How to track ChatGPT traffic to your site

"ChatGPT traffic" is really three different phenomena. Conflating them leads to bad decisions — blocking the wrong bot, optimizing the wrong landing page, or missing the moment a page starts getting cited.

1. AI crawlers hitting your site — GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, and dozens more. They request URLs server-to-server. No JavaScript runs. No GA4 pixel fires.
2. Humans arriving from AI assistants — someone clicked a link in ChatGPT, Perplexity, Gemini, Copilot, or Grok. These show up (sometimes) as referrers or utm_source=chatgpt.com on human browser sessions.
3. Being cited in AI answers — your brand or URL appears in a model response, with or without a click. You infer this from crawl coverage, llms.txt discovery, ranking in AI search surfaces, and correlation with referral spikes — not from a single pixel.

A complete picture needs instrumentation for all three. The sections below start with generic, stack-agnostic advice any team can implement this week. Later we show how Aiola Analytics surfaces the same signals — because we dogfood this on aiola.app and publish what we measure.

AI crawlers identify themselves with HTTP user-agent strings. OpenAI might arrive as GPTBot, OAI-SearchBot, or ChatGPT-User. Anthropic sends ClaudeBot and Claude-User. Perplexity uses PerplexityBot. xAI sends Grok-related bots. The full set changes — our AI crawlers directory lists 58+ with taxonomy as of July 2026.

How to see them (any stack):

• Raw access logs — nginx, Apache, Caddy, Cloudflare Logpush, Vercel log drains. Filter on user-agent substrings: GPTBot, ClaudeBot, PerplexityBot.
• Log CLI one-liner (nginx example):

grep -E "GPTBot|ClaudeBot|PerplexityBot|OAI-SearchBot|ChatGPT-User" /var/log/nginx/access.log \
  | awk '{print $1, $7, $9}' | tail -20

• Count by bot per day to spot trends — a sudden GPTBot spike on /pricing means OpenAI re-fetched pricing pages, not that a human visited.

Crawl types matter. We group bots into four buckets (same taxonomy as Aiola Analytics):

• Answers — user-triggered fetches when someone asks an AI to read a URL (e.g. ChatGPT-User).
• Indexing — building search or retrieval indexes (e.g. OAI-SearchBot, some Perplexity bots).
• Training — dataset collection for model development (e.g. GPTBot under training policy).
• Other — advertising, link-preview, and unclassified bot traffic that is neither an AI answer, indexing, nor training signal.

Your robots.txt can allow indexing while disallowing training for some vendors. Check each bot's documentation before a blanket Disallow — you might block discovery you wanted.

Common gotcha: WAF rules like Cloudflare "Block AI bots" return 403 to crawlers while your marketing site still looks fine in Chrome. Crawlers log the attempt; humans never see the block. Verify with curl -A GPTBot -I https://yoursite.com/ — expect 200 if you want to be cited.

When a person clicks through from ChatGPT, Perplexity, Claude, Gemini, Copilot, or Grok, analytics might show a referrer or UTM parameters — if the client passes them.

Referrer patterns to watch (July 2026):

• chatgpt.com, chat.openai.com — ChatGPT
• perplexity.ai — Perplexity
• claude.ai — Claude
• gemini.google.com, bard.google.com — Gemini
• copilot.microsoft.com, bing.com/chat — Copilot
• grok.x.ai, x.com — Grok (varies by client)

GA4 setup:

1. Open Reports → Acquisition → Traffic acquisition and add a secondary dimension for Session source.
2. Create an Exploration with filters: Session source contains chatgpt OR Referrer contains perplexity (repeat per assistant).
3. Tag outbound links you control with utm_source=chatgpt.com&utm_medium=referral in llms.txt, docs, and support macros — some AI clients preserve UTMs better than referrers.

Limitations: referrer stripping, in-app browsers, and copy-paste visits show as Direct. AI referral numbers are almost always undercounts. Treat them as directional — compare week over week, not as absolute truth.

On aiola.app we saw AI-channel referrals before we had a dedicated card — ChatGPT as our #4 acquisition channel while crawlers were blocked at the edge. Fixing WAF rules and measuring both layers separately changed the story. Measure crawls and referrals independently.

Citation is the hardest layer — you rarely get a webhook when ChatGPT mentions your product. Instead you optimize discoverability inputs and watch proxies.

llms.txt and machine-readable facts. Publish a concise /llms.txt (and optionally /llms-full.txt) with plain-HTML facts: what the product is, pricing, platforms, supported integrations, download URL. Agents quote quotable text — not buried JSON-LD alone.

Crawlability checklist:

• robots.txt allows the bots you intend (verify with curl -A, not assumptions).
• Key pages return 200 to bots — pricing, comparisons, docs, sitemap.xml.
• No accidental 403 from WAF, geo blocks, or bot fight mode on marketing zones.
• Stable URLs and sitemap with real lastmod dates.
• FAQ and comparison tables in HTML — LLMs extract tables cleanly.

Proxy metrics for citations:

• Rising ChatGPT-User / Perplexity-User fetches on specific URLs.
• Discovery-file requests — bots fetching /llms.txt, /robots.txt, /sitemap.xml.
• Branded search lift and direct traffic after AI search launches (noisy but useful).
• Manual prompt tests: ask ChatGPT/Perplexity/Claude "best tool for X" monthly and log whether you appear — Aiola runs this on our own automations.

Citation is a lagging indicator. Crawl health and llms.txt are leading indicators you control directly.

We built Aiola Analytics because indie operators running AI-coded products need the same visibility — without gluing together log grep, GA4 explorations, and a spreadsheet of bot names. This is educational context on how we dogfood it; you can replicate much of it by hand using the sections above.

1. Bot crawler card — Tracks 58+ AI crawlers with the answers / indexing / training / other taxonomy, per-bot trends, page-level drill-down, and bot-hit 404s (URLs crawlers requested that returned not-found). This is layer one without SSH access to nginx.

2. AI Assistants referral card — Surfaces human visitors from ChatGPT, Perplexity, Claude, Gemini, Copilot, and Grok alongside your normal funnel metrics. Layer two next to revenue and conversion data, not in a disconnected GA tab.

3. Discovery & indexing — Monitors fetches of robots.txt, llms.txt, and sitemap; connects Google Search Console and Bing indexing issues so broken pages bots hit show up as actionable fixes. Supports layer three citation readiness.

The AI crawlers directory uses the same bot list and taxonomy — useful even if you do not use Aiola, as a reference for user-agent strings and robots.txt snippets.

Aiola is a desktop operations layer for people who ship software with coding agents — analytics is one panel in that story, not a standalone enterprise suite. If you only need a monthly GPTBot count, server logs may suffice. If you want crawls, referrals, and indexing issues beside app logs, feedback, and agent threads, the integrated view is why we built it.

Day 1 — Verify crawl access. curl -A GPTBot, ClaudeBot, and PerplexityBot against your homepage, pricing, and llms.txt. Fix WAF rules if you get 403.

Day 2 — Baseline server logs. Count AI bot hits for seven days. Note which paths get the most fetches.

Day 3 — GA4 exploration. Build the AI referrer report. Add UTMs to links in llms.txt and docs.

Day 4 — Publish or refresh llms.txt. Plain facts: product name, what it does, pricing, download URL, platforms. Link to comparisons and guides.

Day 5 — Close 404s bots hit. GSC/Bing or Aiola indexing panel — broken URLs waste crawl budget.

Ongoing — Review weekly: crawler trend, referral trend, one manual AI search prompt test. Iterate content agents already fetch (pricing, comparisons, how-to guides).

Tracking ChatGPT traffic is not one dashboard widget — it is three layers with different tools. Start with logs and referrers; add integrated analytics when the manual glue becomes its own part-time job.