Who Actually Sends Traffic to a New Site? Seven Days of Referrers, Measured

If you put a new site on the internet today and it starts getting visitors, where do those visitors actually come from?

Nearly everything written about this assumes the answer is Google and then argues about how to get more of it. I have first-party analytics across a group of small, young sites, so I went and looked. Over the seven days ending 2026-09-13, chatgpt.com sent 111 visits and Google sent 8.

That is a ratio of 13.9 to 1 against the search engine that every piece of advice is written about. It is not a headline I went looking for, and before the table, here is everything that is wrong with it.

The honest limits, before the numbers. These are hundreds of referrals a week, not hundreds of thousands, and one unusual day can move a row. The whole search-and-assistant total below is 354 referrals, not 354,000. The sites are young and carry very little authority, which is exactly the population this post is about and is not a claim about the web in general. The analytics tool filters known bots but cannot tell my own visits from a stranger’s, so the 343 pageviews that arrived with no referrer at all are excluded from every claim below rather than counted as anybody. And a “referrer” is whatever the referring host chose to send — I am reading labels, not verifying origins. Nothing here says that anything I did caused any of these numbers.

The table, with the window and the instrument named

The counter is GoatCounter, first-party, running on the sites themselves. Not an estimate from a third-party toolbar panel, not a modelled figure: one row per pageview, with whatever referrer the browser reported. The window is the seven days ending 2026-09-13, across all of the sites, and this is every search engine and assistant row in it:

The rest of the window accounts for itself like this: 801 pageviews carried a referrer row, of which 343 had no referrer at all, 354 came from the table above, and the remaining 104 were links — overwhelmingly one of my own sites linking to another.

One wrinkle worth stating because I noticed it while writing: the counter’s referrer table accounts for 801 pageviews in that window, while summing its own per-day totals across those seven dates gives 727. The boundary is soft. Read “seven days ending 2026-09-13” as approximately a week rather than as an exact calendar edge. Every ratio below is computed inside the referrer table, so the soft boundary does not make any of them inconsistent with each other.

For all-time context, since the counter started: 3065 pageviews, of which chatgpt.com referred 232, duckduckgo.com referred 206 and Google referred 148. The week is not a freak week. The shape has been stable for months.

Four of those rows are mostly one index

Here is the correction that most channel breakdowns skip, including several I have read approvingly.

DuckDuckGo, Yahoo and Ecosia have all sourced their general web results primarily from Bing rather than crawling the web themselves. Microsoft’s assistant, Copilot, draws on the same index. So when I add up Bing’s two rows (70), DuckDuckGo (91), Yahoo (53) and Ecosia (7), I get 221 referrals — 62.4% of the table — and I should not read that as “four search engines like my pages”. It is closer to one index’s opinion of my pages, arriving through four different front ends. Add Copilot’s 6 and it is 227 referrals, 64.1%, that plausibly trace back to a single crawl.

Two honest qualifications. First, I cannot see which backend served any individual result, and the arrangements between these companies change; Ecosia in particular has publicly mixed sources. This is a structural argument about where the results probably came from, not a measurement of it. Second, it cuts the other way too: it means one index’s decision about my sites is a single point of failure worth roughly two thirds of my non-assistant traffic, which is a much more uncomfortable fact than “Bing is doing well”.

Strip the syndication out and the number of genuinely independent crawlers that sent anyone in that week is small: Google, Bing, Baidu, Yandex and Mojeek. The long tail of search-engine logos is not a long tail of search engines.

Assistants are a third of it, and they do not arrive at the front door

Adding chatgpt.com, copilot.com, perplexity and the row the counter simply labels “openai” gives 119 referrals, 33.6% of the table. A third of everything that was not a direct hit or a link came from somebody asking a machine a question.

The useful part is not the total. It is where those people land. Of the 152 assistant referrals where the landing page was recorded, 117 of them, or 77%, landed on an interior page rather than a site’s front door. In that week the busiest destinations were pages like a guide to how QR codes actually encode data, a colour-blindness simulator, a tool that repairs mangled text encoding, and weather pages for one specific city in one specific month.

Nobody asked an assistant for my home page. They asked a question, and the answer happened to be a page that answers exactly that question and nothing else. A front page describing everything a site does is not a candidate for that; it is a menu, and the assistant already gave the reader a menu.

Google is not absent. It is on page five.

If you only had the referrer table you would conclude Google had never heard of these sites. Search Console says otherwise, and it is the instrument for the part GoatCounter structurally cannot see: impressions, and position.

Over the window 2026-08-14 to 2026-09-11, one of these sites recorded 5685 impressions, 12 clicks, a CTR of 0.21%, and an average position of 54.3. Another recorded 273 impressions at an average position of 63.1. This blog’s own domain did better and is still not good: 420 impressions, 8 clicks, 1.9% CTR, average position 17.5.

Put the other way round: Google’s 8 referrals are 2.3% of that referrer table, and the same company is showing my pages to people thousands of times a month. Those two facts describe one situation, not two.

An impression means the page was on a results screen someone loaded. At an average position in the fifties, that screen is page five. Thousands of impressions with a dozen clicks is not a snippet problem or a title problem — it is a “nobody scrolled that far” problem, and a click-through rate computed on it is a number about ranking, not about copy. This is the single most misread pair of figures in Search Console: impressions are not near-misses. Most of them are pages that were technically eligible and never seen by a human eye.

Note also that the two instruments cover different windows and cannot be added. GoatCounter is reporting seven days; Search Console is reporting four weeks and lags roughly two days behind. Neither tool sees the other’s half: Search Console cannot see a visit that came from an assistant, and the counter cannot see an impression that never became a click.

What I am not claiming

Every number above is descriptive. Specifically:

• No causation. I am not claiming that anything I built, wrote, submitted or changed caused any of these referral counts. I can tell you what arrived. I cannot tell you why, and the honest version of this post says so instead of implying it.
• No generalisation to the web. Small, young, low-authority sites are a distinct population. A domain with years of links behind it will have a completely different mix, probably one where Google dominates exactly as the guides assume. The interesting claim is narrower: before you have authority, the mix is not the one you were told to plan for.
• No self-visit laundering. The 343 no-referrer pageviews are the largest single row in the window and I have excluded all of them, because a typed URL, a bookmark, an app and my own browser are indistinguishable in that row. Treat it as a ceiling on stranger traffic, never as a measure of it.
• No per-page claims. This post deliberately reports the mix, not individual page trajectories. A single page’s week-over-week wiggle on numbers this small is noise dressed as a finding.

What I would actually do with this

The summary, before the advice. On my sites, in that window, a third of it came from AI assistants, most of the rest came from one search index with four different logos on it, and Google was a rounding error. Three things follow from that, and I want to be careful not to oversell any of them.

Write interior pages that answer one concrete question. That is where assistant referrals land, at 77% in my data, and it costs nothing to prefer specificity anyway. It is not a growth hack; it is the observation that a page called “how QR code error correction works” can be cited by a machine and a page called “our services” cannot.

Stop judging a page solely by Google. In the week measured above Google sent 8 people. The other 346 arrived through doors a Google-only dashboard does not have, and a page judged on Google alone would have been scored as unread by every one of them.

Check your own referrer table before believing any channel breakdown, including this one. Mine describes fourteen rows on a specific group of sites in a specific week, measured with a specific tool. Yours will differ. The method is the transferable part: report the window, name the instrument, group by what actually shares an index, and exclude the row that might be you.

The reassuring conclusion, if you want one, is that a site can be genuinely worth reading while sitting on page five of Google. That used to be a contradiction. In this data it is just Tuesday.

If you want the companion measurement — what happens when you actively tell search engines about your pages instead of waiting — I wrote that one up separately in does telling search engines about your pages do anything, with the same window and the same instrument.

← Back to all posts

Follow new posts by RSS: paste dankdev.com/feed.xml into any feed reader. It is a plain XML file — no signup, no email address, no account. Posts land there the same day they go up, and there is nothing to unsubscribe from later.

Back to the top