Blog

AI referral traffic: is ChatGPT recommending you?

A growing share of people no longer search, they ask. If an assistant recommends your competitor instead of you, you never see that visit. Here is how to measure the referrals you do get, and how to become recommendable in the first place.

· Published

The shift from search to answers

For two decades, being findable meant ranking in a list of blue links. Assistant-based search changes the shape of that: the user asks a question, the assistant answers with a handful of named options, and the user clicks one. Recent surveys already report that a large share of consumers use generative AI for local and product discovery, and the number keeps climbing.

The practical consequence: a new acquisition channel exists, and most site owners have no data on it. Generic analytics show a visit from "chatgpt.com" as one more anonymous referral row, if they show it at all.

What an AI referral actually is

When someone clicks a citation inside an AI answer, their browser sends the AI platform's URL as the referrer. That referrer is ordinary data your analytics already collects; it just needs to be recognized and labeled. That is what an AI referrals report does: it groups visits by sending platform (ChatGPT, Perplexity, Gemini, Copilot, Claude, and others), shows bounce rate and duration per platform, and on commerce sites attributes revenue to those visits.

One boundary matters: AI crawlers (GPTBot, ClaudeBot, PerplexityBot and friends) never show up in any JavaScript-based analytics. They do not execute scripts, so they are invisible to every analytics tool, not just privacy-first ones. The referrals you can measure are the humans who clicked out of an answer.

Crawler access decides whether you exist

Before an assistant can recommend you, its crawler must be able to read your site. Three things decide that, and all three are checkable without asking any AI provider:

  • robots.txt: a single Disallow line aimed at a crawler like GPTBot or Google-Extended removes you from that assistant's answers entirely.
  • Server-rendered content: AI crawlers do not run JavaScript. A site whose text only appears after JS execution looks empty to them.
  • Machine-readable structure: clean titles, meta descriptions, and JSON-LD structured data make it easy for a model to describe and cite your page correctly.

A GEO audit turns these into a repeatable checklist: it reads robots.txt, llms.txt, your sitemap, and your most-visited pages, and scores each one. The point is not a vanity number; it is a short list of concrete fixes, such as "GPTBot is blocked in robots.txt" or "your product pages have no structured data".

Privacy note. An audit like this reads your own public pages and applies rules. It never sends your data through an AI provider. Measuring actual AI answers would require doing exactly that, which is why MetriXs deliberately stops at the deterministic layer.

A weekly routine that takes ten minutes

You do not need a platform for this. Once a month is enough to stay on top of it:

  • Check your AI referrals: which assistants send visitors, and do those visitors convert?
  • Re-run a crawlability audit after every site change: new robots.txt rules, a JS framework migration, a new theme.
  • Fix what scores worst. Usually that is one blocked crawler, a missing llms.txt, or thin meta descriptions on your most important pages.

Treat AI visibility the way you treated Google indexing in 2005: unglamorous, mechanical, and disproportionately valuable for those who set it up early.

Frequently asked questions

Can I see AI traffic in regular analytics?
Most tools either do not label it or lump it into generic referrals. In MetriXs, AI-referred visits are grouped per assistant and classified as their own traffic medium, with bounce rate, duration, and revenue attribution on commerce-enabled sites.
Does blocking GPTBot in robots.txt hurt my regular SEO?
No. GPTBot controls ChatGPT training and answer crawlers; Google search uses its own crawlers. Blocking one does not affect the other, which is exactly why it is worth checking deliberately instead of leaving a stale catch-all rule in place.
Do I need an LLM API to measure AI visibility?
No for the deterministic part: crawler access, machine-readable content, and structure are all checkable with plain HTTP requests and rules. Only measuring whether a specific AI answer names your brand requires querying an AI provider, which is a different, privacy-sensitive trade-off.
What is an llms.txt file?
A proposed convention where a site serves a llms.txt file at its root, offering a curated, markdown-formatted overview of its content for AI systems. It is optional, but its presence is a signal that a site is deliberate about AI readability.

Measure your AI referrals and audit your site

MetriXs shows AI-referred visits in your dashboard and runs a deterministic GEO audit of crawler access, content, and structure. No LLM APIs, nothing leaves the EU.

Start free

Cookieless tracking: how it works and why it is GDPR-safeGDPR-compliant analytics: what to look forGoogle Analytics alternative