US market research

Have Personal Shopping Agents Visited Korean Sites? Comparing Bot Logs from Two Sites

Published

This article was first published in Korean by Intrix Lab on October 10, 2026.

We found no visible signs. As of 2026-10-07, we searched the bot logs of intrix.kr and K-DermZIP for the names of personal shopping agents launched in the US in September (Meta's Muse and Instinct), as well as Comet, ChatGPT agent, Google-Agent, and Amazon's Buy for Me. All returned 0 records. However, these agents connect like ordinary Chrome browsers, so the logs cannot distinguish them in the first place. The 0 records mean something closer to "we cannot identify them" than "they did not visit."

What makes personal shopping agents different?

In a recent article, AI search analytics company Profound discussed agents that find, compare, and even buy products on a user's behalf. Examples include Muse, which Meta launched in the US in September, and Instinct, which users access through messaging. Both open real browsers, read pages like people, and use residential internet connections, making them look like ordinary visitors.

Profound said it had found one sign of Muse activity. In some cases, a request named meta-webindexer arrived from a non-Meta IP address just before Muse visited, and those requests increased after launch. Intrix checked whether the same signs appeared on Korean-language sites.

What did we look for in the logs?

We searched both sites' bot logs for all visits containing agent names. The intrix.kr records started on 2026-05-28, while K-DermZIP's records covered May through October 7 using monthly aggregates. The strings Muse, Instinct, Comet, Mariner, Google-Agent, and BuyForMe never appeared.

There were meta-webindexer requests: 176 on K-DermZIP and 150 on intrix.kr, all from Meta's own network. There were 0 "requests from non-Meta IP addresses," the sign Profound associated with Muse. Applying the same periods Profound used, K-DermZIP had 56 requests in the 3 weeks before Muse's launch (08-19–09-07) and 55 in the 3 weeks after launch (09-08–09-27). There was no difference before and after launch.

Did AI visits triggered by user requests increase?

Among visits triggered by user requests that identified themselves, only ChatGPT-User appeared consistently. These visits fetch pages when a user asks ChatGPT a question.

ChatGPT-User by monthMayJuneJulyAugustSeptember
K-DermZIP1,8742,1471,8241,8881,629
intrix.kr-40126536521

K-DermZIP's volume was flat over the five months and fell slightly in September. Of 5,305 K-DermZIP records, 96% were verified as actual OpenAI visits. Visits to intrix.kr increased over the summer and stayed at a similar level from August onward. These figures offered little evidence that agent shopping was spreading.

Why are these agents difficult to identify in logs?

Because these agents do not identify themselves by name. Muse and Comet connect with the same UA as ordinary Chrome. ChatGPT agent identifies itself through a signature header (Web Bot Auth), but the records from these two sites did not retain that header. Visits that do not identify themselves blend in with ordinary visitors.

The next measurement task is therefore to record signature headers. Those records need to exist before we can see what agents read on our sites and where they stop. To check which AI systems actually visit your site, you can start with a free website audit.

How far can these results be applied?

The analysis covers two Korean-language sites. Muse is available only in the US, so it would be natural for it not to visit Korean sites. These results cannot be applied directly to online stores in other countries. Visits that connect like ordinary browsers could not be counted in the first place.

The intrix.kr figures before August predate verification of visit origins. For K-DermZIP, older months were reviewed only as monthly aggregates. This analysis does not guarantee any particular visibility or traffic.

Frequently asked questions (FAQ)

Q. Should our online store prepare for shopping agents now?
The logs of two Korean sites showed no signs yet. However, because the logs cannot distinguish these agents, we cannot conclude that they are absent. Recording signature headers can help identify when they arrive.

Q. Does the lack of growth in ChatGPT-User mean traffic from ChatGPT declined?
There is no basis for that conclusion. ChatGPT-User visits fetch pages to generate answers. They differ from visits by people clicking links. On K-DermZIP, volume was flat from May through September, at around 1,600–2,100 visits per month.

Q. What is meta-webindexer?
It is the name of a crawler operated by Meta. On both sites, its requests arrived in batches every few weeks from Meta's own network. The sign Profound associated with Muse was requests using the same name from non-Meta IP addresses. Our logs contained no such visits.

← Back to insights