What 57 Million Delivered Traffic Sessions Actually Look Like in GA4

If you have ever bought traffic, you already know the problem: everyone sells you an average.

"50,000 visits, 3:00 average session duration." That number is real, and it is also useless, because an average hides the only thing that matters - the distribution. A provider whose traffic is half instant-bounce and half ten-minute sessions has exactly the same average as one where every visitor reads for ten minutes. You cannot tell them apart from the sales page.

We can tell you apart, because we run one of these services and we keep the ledger. Below is what 30 days of our own delivery data actually looks like, published with the failures in it.

What we measured, and what we did not#

Source: our production delivery ledger, every window ending 12 November 2026, 21:21 UTC - one row per attempted visitor session, with the outcome, the dwell time, the page count and the device. No sampling, no cherry-picking: every attempt in the window is counted.

What is not in here, stated plainly: this is our own delivery data, not a read-back from a customer GA4 property. It contains no conversion or revenue data - this study is about session quality, not outcomes. And the visits are simulated by design: they are browser-driven sessions with configured behaviour, not real human readers. We are a traffic delivery tool and we do not describe them as anything else.

The volume and the delivery rate#

A failure rate of 0.134% means roughly 1,342 failed sessions per million. If your dashboard says "sessions" and does not say "landed in GA4", this is the gap you are not seeing.

What actually failed - and why that matters#

Publishing the failures is the most useful part of this study, because failure reasons tell you what a provider is protecting against:

Read that again: nearly three out of four failures (72.3%) were the GA4 endpoint timing out, and together with the closed quota windows and the proxy limit, 97.8% of failures were our own pipeline saying no. These are not "visits that bounced". They are attempts we refused to count as delivered because we could not confirm them.

That is the difference between requested and landed. A provider who counts requests as visits will report a higher number than we do, and will be wrong.

Why averages hide everything: the duration distribution#

From 5,933,529 delivered sessions (3-day window):

The average and the median sit close together - a sign that the distribution is not being dragged by a handful of very long sessions. But the histogram is the interesting part:

83.73% of sessions lasted five minutes or longer. The 10:00 mark is a configured ceiling: 1,511,017 sessions (25.5%) end at exactly 600 seconds, which is why the 90th percentile is 10:00 even though 27.44% of sessions fall in the 10 min+ bucket. The distribution is bimodal at the bottom: either a visitor leaves inside 30 seconds (8.93% - almost exactly our 8.92% bounce rate, and these are the sessions that bounce on the landing page), or the visit is a normal, readable one.

The two figures are not merely close - they are the same population. We checked it directly: every session flagged as a bounce ended within the first 30 seconds (529,515 of 529,515 in the window), and only 320 of the 529,835 sub-30-second sessions were not flagged as a bounce. So in our delivery, "a bounce" and "left within 30 seconds" describe the same behaviour, and it is concentrated at the very front of the funnel rather than spread across the session.

Now compare that with a provider whose sessions are, say, 40% under ten seconds. Their average might still be respectable - and the shape would tell you immediately that you are buying a pile of instant exits. Ask any provider for their duration histogram. If they only have an average, they are hiding the distribution from you.

The device and country cuts of the same data are in our companion piece Where 57 Million Traffic Sessions Actually Came From - including the finding that desktop traffic bounced almost four times as often as mobile traffic in our own delivery.

Bounce rate and pages per session#

A bounce rate of 8.92% is what a normal, engaged traffic mix looks like on a content site. And note the arithmetic: with 1.12 pages per session as the average and only 3.97% of sessions going beyond one page, the multi-page sessions are doing a lot of work in the average. That is normal for a landing-page-heavy mix, and it is worth checking against your own site: if your pages are deep, ask what the provider's multi-page rate is.

Devices: your traffic is mostly mobile#

Four out of five delivered sessions came from a mobile device. If your analytics shows desktop-heavy traffic and you are buying mobile traffic - or the other way round - that mismatch is a configuration problem, and it is visible in the first day of delivery.

Countries#

205 countries in total. Note that the distribution is not a straight mirror of "Western traffic" - South Africa, Poland and Turkey each deliver more sessions than Australia or Germany in this window. If you are buying traffic for a market outside the obvious five, the country mix is where the cheapest volume usually sits.

The five questions to ask any traffic provider#

Distilled from the numbers above:

1. How many sessions landed, and how many were only requested? We report 99.866% delivered. Ask for the denominator.
2. Show me the duration histogram, not the average. One number can hide any shape.
3. What is your bounce rate, and what does a bounce look like in the data? Ours is 8.92%, and it is concentrated in the first 30 seconds.
4. What is the device split and the country split? 81% mobile, 205 countries.
5. What failed, and why? Our top failure reason is our own measurement pipeline saying no (the GA4 endpoint timing out, 72.3%) - which is a very different thing from visitors who bounced.

A provider that answers all five has nothing to hide. A provider that answers only with an average is telling you the one number that cannot be checked.

We turned these five into a checklist you can take to any provider - The Five Questions to Ask Any Traffic Provider - and the same questions are the ones we ask about our own delivery, which is why we publish the answers.

What this study does not claim#

• It is not a comparison of providers. We measured ourselves.
• It contains no conversion, revenue or ranking data.
• It describes simulated visits. They behave like visits - configured country, device, referrer, page depth, dwell time - and they are counted honestly, but they are not human readers, and nobody's should be.
• The numbers cover 13 October to 12 November 2026 (21:21 UTC) and will move.