Letting AI run your owned media: where people still decide, from records of two articles on our journal

What AI did and what people decided

In Rootpublish's records, AI carried out much of the work of making the articles. Approving the firsthand information used in an article, and deciding whether to publish it, remained with the site operator.

In September 2026, two articles on rootpublish.com's own journal (Article 2 and Article 3) were created in WordPress on a PC. Claude Code carried out these steps: choosing the topic, writing the request, generating the draft, checking it in a call separate from generation, translating into English, Chinese and Korean and checking the translations, delivering the articles to the site, and checking after publication.

For Article 2, the operator approved the firsthand information and published the article without editing the body. For Article 3, on the operator's instructions, Claude pressed the buttons in the admin screen to approve the firsthand information and to publish. The AI pressed the buttons, but the operator gave the instruction on whether to press them.

This operation was carried out while the operator could give instructions and approvals on the same PC. It was not an unattended, scheduled operation. In Rootpublish itself, what AI generates is saved as awaiting approval, and a person with publishing rights checks the body and sources and decides whether to publish.

Time taken to generate and check

For Article 3, generating a seven-paragraph draft from a request containing 31 approved pieces of firsthand information took about one minute. Checking that draft against the firsthand information in a separate call took about 15 seconds.

Both used Claude Code on a PC with an existing Claude subscription. The cost per article was not measured, and neither was the time the operator took to read and approve the draft. These figures cannot be used to estimate how much people's working time is reduced.

An overstatement found by the check, and how it was fixed

When a call separate from generation (Claude Opus) checked the draft of Article 3, it found one overstatement.

Under Rootpublish's specifications, only the search improvement feature observes the same page for seven days after publication and judges on measured results. The draft, however, described it as if it were a general post-publication policy. Before the draft was imported into WordPress, the sentence was rewritten to state the scope and conditions.

The check of Article 2 produced no findings. For three trial articles made before that, the check found two issues: a sentence that defined the scope of automatic updates without any basis, and a sentence that wrongly gave the time of an inspection as "at publication". Both were fixed in WordPress before publication.

The overstatement found in Article 3 was not a numerical error but wording that widened a feature's scope and conditions. How to fix a finding has to be decided by reading it against the original firsthand information.

Limits seen in checking translations

To check the translations, a Claude Opus call separate from the translation compared the original with each translation and reported omissions, additions, changes in meaning, terminology and unnatural phrasing.

For the English, Chinese and Korean translations of Article 3, 8 of 10 findings were adopted. The 2 not adopted mistook a mark added by the process that extracts text for comparison for a translation error. Across the translation checks as a whole, because text was compared with the tags removed, false findings that list numbers and link URLs were missing came up repeatedly.

Claude decided whether to adopt each finding by comparing it with the original; no person or native speaker has reviewed them yet. One finding suggested changing the Korean word for "publish" from 「공개」 to 「발행」. Because the wording has to stay consistent with articles already published, we decided to wait for a native speaker's judgment. A decision remained that the AI check alone could not settle.

Where publishing stalled

Publishing Article 2 had to be redone several times, for two reasons. First, the in-page publish confirmation opened below the button, off screen, and was missed. Second, the browser screen in which the confirmation had been opened was hidden, so the click to publish did not get through. As a fix, the confirmation now appears at the button's position, and focus moves to the confirm button.

There had been a problem before that as well. On screens that do not show the browser's confirmation dialog, pressing 「承認して公開」 ("Approve and publish") sometimes did nothing. The publish confirmation was therefore moved into the page.

In these records, the position of the publish confirmation and whether the screen was visible sometimes kept the publish action from getting through. Whether publishing actually completed needs to be checked on screen.

Rebuilding the admin screens while operating

On September 25, 2026, we updated the Rootpublish plugin three times in one day (0.1.14 to 0.1.16). We did this because, while we were running our own journal, the operator said it was hard to tell what to do and unclear where to start planning or reviewing.

An intermediate version added more explanation to each screen, which instead made the screens more cluttered. In the end, the admin screens were consolidated into four destinations — 「やること」 (To do), 「記事」 (Articles), 「会社情報」 (Company information) and 「設定」 (Settings) — and each screen states in one line when and for what it is used.

In addition, each article's stage (planning, request to AI, draft review, publication, measuring results) and its next action were brought together into one list. Draft firsthand information can now be reviewed on the screen and then approved in bulk. In our own operation, the concern that it was hard to tell what to do next was what prompted the rebuild.

Pages Google indexed after publication

On September 25, 2026, we submitted the sitemaps to Search Console. When we checked early on September 27, 29 of the 36 pages in the sitemaps had been indexed.

The 7 pages not yet indexed were four Korean articles, the English and Chinese article list pages, and one Japanese article. Search Console search data is reported with a delay of a few days, so impressions and clicks in search could not be checked yet.

Google explains that, for AI Overviews and AI Mode, meeting the requirements for appearing does not guarantee crawling, indexing or display. Rootpublish does not guarantee search rankings or citations by AI either. Being indexed does not mean that results have been achieved.

What these records cannot tell us

These records are a small sample: two articles on our own journal and their translations. Neither people's working time nor the cost per article was measured. Results such as search rankings or citations by AI are not known yet.

So we cannot say that handing operations to AI makes them run automatically, that working time goes down or that search performance improves. What these records show is only that there were steps AI could carry out, and that people's judgment and fixes to the screens were needed at the points of checking and publishing.

Steps to decide where people make decisions

From here on is a suggested procedure based on the records above and the suggestions in the Rootpublish editorial guide. It is not a method whose effect on search rankings has been measured.

First, write out the steps from planning to post-publication checks. Then separate the work AI will carry out from the work people will decide. In our records, approving firsthand information and deciding on publication were left to people.

Second, decide which firsthand information will ground the article, and use only approved information. The editorial guide recommends recording the point in time, conditions, source and what may be published, and not filling unconfirmed items with guesses. Rootpublish does not use unconfirmed or private information as grounds for an article, and for each paragraph of a draft it records the IDs and versions of the information referred to. However, the version check does not guarantee that the text matches the meaning of its grounds.

Third, put a check separate from generation in place, and decide who decides whether to adopt its findings. For findings that need consistency with published articles or a native speaker's judgment, such as translation terminology, having criteria for passing them to a person makes decisions easier.

Fourth, decide what to check before publication. The editorial guide suggests checking each claim against its grounds, figures and conditions, what may be published, and the reader's next step. The overstatement found in our records was also wording that widened a feature's scope and conditions.

Fifth, decide who presses the publish button and how to confirm that publishing has completed. Even when the operation is left to AI, as with Article 3, a person should give the instruction on whether to press it.

Finally, decide what to check after publication and when, and record when you checked. Google asks for a focus on accuracy, quality and relevance when generative AI is used, and also says that mass generation that adds no value for users may violate its spam policies. These steps assume that publishing articles whose grounds can be checked takes priority over increasing the number of articles.