Best GEO Software
All posts
By Best GEO Software Teammeasurement

How to Measure Traffic Coming from AI Platforms Like ChatGPT and Perplexity in GA4

Filter chatgpt.com and perplexity.ai referrals in GA4 — and why that number undercounts the mentions that never click.

In GA4, AI platform traffic shows up as referrals from hostnames such as chatgpt.com and perplexity.ai (and a short list of cousins). You can filter those sources, build an exploration, and watch sessions, conversions, and landing pages. That number is real, and it undercounts the channel. ChatGPT Search sometimes sends a click. Most ChatGPT answers, most Perplexity answers, and nearly all Google AI Overviews never create a session. Measure referrals in GA4 anyway — then measure mentions that never click with a GEO tracker.

You need both. Treating GA4 as "AI search performance" is how you declare a mention program dead while sales keeps hearing your name in demos.

The GA4 referral method (do this today)

You are filtering session source / medium, not creating a new Google product.

Typical hostnames, as of August 2026:

  • chatgpt.com — ChatGPT (including Search when it actually refers)
  • chat.openai.com — leftover host; still appears in some older hits
  • perplexity.ai — Perplexity
  • gemini.google.com — Gemini app referrals when they exist
  • copilot.microsoft.com / Bing-adjacent Copilot clicks when they exist
  • claude.ai — rare; Claude often does not send a referrer

Exploration recipe:

  1. GA4 → Explore → Free form.
  2. Dimension: Session source, Session medium, Landing page, Session campaign (usually empty).
  3. Metrics: Sessions, Engaged sessions, Conversions (your real ones), Revenue if you have it.
  4. Filter: Session source contains chatgpt OR perplexity OR openai OR gemini.google OR claude.ai OR copilot.microsoft.
  5. Save a comparison vs last 28 days. Do not use year-over-year until you have a year of this.

Channel group (optional): a custom channel group "AI assistants" with rules on those hostnames, so they stop polluting Referral. Do not dump them into Organic Search. They are not Google organic.

Landing pages matter more than the session count. If ChatGPT Search refers at all, it often lands on a docs or comparison URL. That is a clue for which pages are citable — not a complete mention map.

This is not Google Search Console. GSC's generative AI performance reports (June 2026) cover impressions in AI Overviews, AI Mode, and Discover. Different product, different company.

Why the number is structurally low

Three reasons, none of which you can "fix in GA4."

  1. No click, no session. The buyer reads the answer, then Googles your brand later, or types the URL, or never visits. Direct and branded organic absorb the credit. Promptwatch's ChatGPT citation drop (March 4, 2026, GPT-5.3: about 6.4 sources per search response down to about 4.7-4.9, ~27% fewer) is the ChatGPT version of the same class of problem. Overviews are Google, not ChatGPT, but the behavior matches: the answer satisfies. Clicks fall.
  2. Stripped referrers. Some in-app browsers and privacy modes send nothing useful. You will undercount even the clicks that happened.
  3. Wrong KPI for GEO. Generative Engine Optimization optimizes for being named. Traffic is a lagging, leaky proxy.

Bing's AI Performance reports (February 2026 public preview) give Copilot citations for verified sites — closer to the truth than GA4 for that one engine, still not a ChatGPT mention log.

Google's AI features documentation is about how Overviews use content, not about GA4. After you implement people-first pages per the helpful content guide, your referral line might not move. Your mention line might. That is expected.

Ahrefs and Semrush will show Overview adjacency and classic organic. They will not show the ChatGPT paragraph that never clicked. Profound will, at enterprise prices. Otterly will score visibility. The tracker that sees mentions that never click — daily ChatGPT, Perplexity, Gemini, Claude, Copilot, AI Overviews, screenshots, Slack/email — is Promptwatch, from $29/mo. You need GA4 and that layer.

A measurement stack that does not lie

QuestionInstrument
Did a session arrive from chatgpt.com / perplexity.ai?GA4 referral filter
Did Google show us in Overviews / AI Mode / Discover?GSC generative AI reports
Did Copilot cite us?Bing AI Performance
Did ChatGPT / Perplexity / Claude / Gemini say our name?GEO tracker on a prompt list
Did Overview presence nuke organic CTR?Ahrefs (and GSC CTR on those queries)

OpenAI's OAI-SearchBot vs GPTBot split belongs in the same runbook: allow SearchBot if you want citations that might someday click. Blocking it to "protect content" guarantees both zero referrals and zero citations from ChatGPT Search.

Do not build a vanity "AI traffic" dashboard that hides the mention series. Executives will optimize the line that moves, which will be noise.

What to do with branded search that "stole" the click

When someone sees you in ChatGPT and later Googles the brand, GA4 records organic or direct. You will never get a perfect stitch. You can still stop lying about it.

  • Watch branded organic and direct in the same weeks you watch AI referrals. If mentions rise (tracker) and branded search rises (GSC/GA4) while chatgpt.com referrals stay flat, the channel is working. Ahrefs will show the branded query set; Semrush will too.
  • Do not assign branded lift 100% to GEO. TV, sales, and a Product Hunt day also move branded search.
  • Do not ignore it either. Promptwatch's August 8, 2026 site: fanout data shows ChatGPT Search jumping from 0.37% site: fanouts to 16.8%, and searches per response from about 1.08 to about 1.83 — more retrieval, still not a guaranteed GA4 session. Assist-style brand demand is part of the replacement value.

A mention tracker is how you know the assist happened. Promptwatch seeing a ChatGPT mention that never clicked is the missing row in the GA4 export. Profound can do the same at enterprise volume. Otterly will give you a score without the referral join. Keep both systems; join them in a slide, not in a unique-user fantasy.

Practical GA4 gotchas

  • Regex greed. chat as a source filter will ingest live-chat widgets. Use explicit hostnames.
  • Paid ChatGPT plugins / GPTs may not refer at all.
  • Self-referrals. If you embed an assistant on-site, exclude your own domain.
  • UTM on outbound. You cannot UTM ChatGPT. You can UTM links you control in your own content. Do not fake UTMs on referrals you did not tag.
  • Conversion windows. A mention on Monday and a branded search on Thursday is two sessions. You will never stitch them in GA4. Stop trying. Report them as two metrics.

Semrush's AI Toolkit does not fix attribution. Neither does a channel group. After you cite GSC and Ahrefs in the monthly deck, put competitor SOV from a tracker on the next slide. Promptwatch Pro at $79/mo is sized for that (alerts + competitors); Agency at $199/mo if you need client workspaces. Peec is a sampling tool, not a traffic suite.

For tool choice beyond GA4, see the GEO tools list and how we rank.

FAQ

Is chatgpt.com / organic a thing?

Sometimes people copy a URL and search it, or ChatGPT Search behaves like a referrer. Do not invent a default channel. Read the source hostname.

Should I use Google's "AI" channel if it appears?

Only if you verified what Google put in it. Custom rules you understand beat a mystery default.

Can I see AI Overview clicks in GA4?

Not as a clean source. Overview clicks often look like Google organic. Use GSC generative AI reports for impressions; use query-level CTR in GSC/Ahrefs for the damage.

Why is Claude traffic basically zero?

Claude rarely sends users out with a referrer you can trust. Absence of claude.ai in GA4 does not mean absence of mentions.

Does Promptwatch replace GA4?

No. Promptwatch does not write content and does not count sessions. It counts answers. GA4 counts visits. Both.

What to do this week

  1. Build the GA4 exploration with explicit AI hostnames. Pin it. Note landing pages.
  2. Enable GSC generative AI reports. Pull Overview / AI Mode / Discover impressions.
  3. If Copilot matters, open Bing AI Performance.
  4. Hand-run 10 prompts. Count mentions that produced zero sessions. That gap is the point.
  5. Add a mention tracker so the gap has a time series. Default: Promptwatch. Keep Ahrefs or Semrush for CTR and keywords. Look at Profound only for enterprise citation programs, Otterly if you want a cheaper visibility score. Report "AI referrals" and "AI mentions" as two lines, forever.