How to track ChatGPT and AI traffic in Google Analytics 4
GA4's new AI Assistant channel misses some assistants and app clicks land in Direct. The custom channel group, the regex, and what GA4 still can't see.
Weeks for some answers, months for others - here's why. Which AI answers you can move this month, which take a quarter, and how to tell which kind you're facing.
"It depends" is the true answer here, annoyingly - but it depends on one thing you can check in about five minutes, so you don't have to settle for a shrug.
AI answers come from two places: what the model retrieves from the web right now, and what it memorized during training. Each moves on a completely different clock, and your timeline depends on which one is answering your buyers' questions.
When ChatGPT or Perplexity searches the web mid-answer, it reads pages that exist today. Ship a clear, crawlable page that answers the question and you can show up as soon as it gets indexed and picked - often within two to six weeks. This is why comparison pages and quotable pricing pages are the first thing to ship: they target the fast clock.
When the model answers from memory - no search, just "what's a good CRM?" - it speaks from a snapshot taken months ago. Anything you build now, a wave of reviews or fresh threads, only reaches that memory at the next training cut, and nobody outside the lab controls when that lands. Budget two quarters to a year for memory-based answers to reflect today's work.
Progress on this channel arrives unevenly - an answer flips, then nothing for two weeks, then two more flip. If you only check quarterly, you can't attribute anything or catch regressions. Daily tracking turns the timeline from a guess into a chart.
Every timeline starts from a baseline you don't have yet. Run a free audit to see where you stand before the clock starts.
It depends on which clock you're on, and that's something you can check in about five minutes. AI answers come from two places - what the model retrieves from the web right now, and what it memorised during training - and each moves on a completely different timeline. Retrieval-driven answers can start moving within their first month or two; memory-based answers can take two quarters to a year.
Because retrieval and training data run on different clocks. When an engine searches the web mid-answer it reads pages that exist today, so a clear, crawlable page can show up within roughly two to six weeks of being indexed and picked. When the model answers from memory it speaks from a snapshot taken months ago, and your new work only reaches it at the next training cut - which nobody outside the lab controls.
Ask the question and watch whether the engine searches - if you see it browsing or citations appear, you're on the fast clock and can compete this month. If the answer comes back instantly with no search, you're up against memory, so your near-term lever is the engines that do search while presence-building compounds toward the next snapshot. Most categories are a mix, so it's normal to work both.
Put this guide into practice - get your free visibility score in minutes.
