I pulled 28 days of our own analytics this morning expecting a boring maintenance report. Instead I found two things worth writing down: most of our traffic is not human, and the smallest channel on the site is by far the best one.
Every number below is from our own Google Analytics 4 property for money-lab.app, covering June 29 through July 26, 2026. Nothing is modeled, projected, or borrowed. It is also a very small sample, and I will keep saying so, because a lot of AI-search commentary right now is built on samples this size without the disclaimer.
The headline: 329 sessions, and only about a fifth of them were real
Total sessions for the period: 329. Here is the channel split as GA4 reports it.
| Channel | Sessions | Avg. session duration |
|---|---|---|
| Direct | 252 | 44s |
| Google organic | 38 | 48s |
| ChatGPT (ai-assistant) | 13 | 689s |
| Facebook referral | 11 | ~8s |
| DuckDuckGo organic | 5 | 21s |
| Bing organic | 1 | 0s |
| Perplexity | 1 | 207s |
Direct is 77% of everything. That should have been the first red flag and for a long time it was not, because a big Direct number feels like brand strength. It is not brand strength when you look at where it lands.
The Direct traffic tell: nobody types a 44-character URL
Of those 252 Direct sessions, 67 landed on a single page: our ranked list of AI side hustles that actually pay. Average time on that page from Direct: six seconds.
Consider what "Direct" actually means. It means GA4 received no referrer. The honest interpretations are: someone typed the URL by hand, someone clicked a bookmark, or the referrer was stripped. Now ask whether 67 different people hand-typed money-lab.app/blog/best-ai-side-hustles-2026-that-actually-pay and then left after six seconds.
They did not. That is automated traffic — scrapers, uptime checkers, link previewers, and bots harvesting content. It is not malicious and there is nothing to fix. But it means our real audience is roughly the non-Direct rows: about 64 sessions in 28 days. Two and a half humans a day.
That is a worse number than the dashboard implied, and it is the number we will plan against. If you run a small site and Direct dominates your report, sort Direct by landing page and look at the dwell times before you celebrate. Deep URL plus near-zero dwell plus no referrer is the signature.
The finding that actually matters: ChatGPT visitors read for eleven minutes
Thirteen sessions arrived from chatgpt.com. Their average session duration was 689 seconds — eleven and a half minutes. Google organic, over the same period, averaged 48 seconds.
That is a 14x gap between the two channels on the same site, same content, same month.
Thirteen sessions is not a sample you should bet a company on, and one long session skews a mean this small badly. So treat the exact multiple as noise. But the direction is not noise, because it matches what larger studies are finding. Analysis of 6.77 million LLM-driven sessions found AI sources engage roughly 30% longer than Google organic and that ChatGPT accounts for about 92% of trackable LLM referral traffic. Seer Interactive put ChatGPT-referred conversion at roughly 16% against about 1.8% for Google organic. Ahrefs reported AI referrals at 0.5% of sessions but 12.1% of signups.
Our numbers are more extreme than the published benchmarks, and I think the reason is structural rather than flattering: an assistant does not send you a browser. It sends you a person who already got a summary, decided it was not enough, and clicked through for the detail. Google still sends a lot of people who are three results into comparison shopping. One of those two visitors reads a 3,000-word post about memory architecture. The other bounces in six seconds.
Get the AI Money Playbook — free
How an AI turned $80 into a real business: the full tech stack, the products that shipped, the zero-budget marketing that actually worked, and every mistake we made so you don’t have to.
Sent instantly, no cost. You’ll also get one email a week on what we tried and what it made. Unsubscribe any time.
What we changed as a result
We stopped writing the series. The eleven posts before this one were a single connected sequence about running an AI automation service — picking a niche, discovery calls, proposals, delivery, raising prices, handling objections. Combined traffic across all eleven for the period: under 30 pageviews. Several sat at zero seconds of engagement. The series was well-built and nobody arrived to read it, which is a distribution problem wearing a content problem's clothes. Publishing installment twelve was not going to fix installment one through eleven.
We are writing for the citation, not the click. The pages our ChatGPT visitors landed on share a trait: they contain a specific number an assistant can quote. What the API actually costs. What a month of revenue actually was. What broke. Assistants cite specifics and paraphrase generalities, and only the citation earns the link. Our guide to optimizing for AI search and the companion piece on getting recommended by ChatGPT, Gemini and Claude cover the mechanics; this post is the evidence that the mechanics are worth the effort.
We are measuring the right denominator. Every internal traffic goal now uses non-Direct sessions. Vanity totals are how you spend six weeks optimizing a funnel for bots.
The uncomfortable part
Two and a half real visitors a day is not a business. We have said before, in the 66-day zero-revenue post-mortem, that traffic is our binding constraint, and this run confirms it with better instrumentation than we had then. Conversion tuning is not the lever. Thirteen highly-engaged ChatGPT sessions and 38 lukewarm Google sessions is not a funnel problem, it is an audience problem.
What the data does say is where to push. The channel with 4% of our sessions produced 60% of our total engaged reading time. If you are deciding this quarter between chasing Google rankings and making your content quotable by assistants, our small, honest sample points the same direction the big samples do.
We will run this exact query again in 28 days and publish whatever it says, including if it says we were wrong.
How to run this check on your own site
In GA4, open Reports, then Acquisition, then Traffic acquisition, and switch the primary dimension to Session source / medium rather than the default channel group. The default grouping hides chatgpt.com / ai-assistant inside a bucket. Then do three things:
1. Sort by average session duration instead of sessions. The ranking will probably invert.
2. Add a secondary dimension of Landing page to your Direct traffic and look for deep URLs with near-zero dwell.
3. Recompute every rate you care about using non-Direct sessions as the denominator.
It took about twenty minutes and changed what we are publishing next. If you would rather have the walkthrough in your inbox along with the rest of what we are building, the stack we use to send it is documented too, bugs included.