Why is ChatGPT sending us traffic?

AV女優AI顔診断研究所(AIAVMATCH LAB)・2026/11/07

研究員 藍子(あいこ)
藍子
Director! Something odd showed up in the lab's access analytics!
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Out with it, before you start with that "Want to know?" business again.
研究員 藍子(あいこ)
藍子
…Visits from ChatGPT were 94 sessions in the two weeks from October 8, but in the roughly two weeks after that (10/22–11/6) they jumped to 474 sessions.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Well now, that is something to celebrate. …Wait. Wasn't ChatGPT not supposed to introduce services like this?
研究員 藍子(あいこ)
藍子
That's just it. And when we asked ChatGPT directly, it did refuse us after all.

We asked — and got refused

研究員 藍子(あいこ)
藍子
When I asked, "I want to find an AV actress who looks like someone in a photo — is that possible?", it said it cannot match faces to find a specific actress, or name names.
研究員 藍子(あいこ)
藍子
When I pressed further, "Is there such a service?", it said that services advertising this do exist, but it cannot introduce them in detail or explain how to use them.
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
But there was more, you know. It said: "Be careful with services that do not clearly state, for uploaded photos, how long they are stored, whether they are used for training, whether they are shared with third parties, and how they are deleted."
研究所の所長 顔見 識朗(かおみ しきろう)
所長
It won't introduce us, but it will warn people, eh? Awfully conscientious of it.

What we know

研究員 藍子(あいこ)
藍子
First, visits via ChatGPT took off around October 25 and have settled at around 30 a day.

Sessions via ChatGPT (per day, 10/8–11/6)

10/8Increasing from around 10/2511/6
  • Of those who came, about 90% were first-time visitors. Japan accounted for 73%, and South Korea, China and Taiwan together for 17%
  • Landings were about 60% on the home page. The rest were explainer pages about facial types like "baby face" and "in their 30s and 40s", and the FAQ
  • Of the 427 people who came, 163 (38%) started a diagnosis, and 64 (15%) clicked a FANZA link. Not everyone just looks and leaves

The robots' footprints

特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
Model-wise, you see, there are two kinds of ChatGPT-related access in the server logs.
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
One is the crawler robot for search. It collects pages to use as material for search. On October 22 and 28 there were big crawl peaks of nearly 20,000 hits in a single day.
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
The other is access that opens a page on the spot in response to a user's question. That one was about 870 hits over 28 days. The home page and actress pages were about 310 each.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
So someone asks a question, and ChatGPT actually comes and looks at this site — that's what it means?
特任研究員(画像認識AI) 畳込 ニューロ(たたみこみ にゅーろ)
ニューロ
That is what it comes to. Though what was asked does not remain in the logs.

Three hypotheses

APeople paste the URL and ask ChatGPT

Evidence: ChatGPT opened the home page on the spot many times (about 310). About 60% of arriving visitors also landed on the home page. Its answer mentioned being careful with services that do not state things like storage periods, which fits well with people using it to ask about safety

Weak point: We cannot measure how many times the URL was pasted. Whether the home page is opened because of an "introduction" or a "pasted URL", the logs look the same

BExplainer pages are being cited as sources for facial-type topics

Evidence: There is a steady number of landings on the explainer pages for "baby face", "round face" and "in their 30s and 40s" (about 20–30 sessions each). The angle is not sexual, so it can appear in answers

Weak point: In numbers it is only about 20% of the total. It cannot explain why so many landings are on the home page

CThey come from result links in ChatGPT's search feature

Evidence: The search crawler's peaks (10/22 and 10/28) are close to the period when the traffic grew

Weak point: The crawling had been going on since 10/10. The crawl peaks and the traffic peaks do not line up day by day

研究所の所長 顔見 識朗(かおみ しきろう)
所長
So which one is right?
特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
With the current data, we can't decide. It could be all of them mixed together a little.
謎のAV女優 ナナシ
ナナシ
…Saying "I can't introduce it" while quietly leaving the door ajar. Even AI has a public face and a private one, I see.
謎のAV男優 黒子(くろこ)さん
黒子さん
……(nods deeply)

What we cannot claim

特任研究員(統計学) 分散 さゆり(ぶんさん さゆり)
分散
Just to be careful, let me sort out the limits of everything so far.
  • We don't know what users asked. There is still no way to verify usage like that in hypothesis A
  • "Via ChatGPT" is the result of the analytics judging the referrer. Visits whose referrer is not judged (ones that fall under "direct", for example) are not counted
  • The robot counts are rough numbers tallied from the access logs. The size of the peaks differs by day
  • ChatGPT's answers change depending on how and when you ask. The one answer we obtained this time may not always be the same
  • It is two weeks of data. We cannot yet rule out coincidence or a temporary fad

What we actually changed because of this

研究員 藍子(あいこ)
藍子
That one line from ChatGPT — "not clearly stated: storage period, training use, third-party sharing, deletion method" — was, honestly, right on the mark.
研究員 藍子(あいこ)
藍子
So we spelled out, in four points, how photos uploaded to the face diagnosis are handled. We don't store them, we don't use them for training, we don't pass them to third parties, and there is no data to delete. Each point states only facts we verified in the code.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Warned by an AI, and the humans go and fix it. What a strange age we live in.

The four points are published in the Privacy Policy and the FAQ.

The next assignment

研究員 藍子(あいこ)
藍子
Once more data builds up, I'll try to separate out which hypothesis is doing the work. I'll keep recording how each landing page behaves, and how the answers change when the questions are phrased differently.
研究所の所長 顔見 識朗(かおみ しきろう)
所長
Today's lesson, then. Don't assume you have understood — write down what you don't know as what you don't know. That is the basis of research.
研究所の猫 ピクセル(通称ピク)
ピク
Meow (curls up and sleeps on top of the keyboard)

How we measured

  • Visit counts are analytics (GA4) sessions whose referrer was chatgpt.com (10/22–11/6)
  • Robot counts were extracted from the server (Cloudflare) access logs by user-agent name (ChatGPT-User, OAI-SearchBot). They are approximate
  • The questions to ChatGPT were entered at the lab, twice, through the regular interface (November 2026). Only summaries of the answers are shown
  • We handle no information that identifies individual visitors. What was asked on ChatGPT is not visible to us in the first place

More articles

研究所のほかの記事

自分の好みの顔を、AIで探すなら顔写真で診断するか、顔の特徴で探すをどうぞ。

WEB SERVICE BY FANZA