Trivipedia: 560 Articles Mass-Produced With Claude Code in 3 Months for ¥12,948 a Month. Reading an AI-Automated Media Site by Its Own Reports

The trivia site Trivipedia earned ¥12,948 in August 2026: ¥11,255 in confirmed AdSense revenue plus ¥1,693 from a sale on note. Among the 245 cases on this site that disclose monthly revenue, that ranks ninth from the bottom, roughly 1% of the median for the 16 Japanese blogs (¥1.2 million). On the numbers alone it looks like a thin case to cover. We are covering it because this is the most meticulous Japanese record we have found of someone testing, on their own site, the claim that mass-producing articles with AI stacks up ad revenue. A Claude Code automation pipeline published 90 articles in June, 174 in July and 286 in August, and pageviews roughly doubled each month. Ad revenue still stayed in the ¥10,000s, and search click-through rate fell from 4.22% to 2.80%. The operator hides none of it, down to the sitemap that stopped updating for two weeks.

Three months of actual numbers

The cumulative counts are the operator’s own (291 at the end of July plus 286 in August exceeds 560, but we keep the reported figure). August also included one sale of the automation kit sold on note, ¥1,980 less a ¥287 fee for ¥1,693. Affiliate revenue showed ¥15,000 in unconfirmed hosting commissions (two conversions) in June and ¥0 in July and August.

The site itself opened in December 2017 as a miscellaneous blog, published 27 posts in 2018 and 2019, and then sat idle for seven years. In June 2026 it was renamed, and the monthly reports begin at the point the operator describes as “exactly one month since starting full-scale operation with AI automation.” The operator is Yuki Mitsuyama, who writes that alongside running companies he handles planning, building the writing system, operations and article review for this site alone.

¥0.2 per pageview, ¥20 per article per month

The June report states an RPM (ad revenue per 1,000 pageviews) of ¥164.81. Dividing August’s ¥11,255 by 54,786 pageviews gives about ¥0.21 per pageview, or an RPM around ¥205, so the rate barely moved in three months. Pageviews grew. The price of a pageview did not.

A second division is the core of this case. August’s ¥11,255 in AdSense divided by 560 cumulative articles means each article generates about ¥20 of ad revenue a month. 560 articles for ¥10,000 a month. What mass production bought was pageviews, not price. Under the same ad model, a site with 3.5 million monthly pageviews earning ¥200,000 to ¥300,000 hit a ceiling of ¥0.5 per pageview. Trivipedia sits at less than half that rate, at one sixtieth of the scale.

Dependence on one article thinned, but rankings slid

In June, 61% of pageviews came from a single article, a list of chuunibyou words, move names and lines (6,728 pageviews). In August the same article still led with 9,323 pageviews, but its share had fallen to 17%, followed by a list of real-life unusual baby names at 2,331 and a list of aliases and epithets at 2,256. Mass production has at least worked in the direction of breaking the one-article dependency.

Search quality went the other way. Impressions grew 4.7x in three months while CTR fell from 4.22% to 2.80% and average position slipped from 6.06 to 7.3. The operator’s own diagnosis is that “new articles are being shown in large numbers at still-low positions, dragging the average down,” which names precisely the phase every volume-driven site passes through. Impressions rise, but the pages are not yet where they get clicked. Our own read, added to that, is that Trivipedia at this stage is a process for building inventory on search result pages rather than a revenue business, and revenue is a by-product.

The pipeline, and the automation that was spinning in place

The automation runs “topic generation, competitor research, writing, SEO check, WordPress draft posting” through Claude Code and produces three to five articles a day on active days. Thumbnails come from GPT Image 2, and posting to X and Threads is automated. In August the categories were reorganized from 5 to 11, quizzes and more than 70 in-browser tools (calculators and converters) were embedded in articles, and seasonal event articles gained automatic publishing along with auto-generated related links. The operator writes that AI is used for research and drafting, with checks against primary sources and public institutions done by a person.

The failures are in the same report. The sitemap had gone about two weeks without updating and roughly 170 articles were not registered. The automation was also found to have been “spinning through the same work endlessly.” These are the kind of accidents that do not happen when a person writes by hand, and the kind that happen only in a month that ships 286 articles. Social traffic rose from 50 sessions in July to 1,086 in August, but follower counts are not given.

The plan for September is to slow new-article production, rewrite one by one the articles like the baby-names list that show often but rank low, and strengthen monitoring and automatic recovery. Saying “quality over quantity” is easy. The fact that the operator reached that judgment in month three of mass production is the part that stays on the record.

The numbers that are not there

Costs are unknown. The July report notes that server and API costs are covered by revenue, but the August report gives no cost breakdown, so profit cannot be computed. Whether the API bill for generating 286 articles a month with LLMs, generating images, and auto-posting to X and Threads fits inside revenue in the ¥10,000s cannot be judged from outside. The operator’s hours are not given either. The note kit sells “automation know-how,” and a media site’s revenue sitting next to know-how sales in the same report is a classic shape of blog income reporting.

How much of this carries over

Japanese cases with real revenue figures for AI-made content include an AI explainer YouTube channel with 36,500 subscribers whose total revenue over two years was in the high hundreds of thousands of yen, an hourly rate of ¥416, and a record of 9,345 AI-generated images placed on Adobe Stock earning $589 a month. All three produced several to dozens of times the volume of a human, and none reached the revenue of the human-made cases.

• Mass-producing articles makes sense only when the goal is breaking a one-article dependency rather than raising per-article rates, and costs are planned on the assumption the rate will not rise
• The phase where impressions grow but revenue does not is survivable only with a monthly practice of measuring CTR and position decline as the passage of new articles toward their eventual rankings
• An automation pipeline avoids collapse only when monitoring for silent failures like sitemap outages and unregistered articles is built with the same priority as the generation itself

The limits are clear. This is a three-month record in the ¥10,000s, and whether pageviews keep doubling or the rate ever rises is unknown. Without costs and hours, whether it is profitable as a side project cannot be determined. The operator is also a company executive whose livelihood does not depend on revenue at this scale. And the largest limit is what the case demonstrates. It does not show that AI mass production produces no revenue. It shows that AI mass production can buy pageviews, and that the price of a pageview is set by something else.