A third of our visitors now arrive from an AI assistant

Here is every channel GA4 reported for the window, complete — the table said “1–6 of 6”, so nothing below is truncated to make a point:

That is 46.1 sessions a day. This is a small site and we would rather say so than let a percentage imply otherwise — 32.25% of a modest number is still a modest number. What makes it worth writing down is the position of that row, not its size.

One oddity, left in rather than tidied away: those six rows add up to 1,289, which is one session short of the 1,290 GA4 prints on its own total line. GA4 calculates every percentage in the report against 1,290, so we have kept its figure and its shares exactly as the tool reported them. We mention it because a report whose numbers reconcile perfectly is usually a report somebody has quietly rounded.

“AI Assistant” sounds like a category. On this property it is very nearly a single name:

chatgpt.com is 99.3% of the channel and 32.02% of the entire site. Perplexity does not appear in the table at all. Gemini and Copilot appear with 2 and 1 sessions, which is shape, not a finding — you cannot conclude anything from two visits and we are not going to try.

This is why the second dimension matters. Had we stopped at the channel total we would have written “answer engines send us a third of our traffic” and been misleading by omission. One product sends us a third of our traffic. That is a different sentence with different consequences.

Referral counts on their own are easy to over-read, so here is the same cohort measured on engagement rather than volume, against google over the identical window:

It stays longer and does more, and it engages very slightly less often. It is not obviously worse traffic, which is the thing people usually assume about a referral they did not earn through a ranking.

1. That we know why. We can see a referral arrive. We cannot see the prompt that produced it, what was asked, or what else was suggested in the same answer. That conversation is not in anyone’s analytics.
2. That we caused it. SlowDen has done deliberate work on being readable by machines, and it would be a tidy story. We have no measurement connecting the two, and picking a cause to fit a number we like is how sites end up optimising for something that was never the mechanism.
3. That this generalises. One property, one 28-day window, one catalogue of free browser games. If you want to know whether it is true of your site, the second FAQ below tells you exactly where to look rather than asking you to take our word for it.
4. That it is unambiguously good news. See the next section.

Direct traffic is 37.91% of sessions here and chatgpt.com is 32.02%. Those two together carry roughly 69.93% of everything. One of them is a third-party product that can change how it links out, when it links out, or whether it links out at all, with no notice, no appeal, and no diagnostic panel for us to read afterwards.

Search has always had a version of this problem, but it also has an entire public apparatus around it — documented guidelines, a search console, indexing reports, and reasons attached to decisions. Nothing equivalent exists for an assistant’s referral behaviour today. A site leaning on it is leaning on something it cannot inspect.

The honest answer is: measure your own, and keep your pages readable. A language model can only quote text that exists as text. A game embedded in a canvas or an iframe is, to anything that does not run the game, an empty box — whatever is written around it is the entire surface.

On SlowDen that means every game page carries written material rather than just a player: 311 of the 311 games have a description, 122 have how-to-play instructions, and 86 carry genuine questions with answers. Three examples, listed alphabetically rather than ranked — we count no plays here, so there is no “top” list to give you:

• 195-0 World Draft — by 195-0 World Draft. Its page carries a description, a how-to-play section and 3 real questions with answers.
• 2048 Cupcakes — by 2048 Cupcakes. Its page carries a description, a how-to-play section and 3 real questions with answers.
• AI Party Game Generator — by AI Party Game Generator. Its page carries a description, a how-to-play section and 3 real questions with answers.

If your own analytics shows nothing at all in this channel, the first thing to rule out is the boring one: whether the page is being measured in the first place. We managed to ship blog posts with no analytics tag on them and wrote up how that happened. If you are still deciding where to put a finished game, we also compared the free places to publish an HTML5 game.

Google Analytics 4, property G-2SGFC0Z298, Reports → Acquisition → Traffic acquisition, dimension Session primary channel group with Session source as the secondary. Read on 20 September 2026. The window caption was read back from the chart verbatim before any figure was quoted: “Last 28 days 23 Aug - 19 Sept 2026”.

Every figure above was read out of the table as text. None was read off a screenshot — on the same afternoon a scaled screenshot of the same report rendered an engaged time of 53s as 39s, which is the kind of error that survives a review because it looks entirely plausible.

Catalogue figures (311 games, 200 partner and 111 developer-submitted, 217 studios) were counted from the live catalogue on 20 September 2026 and change as games are added. SlowDen made none of them.

Does ChatGPT actually send traffic to websites?

On this site, yes, and it is not a rounding error. Over the 28 days 23 Aug - 19 Sept 2026 we recorded 413 sessions with chatgpt.com as the referring source — 32.02% of all traffic to slowden.com, and 2.2× what google sent over the same window (192). We cannot tell you whether that generalises to any other site; it is one property, one window, and we are reporting it rather than projecting from it.

How do I see AI assistant traffic in my own Google Analytics?

In GA4, open Reports → Acquisition → Traffic acquisition and leave the dimension on Session primary channel group. If your property is recent enough you will see a channel called AI Assistant sitting alongside Organic Search and Direct. Then change the secondary dimension to Session source — that is the step that matters, because the channel total hides whether you are being sent traffic by one product or by ten. Read the chart caption back before you quote any figure from it, and read the numbers out of the table rather than off a screenshot.

Why does an AI assistant link to a browser game site at all?

We do not know, and we are not going to guess in public. We can see the referral arriving; we cannot see the conversation that produced it, which prompt was typed, or what else was suggested alongside us. Anyone who tells you they have decoded this from their own analytics is reading more than the data contains. What we can say is what the traffic did after it arrived: it engaged at 67.07% with 46s of average engagement time, against 68.23% and 33s from google.

Is a third of your traffic from one source a good thing?

It is a real number and a real risk at the same time, so we are recording it as both. Direct (37.91%) and chatgpt.com (32.02%) together carry about 69.93% of all sessions here. One of those two is a third-party product whose referral behaviour we do not control, cannot see into, and have no way to appeal if it changes tomorrow. A site that depends on it is one product decision away from a very quiet month.

Did you do something to make this happen?

Not that we can demonstrate. SlowDen does publish plain readable text on every game page — 311 of the 311 games carry a written description, 122 carry how-to-play instructions and 86 carry real questions and answers — and that is the kind of material a language model can actually quote. But we have no measurement connecting any of it to this channel, and saying otherwise would be inventing a mechanism to fit a number we like.