kizasi

» see kizasi.jp as it lived, on the wayback machine

~ the eulogy ~

The slogan on the front page translated to something like: from blogs, know the topic, find the sign. Kizasi means sign or omen in Japanese, the early indication that something is about to happen. That was the product thesis. If you could read every Japanese blog post written today, you would know what the country was going to be talking about tomorrow.

It started in September 2005 as a business unit inside CAC, a Tokyo systems integrator, and worked well enough that CAC spun it out as its own company in January 2007 with 111 million yen of capital. The service collected roughly 250,000 blog entries a day, indexed them, and ran a time-series co-occurrence engine over the result. Not just which words were spiking, but which words were spiking next to which other words, which is the part that told you why. Most of the crawler and analysis stack was written in Python, which the team presented at a Python workshop in April 2006.

Yahoo Japan took a stake and shipped a consumer version of the output as Yahoo! Trend Word. That product closed on 26 November 2008, but the shareholding stayed in place afterward. Kizasi kept going on the business side, selling word-of-mouth research reports off the same index and publishing a developer API called kizAPI that returned related terms and rankings to anybody who wanted them.

~ the dossier ~

~ the engine ~

A quarter of a million posts a day sounds small next to a real web crawler, and it was. The bet was that it did not need to be big, it needed to be the right corpus. Japanese blogs in 2006 were where ordinary people wrote several paragraphs about a television show or a product on the same day they encountered it, in full sentences, timestamped, with an RSS feed attached. That is an unusually clean signal, and for about four years there was nowhere else to get it.

Co-occurrence over time is also a good idea that mostly got abandoned rather than improved on. Knowing that a word is trending is close to useless by itself. Knowing that a brand name has started appearing next to a different set of adjectives than it did last month is the thing a client will pay for, and it is what kizasi was built to produce.

~ the yahoo years ~

Trend Word lasting two years is the quiet data point in the story. Yahoo Japan had every distribution advantage there was and still could not turn buzz analytics into a consumer product anybody came back to. Nobody ever solved that. Twitter's trending topics is the only version of the idea that stuck, and it stuck by being a sidebar on something people were already using rather than a destination.

After Trend Word closed, kizasi drifted toward selling reports and then toward advertising technology, which is where the money in Japanese web infrastructure went in the 2010s. That drift is what the death notice is about. CAC Holdings wrote that the core ad-tech business had deteriorated with no prospect of recovery, and dissolved the subsidiary at an extraordinary shareholders meeting on 31 August 2019. The notice went into the official gazette on 2 September and liquidation was expected to finish by December.

~ what the blogs knew ~

The whole category died with the corpus rather than with any of the companies in it. Technorati was doing this in English, Bloglines and BlogPulse alongside it, all of them built on the same assumption: that people would keep writing several paragraphs a day on a personal domain with a feed. By 2012 they were writing one sentence on Twitter instead, with no feed, behind an API you had to license. There was nothing left to index.

I was on the reader side of that shift in Turkey and never once thought about who was measuring it. Blog aggregation reached me as Google Reader and as the blogroll in the sidebar, and it did not occur to me that somebody in Tokyo was running a Python job over a quarter of a million posts a night to figure out what any of it meant. The domain answers nothing today over either protocol, which is a later ending than the company got.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »