How to Read GA4 as a Publisher: The 4 Metrics That Pay

If you run an ad-funded site, most of GA4 is noise. The dashboard is built for lead generation and ecommerce, so it foregrounds users, sessions and conversions, none of which pays your invoice on its own. Your revenue comes from ad impressions, and impressions scale with pageviews and viewability, not with how many people showed up. There are four metrics in GA4 that map cleanly to what Raptive, Mediavine or AdSense pay you: engaged sessions and engagement rate, views and views-per-session, average engagement time, and traffic by channel and source. All four are available out of the box under Reports > Engagement and Reports > Acquisition. Read those four well, in the right combination, and ignore almost everything else.

How should a publisher read GA4?

Read it as a revenue instrument, not a popularity contest. The most common mistake is opening the home overview, seeing a big user or session number, and stopping there. Sessions are the container, but ad impressions live inside the container. A visitor who lands on one page, reads for eight seconds and leaves has generated one pageview and next to no ad money. A visitor who reads three pages over four minutes has generated three or more ad slots, several viewability windows, and any refresh impressions your ad partner runs on dwell.

So the correct reading order is quality first, volume second, money third. Start with engagement rate to check the traffic is real and awake. Move to views and views-per-session to see how much ad inventory that traffic actually produced. Layer in average engagement time to gauge viewability and refresh potential. Then split all of it by channel and source to understand where the good traffic comes from, and increasingly, what AI is doing to it. Everything else in GA4 is either a supporting detail or a distraction.

Which GA4 metrics actually affect ad revenue?

Four, and they line up with the mechanics of display advertising.

Engaged sessions and engagement rate

GA4 defines an engaged session as one that meets any one of three conditions: it lasts longer than 10 seconds, it fires at least one key (conversion) event, or it has at least two pageviews or screenviews. Engagement rate is engaged sessions divided by total sessions, and it is the arithmetic complement of bounce rate, so the two always sum to 100 percent (Google Analytics Help). For an ad-funded publisher this is the truest quick read of traffic quality, because a session that bounces in under 10 seconds with a single pageview earns you almost nothing. The 10-second threshold is not fixed: you can raise it to as much as 60 seconds per web data stream if you want a stricter bar (Playwire).

Views and views-per-session

This is the metric most publishers underweight. Your ad slots are attached to pages, so a doubling of views-per-session is, roughly, a doubling of ad inventory from the same audience. Two sites with identical session counts can differ enormously in revenue if one runs at 1.2 views per session and the other at 2.8. Views and views-per-session sit in Reports > Engagement > Pages and screens, which is also where you find your best-earning pages.

Average engagement time

Dwell time is not a vanity stat for a publisher. Longer time on page means more of your ad slots enter the viewable window, which is what advertisers actually pay for, and it opens the door to refresh impressions on partners that support them. Average engagement time is on the Engagement overview and should be read alongside views: a page with high views but low engagement time may be pogo-sticking traffic that never lets the ads render.

Traffic by channel and source

The first three metrics tell you how good your traffic is. This one tells you where it comes from, and it is where the entire AI question gets answered. It lives in Reports > Acquisition > Traffic acquisition, broken down by default channel group and by session source / medium.

A worked example: why sessions lie

Take two channels sending 10,000 sessions each in a month.

Same session count, double the revenue. If you had judged these two sources on sessions alone you would have called them equal. The revenue read is views multiplied by your page RPM, and the quality read is the engagement rate that predicts it. This is the arithmetic that should govern where you invest content and promotion.

How do I find my best-earning pages in GA4?

Go to Reports > Engagement > Pages and screens. This report ranks your URLs by views and shows average engagement time per page alongside them. For a publisher, your best-earning pages are not simply your most-viewed pages; they are the pages with the best combination of views and engagement time, because that combination is what turns into viewable, refreshed impressions.

A practical routine:

1. Sort by views to find your traffic drivers.
2. Add average engagement time as a second lens. A high-views, low-engagement page is fragile revenue that may collapse if a search feature changes.
3. Multiply views by your ad partner's page RPM to rank pages by estimated revenue rather than raw popularity. GA4 will not do this maths for you unless you have Ad Manager or Publisher Ads linked, so most publishers do it in a sheet.
4. Cross-reference the top pages against the traffic acquisition report to see which channel feeds each one. A top earner fed entirely by Organic Search is exposed to exactly the AI-summary decoupling described below.

This is also the report where you spot a page whose views are holding but whose engagement time is sliding, an early warning that the audience is changing even before the totals move.

Where do I see which sources send my traffic in GA4?

Reports > Acquisition > Traffic acquisition, viewed by default channel group and by session source / medium. GA4 sorts traffic into rule-based default channels derived from source and medium: Organic Search, Referral, Direct, Paid Search, Organic Social, Email and others. One thing to know before you rely on it: the default channel group cannot be edited (Google Analytics Help). If you want a bespoke grouping, for example a single bucket that captures all AI referrers, you have to build a custom channel group rather than modify the default one.

The new AI Assistant channel, and what it misses

On 13 May 2026 Google added a native AI Assistant default channel to GA4, with broad rollout across properties following in early June. When GA4 detects a referrer matching a recognised AI assistant, it stamps the medium as ai-assistant and files the session under AI Assistant with no configuration required. As of June 2026 the recognised list is ChatGPT, Gemini, Deepseek, Copilot and Grok (Search Engine Journal). Note that Google's original launch post also named Claude, but Claude does not appear in the current published definition, so treat the list as fluid and check it against your own data. That is genuinely useful, but it has three limitations a publisher must hold in mind.

• Perplexity is not on the list. Perplexity visits stay in Referral, so your AI Assistant number is already an undercount.
• It is forward-only. GA4 does not retroactively reclassify historical data, so any AI traffic that arrived before 13 May 2026 remains labelled Referral or Direct. This is precisely why year-on-year AI comparisons in GA4 are unreliable.
• Referrer stripping sends a large share to Direct. When a user clicks a link inside a ChatGPT, Perplexity or Gemini mobile app or embedded browser, the referrer header is frequently stripped, so GA4 sees no referrer and files the session under Direct. That systematically undercounts AI traffic on top of the two gaps above.

How to recover the AI traffic GA4 hides

Because the default channel misses Perplexity and the stripped-referrer app clicks, build a custom channel group or an Exploration with a regex on session source or page referrer that matches the AI hosts directly. A working pattern is chatgpt.com|openai.com|perplexity.ai|gemini.google.com|copilot.microsoft.com|claude.ai and similar. That catches the referrers the default channel ignores. It will not fully recover the app clicks that arrive with no referrer at all, which is the honest limit of any GA4-only approach and the reason a purpose-built read exists. This is the single hardest part of reading GA4 as a publisher today, and it is exactly the gap Ramprt is built to close: it reads your existing GA4 read-only, with no new tracking script, and applies AI detection on top so the AI traffic in Referral and Direct stops being invisible. There is a live free demo at ramprt.io/demo.

What is AI actually doing to my search traffic?

The channel report tells you what arrives. The AI-Overview effect explains what is quietly not arriving. The Pew Research Center analysed 68,879 Google searches from 900 US adults in March 2025 and found that searches which produced an AI summary cut the click-through rate to traditional results roughly in half: users clicked a traditional search result in 8 percent of visits when an AI summary was present, versus 15 percent when it was not (Pew Research Center). Clicks on the links inside the AI summary itself happened in just 1 percent of visits. So the summary rarely sends the click on to you, and it halves the odds of a click to the normal results below it. Google has publicly disputed methodologies of this kind, arguing that AI Overviews send higher-quality clicks, so treat the exact figures as directional rather than gospel, but the direction is consistent across multiple independent studies.

This is not a fringe case. Pew found AI summaries appeared on about 18 percent of all Google searches studied, and 58 percent of users encountered at least one in March 2025. Users were also more likely to end their browsing session entirely on a page with an AI summary, 26 percent versus 16 percent without one (Pew Research Center). And when the summary does cite sources, the traffic is diluted: 88 percent of AI summaries cited three or more sources. The practical read for a publisher is that impressions in Search Console can rise while clicks fall, and rankings can hold while organic sessions decline. That decoupling shows up in your GA4 Organic Search channel as a falling session count against a stable, valuable set of pages, which is why you read the channel report and the Pages and screens report together.

Is the AI traffic that does arrive worth anything?

Yes, and it is worth measuring rather than assuming. Third-party data suggests AI-referred visitors monetise and engage above baseline: Similarweb reports that ChatGPT referral traffic converts at about 7.1 percent, second only to paid search, and that visitors from AI tools tend to spend longer on-site than organic search visitors (Similarweb). The scale is real too: AI platforms generated over 1.13 billion referral visits in June 2025, up 357 percent year on year, though Google search still drove roughly 191 billion referrals over the same period (TechCrunch, on Similarweb data). You can test the quality claim in your own data: compare average engagement time and engagement rate for the AI Assistant channel, plus your custom AI regex segment, against your site average. If AI visitors dwell longer, they are producing more viewable impressions per session, and that is money.

What about the crawlers taking my content?

GA4 only sees humans, so it will never show you the bots reading your site to build AI answers. That question lives in your server logs, and there are two things worth knowing so you read the logs correctly.

First, there is a clean split between training crawlers and answer crawlers. OpenAI's GPTBot trains models, while OAI-SearchBot powers ChatGPT search results and ChatGPT-User handles live user-initiated fetches. Anthropic splits ClaudeBot (training), Claude-SearchBot (search index) and Claude-User (live fetch). Perplexity splits PerplexityBot (indexing) and Perplexity-User (live). Blocking the training crawlers does not remove you from the AI search surfaces that can still send referral traffic, so a blanket block can cost you the very AI referrals you were trying to measure.

Second, do not trust the user-agent string. It is self-reported and trivially spoofable. The reliable check is matching the requesting IP against the official verified-crawler ranges each vendor publishes. OpenAI publishes JSON range files at openai.com/gptbot.json, openai.com/searchbot.json and openai.com/chatgpt-user.json (OpenAI); Perplexity publishes equivalents at perplexity.com/perplexitybot.json and perplexity.com/perplexity-user.json (Perplexity). Anthropic historically relied on reverse DNS but now also publishes an IP list for ClaudeBot verification (Anthropic), and Google publishes its crawler ranges too. The user-agent is a hint; the verified IP is the proof.

One robots.txt point that trips up publishers: Google-Extended is a control token, not a crawler. It only governs whether your content is used to train Gemini and generative products. It does not stop you appearing in AI Overviews, because those draw on the live Google Search index, and it uses no separate request-time user-agent, since crawling is still done by Googlebot (Google Search Central). Google has confirmed that disallowing Google-Extended has no effect on Search rankings or indexing, so blocking it keeps you ranking and AI-Overview-eligible while opting out of model training only.

Which GA4 numbers can I safely ignore as a publisher?

Plenty, and ignoring them frees you to read the four that matter.

• Total users in isolation. It tells you nothing about impressions. Two visitors reading five pages each beat ten visitors bouncing.
• Raw session counts on their own. Always pair sessions with engagement rate and views-per-session. Sessions without those two are a headline with no story, as the worked example above shows.
• Year-on-year AI channel comparisons in GA4. Because the AI Assistant channel is forward-only with no historical backfill, any AI-versus-last-year figure inside GA4 is comparing a labelled present against an unlabelled past. It is not a real trend.
• Conversion (key) events, unless you have set up meaningful ones. For a pure ad-funded site with no signup or sale, out-of-the-box conversions rarely reflect revenue. Do not let an empty conversions column make you feel your traffic is failing.
• Bounce rate as a separate line of enquiry. It is just 100 minus engagement rate. Read engagement rate and you have already read bounce.

The discipline is simple. Every time GA4 shows you a number, ask whether it changes your estimate of viewable ad impressions or the value of the traffic behind them. If it does not, it is context at best and a distraction at worst.

Putting it together

Read GA4 in one loop: engagement rate to confirm the traffic is awake, views and views-per-session to size the ad inventory, average engagement time to judge viewability, and channel and source to see where it comes from and what AI is doing to it. Translate any shift into money by multiplying views by your page RPM rather than counting sessions. And accept that GA4 alone cannot fully see AI: the AI Assistant channel misses Perplexity, ignores everything before 13 May 2026, and loses app clicks to Direct. That last gap is the one to solve deliberately, whether with a custom channel group and regex, or a purpose-built read layered on your existing GA4. If you want to see that AI view without touching your tracking, the Ramprt free demo shows exactly what AI is doing to your traffic and revenue.