructure.

On Vercel, Cloudflare or Netlify you can filter this in the platform's log view. On traditional hosting, grep the access log. Either way, check the same robots.txt file while you're there — if it disallows these agents, they can't reach you and none of the rest of this matters. What a robots.txt file does covers the syntax.

Layer 3 — Citation appearance (are you actually named?)

Neither of the layers above tells you whether you appeared in an answer that nobody clicked. For that you either check manually or pay a tool.

Manual, which is free and surprisingly effective: take your 20 highest-value queries, run each through ChatGPT, Perplexity, Gemini and Google once a month, and record three columns — were you cited, who else was, and which of your URLs was used. An hour a month gives you a real trend line and, more usefully, shows you exactly which competitor pages are winning so you can see what they do differently.

Paid tools — Otterly, Peec AI, ZipTie, LLMrefs and similar — automate that sampling across platforms and give you share-of-voice charts. They're worth it once you're tracking more queries than you'll patiently check by hand. Treat their numbers as directional: every one of them works by sampling prompts, and AI answers vary between runs for the same question.

What Search Console will and won't show you

This trips people up, so state it plainly: Google Search Console has no AI-specific report. Clicks and impressions from AI Overviews and AI Mode are folded into the standard Web performance data rather than broken out, and Google has said it has no plans to separate them.

What you can still do with it:

Watch impressions rising while clicks flatten on informational queries. That pattern is the fingerprint of an AI Overview answering the question above your result.
Compare click-through rate by query type. Queries where you hold position but lose CTR are the ones being absorbed.
Use it, as always, for the underlying ranking health — because AI Overviews are built on Google's core ranking systems, whatever improves your ranking improves your odds of being drawn on.

Putting it together

A workable monthly routine, about ninety minutes:

Pull GA4 traffic acquisition filtered to the AI hostnames; note sessions and, more importantly, landing pages.
Pull server logs for the eight user-agents above; note which pages the -User bots opened.
Run your 20 priority queries manually; log citations and competitors.
Check Search Console for impression-up/click-down queries.
Cross-reference: the pages that appear in all four are your proven AI-visible assets. Study what they have in common and apply it to the rest.

What you'll usually find is that the cited pages share a shape — a direct answer near the top, real question-and-answer blocks, specific checkable facts, and enough authority to be retrieved at all. The last of those is the slow part. Backlinkster handles it directly with one-for-one in-content link swaps between site owners in related niches, each verified live by code, five a month on the free plan.

Frequently asked questions

How do I track AI search traffic in GA4? Go to Reports → Acquisition → Traffic acquisition and set the dimension to Session source / medium. Look for chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com and claude.ai. Save it as a comparison so the small volumes don't get lost in the default view.

Does Google Search Console show AI Overview traffic? Not separately. Clicks and impressions from AI Overviews and AI Mode are included in the standard Web performance report but not broken out, and Google has said it doesn't plan to split them. The usable signal is impressions rising while clicks stay flat.

What user-agent does ChatGPT use? Three, with different jobs: GPTBot crawls for model training, OAI-SearchBot builds the ChatGPT search index, and ChatGPT-User fetches a page live when a user's question requires it. Seeing ChatGPT-User in your logs means a real conversation just pulled up your page.

Why is my AI referral traffic so low? Because most AI answers are read without a click — the assistant summarises and the user moves on. Low referral numbers don't mean you aren't being cited. Use server logs and manual query checks to see the citations that never produced a visit.

Can I see which pages AI systems cite? Partially. GA4 landing pages show which of your URLs produced clicks, and server logs show which pages the live-fetch bots opened. For citations with no click, you have to run the queries yourself or use a monitoring tool that samples prompts on a schedule.

Do I need a paid AI visibility tool? Not to start. Manually running 20 priority queries once a month across ChatGPT, Perplexity and Google gives you a real trend line for free. Paid tools earn their keep once you're tracking more queries than you'll check by hand.

The bottom line

There's no single dashboard for AI search, and anything claiming to be one is sampling. Build the three layers instead: GA4 for clicks, server logs for crawler and live-fetch activity, a manual query sweep for citations nobody clicked. Then act on the overlap — the pages that show up across all three are telling you what a citable page looks like on your site, and that's a more useful output than any headline number.