How I Changed My GA4 Settings to Become a GEO Power User

By Jourdan Rombough, Optimization Theory · Updated October 22, 2026 · Field-tested on a three-site network

GA4 can tell you that ChatGPT sent a visit. It cannot tell you whether that visitor read anything, came back, converted, or was a bot. That gap is the difference between a GEO report and a GEO guess.

I just rebuilt the analytics on my own network (optimizationtheory.com, vuss.io and infographi.ac) to answer the questions SEOs are now getting from clients: which AI assistants cite us, what those visitors do next, which content actually gets read, and how much of the "traffic" is real. This is every setting I changed, the code I used, and the reports I built on top. If you already know where the GA4 admin menu is, this is for you.

Sources: AI share range across Contentsquare, Ahrefs, SE Ranking and Conductor studies compiled by Mission Growth. ChatGPT share and growth from BrightEdge. Click rates from Pew Research Center.

What GA4's native AI Assistant channel misses

Google added an AI Assistant channel to the default channel group in May 2026. Recognized assistants get the medium ai-assistant and the campaign (ai-assistant). It is a good start, and it is not enough for GEO work:

• ✕No history. It is not retroactive. Everything before the rollout stays in Referral.
• ✕Unpublished source list. Google has not published which assistants it recognizes, and reports disagree on whether Claude and Perplexity are included. Check your own Traffic acquisition report for claude.ai and perplexity.ai with a medium of "referral."
• ✕App clicks. Clicks from mobile and desktop assistant apps often arrive with no referrer and land in Direct.
• ✕Google's own AI. AI Overviews and AI Mode clicks still report as google / organic.
• ✕No behavior layer. It labels the session. It does not tell you which assistant a returning customer first came from.

Sources: Semrush and Zeevron on the rollout and its gaps. The rest of this guide closes each one.

Keep 14 months of event data, not 2

The default event data retention is 2 months. Any exploration, funnel or path report older than that comes back empty. GEO is a trend story, and you cannot show a client that AI referrals doubled year over year with 60 days of data.

Tune enhanced measurement for content sites

Open your web stream (Admin › Data streams › your stream). Enhanced measurement is the block of automatic events GA4 collects without any code.

• Page views: on, including browser history changes
• Scrolls: on. This fires at 90% depth. Tag Manager adds 25, 50 and 75.
• Outbound clicks: on. You want to know when readers click your citations and sources.
• Site search: on. Include the query parameters q,s,search,query,keyword. WordPress uses s.
• Form interactions: off. It fires on every form focus, including search boxes and newsletter bars. Track successful submissions instead.
• Video engagement and file downloads: on

Stop payment pages from stealing your organic credit

When a visitor who arrived from Google or ChatGPT goes to PayPal or Stripe to pay and comes back, GA4 can start a new session credited to the payment site. The channel that earned the sale loses it. "Sign in with Google" does the same.

Add these as "Referral domain contains":

• paypal.com
• paypal.me
• stripe.com
• accounts.google.com
• mail.google.com

Take yourself out of the data

When you test your own pages in ChatGPT and click through, you become AI referral traffic. On a small site that alone can make GEO look like it is working.

Name the rule, keep traffic_type as internal, and add your IP.

Then check Admin › Data filters. GA4 creates the "Internal Traffic" filter in Testing mode, which labels your visits without excluding them. Confirm the label in DebugView, then switch it to Active.

Running more than one site? Build a Network property

If you own sites that link to each other, each property sees the others as plain referrals. You never see the full journey, like a reader who lands on one site from Perplexity and requests a quote on another.

The fix is a separate Network property that every site also sends to, with cross-domain measurement on. Each site keeps its own clean property, and the Network property is the roll-up.

Add one web stream, then open Configure tag settings › Configure your domains and list every domain:

Give the Network property the same retention, unwanted referrals, internal traffic rules and custom definitions as the others, and add a site_name parameter to every event so one report can switch between sites.

Move tracking into Tag Manager, with Consent Mode v2

Site Kit and similar plugins are fine for page views. They cannot capture the AI source of a session, real reading behavior, tool use or where on the page a click happened. Tag Manager can.

I built each container as a JSON file and imported it (Admin › Import Container), which loads every tag, trigger and variable in one pass:

Each container fires two Google tags, one for the site property and one for the Network property:

On WordPress, GTM4WP places the container and pushes post type, categories, author, publish date and logged-in state into the data layer.

Consent Mode v2 on a free consent plugin

With Consent Mode, a denied visitor still sends a cookieless ping and Google models the gap. Without it, an opt-out erases the visit, and AI referral numbers that are already small get smaller.

I use Complianz (free) with United States opt-out rules and Global Privacy Control respected. The free version does not pass Consent Mode signals to Tag Manager, so each container has a Custom HTML tag on the Consent Initialization trigger that reads the banner choice and the browser's GPC flag and sets the defaults before any Google tag fires.

Build an AI Assistants channel that beats the default

A custom channel group fixes the history and coverage problems. Unlike the native channel, a custom group re-buckets historical data whenever you select it in a report or exploration.

.*(chatgpt|openai|perplexity|gemini\.google|bard\.google|copilot|edgeservices\.bing|claude\.ai|you\.com|phind|poe\.com|meta\.ai|deepseek|grok|x\.ai|mistral|kagi|duck\.ai|qwen|character\.ai).*

Capture the AI source on every hit, not just the session

A channel tells you how a session started. For GEO you also want to know, on every event, which assistant the reader came from, and for every user, which assistant first introduced them. That takes three pieces in Tag Manager:

• ai_sourceEvent parameter on every hit in the session: chatgpt, perplexity, gemini, copilot, claude and so on. Read from the referrer or utm_source, then held for the rest of the session.
• first_ai_sourceUser property saved once and kept. Shows which assistant introduced readers who return weeks later through search or direct.
• ai_landingEvent fired on the first page of an AI-referred session, with the landing page. Your list of pages AI assistants actually send people to.

Here is the Custom JavaScript variable behind ai_source, trimmed for readability. Add more assistants to the map as they appear in your referral report.

function () {
  try {
    var saved = sessionStorage.getItem('ai_source');
    if (saved) return saved;
  } catch (e) {}
  var map = [
    [/(^|\.)(chatgpt\.com|chat\.openai\.com)$/, 'chatgpt'],
    [/(^|\.)perplexity\.ai$/, 'perplexity'],
    [/(^|\.)gemini\.google\.com$/, 'gemini'],
    [/(^|\.)(copilot\.microsoft\.com|edgeservices\.bing\.com)$/, 'copilot'],
    [/(^|\.)claude\.ai$/, 'claude'],
    [/(^|\.)(grok\.com|x\.ai)$/, 'grok'],
    [/(^|\.)(deepseek\.com|chat\.deepseek\.com)$/, 'deepseek']
  ];
  function match(host) {
    host = (host || '').toLowerCase().replace(/^www\./, '');
    for (var i = 0; i < map.length; i++) if (map[i][0].test(host)) return map[i][1];
    return null;
  }
  var src = null;
  if (document.referrer) {
    var a = document.createElement('a');
    a.href = document.referrer;
    src = match(a.hostname);
  }
  if (!src) {
    var m = location.search.match(/[?&]utm_source=([^&#]*)/i);
    if (m) src = match(decodeURIComponent(m[1]))
      || (/chatgpt/i.test(m[1]) ? 'chatgpt' : null);
  }
  if (src) {
    try {
      sessionStorage.setItem('ai_source', src);
      if (!localStorage.getItem('first_ai_source')) localStorage.setItem('first_ai_source', src);
    } catch (e) {}
  }
  return src || undefined;
}

Send it as an event parameter in your Google tag's shared event settings, and send localStorage.first_ai_source as a user property. Register both as custom dimensions (step 12).

How to track traffic from AI Overviews (and why you mostly cannot)

This is the honest part. Clicks from Google AI Overviews and AI Mode arrive as google / organic, with nothing in the referrer to separate them. Pew's study of real browsing found people clicked a standard result on 8% of searches with an AI summary versus 15% without, and clicked a link inside the summary on just 1%. The traffic is small, and it is hidden inside organic.

What you can do:

• ✓Search Console's generative AI report. Announced June 2026 and rolling out from the UK, it shows AI Overviews and AI Mode impressions by page, country, device and date. It does not yet show clicks, CTR or queries. Pair its pages with GA4 landing pages to spot pages earning AI visibility.
• ✓Text-fragment landings (experimental). When Google deep-links to a passage, the URL carries #:~:text=. Browsers hide it from page scripts, but Chrome's navigation timing entry can expose it. I log it as a text_fragment_landing event with the quoted text. It undercounts, but every hit shows the exact sentence Google pulled.
• ✓Engagement deltas. Compare engaged read rate for organic landings on queries where an AI Overview appears against those where one does not. A drop in clicks with stable or rising read rate is the AI Overview signature.

Sources: Pew Research Center (900 US adults, 68,879 searches, March 2025) and Passionfruit on the Search Console report.

Track real reads, not engaged sessions

GA4's "engaged session" is 10 seconds, two page views or one key event. For content, that tells you almost nothing about whether an AI-referred visitor found the answer they were promised. Here is what I track instead, all through one tracking script in Tag Manager:

• engaged_readFires once at 75% scroll plus 60 seconds of active, visible time. Time only counts while the tab is visible and the reader moved, scrolled or tapped in the last 30 seconds.
• scroll_depth25, 50 and 75 percent. GA4's own scroll event covers 90.
• network_clickClicks between your own sites, with destination and page zone: header, nav, content, sidebar or footer.
• tool_openCalculators, simulators, quote builders and quizzes, counted once per session so multi-page tools do not inflate numbers.
• generate_leadA successful form submission from the form plugin's own success event, not a button click.
• contact_clickContact, pricing and quote clicks each get their own event with cta_location, so you know which part of the page drives action.

Mark only the ones that represent business value as key events. On a service site that is leads, quote starts and contact clicks. On a publisher it is engaged reads, newsletter sign-ups and clicks to money pages.

Register custom dimensions for content metadata

Parameters stay invisible in reports until you register them. Go to Admin › Custom definitions › Create custom dimension, choose the scope, and type the parameter name. I registered 29 dimensions and 3 metrics in every property.

For numbers you sum or average, create custom metrics and set the right unit. Read time should be Seconds:

Flag ghost and bot traffic instead of deleting it

AI referral traffic is a fraction of a percent of visits. A handful of bot sessions with a spoofed chatgpt.com referrer can double it. Filtering looks like the fix, but GA4 data filters are permanent: an excluded hit is gone, and so is your evidence. I label every hit instead:

With the BigQuery export on, this query shows how much of your data arrived without the Tag Manager label in the last 28 days:

SELECT
  IFNULL(
    (SELECT value.string_value FROM UNNEST(event_params) WHERE key = 'hit_source'),
    'ghost (no gtm)'
  ) AS hit_source,
  COUNT(*) AS events,
  COUNT(DISTINCT user_pseudo_id) AS users
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN FORMAT_DATE('%Y%m%d', DATE_SUB(CURRENT_DATE(), INTERVAL 28 DAY))
  AND FORMAT_DATE('%Y%m%d', CURRENT_DATE())
GROUP BY hit_source
ORDER BY events DESC;

The GEO reports I actually use

None of the tracking matters until it answers a client question. These are the reports I build on top, in GA4 Explorations and in a Looker Studio dashboard on the Network property.

In Looker Studio, two calculated fields do most of the work. The first groups sources into platforms:

CASE
  WHEN REGEXP_MATCH(Session source, ".*(chatgpt|openai).*") THEN "ChatGPT"
  WHEN REGEXP_MATCH(Session source, ".*perplexity.*") THEN "Perplexity"
  WHEN REGEXP_MATCH(Session source, ".*(gemini|bard).*") THEN "Gemini"
  WHEN REGEXP_MATCH(Session source, ".*(copilot|edgeservices\\.bing).*") THEN "Copilot"
  WHEN REGEXP_MATCH(Session source, ".*claude\\.ai.*") THEN "Claude"
  WHEN REGEXP_MATCH(Session source, ".*(grok|deepseek|meta\\.ai|mistral|you\\.com|phind|kagi).*") THEN "Other AI"
  ELSE "Not AI"
END

The second turns the labels from step 13 into a single Data Quality control, defaulted to Clean:

CASE
  WHEN hit_source IS NULL OR hit_source = "(not set)" THEN "Ghost (no GTM)"
  WHEN traffic_quality = "bot_suspect" THEN "Bot suspect"
  WHEN traffic_quality = "spam_referrer" THEN "Spam referrer"
  WHEN traffic_type = "internal" THEN "Internal"
  ELSE "Clean"
END

• Search Console linkAdmin › Product links. Brings queries and organic landing pages into GA4 next to engagement data.
• BigQuery linkFree daily export of raw events. Turn on billing in Google Cloud or sandbox tables expire after 60 days.
• Looker StudioOne report on the Network property with a site switcher: overview, GEO, Search Console, content, funnels, tools, cross-site flow and data quality.

GA4 audit checklist for SEOs

Run this before you trust any number, yours or a client's:

1. 1Event data retention is 14 months on every property
2. 2Exactly one GA4 tag per page: no plugin plus hardcoded gtag plus theme option
3. 3Static pages, microsites and HTML tools outside the CMS are tagged too
4. 4The live HTML matches what you installed, after host and CDN cache purges
5. 5Payment and login domains are listed as unwanted referrals
6. 6Internal traffic is defined, tested, then set to Active
7. 7Consent Mode sends gcs on every hit, and opt-outs show G100
8. 8A custom AI Assistants channel sits above Referral and is the primary group
9. 9ai_source and first_ai_source are populated and registered
10. 10Key events are limited to actions with business value
11. 11Ghost and bot hits are labelled, and reports default to clean data
12. 12Search Console and BigQuery are linked

• ✕Filters switched to Active on day one. Test first. An active filter that matches the wrong thing deletes real data forever.
• ✕Key event overload. If every scroll is a key event, nothing is. AI traffic will look like it converts brilliantly.
• ✕Reporting AI traffic without excluding yourself. Your own ChatGPT tests can be a meaningful share of a small site's AI referrals.

Frequently asked questions