Practical use and limits
Use it for: Use the sequence as a release gate: inventory the public site, repair weak pages, require original evidence, verify both languages, run the audits and production build, then inspect representative deployed URLs before applying again.
Limits: This is one site's engineering record, not an AdSense approval formula. The checks cannot determine reader value, search demand, index state, account eligibility, policy interpretation, or a future review outcome.
Treat the rejection as a diagnosis, not a recipe
Google's notice did not identify one sentence or one missing page. It said the site was not yet ready because of low-value content and pointed to minimum-content, unique-content, user-experience, and webmaster guidance. That kind of response does not establish a universal article count, word count, or waiting period. The defensible interpretation is site-wide: a reviewer must be able to see why the pages exist, what they add beyond other results, who is responsible for them, and whether the experience is complete enough to use.
Record a baseline before changing the site
Before this case study was added, the repository contained 35 English article records and 35 matching Chinese records, producing 70 bilingual article URLs. The mix was 19 practical articles, 10 research reports, and 6 news posts. Money and Engineering each had 10 articles, AI had 9, Tools had 4, while Gaming and 3D Printing had only one each. This inventory exposed two important risks: broad topic coverage without equal depth, and dated news that requires maintenance after publication.
Improve weak pages before manufacturing new ones
The first pass did not create dozens of posts. It expanded three thin English articles and their Chinese counterparts, restored traceable sources, aligned section order and meaning across languages, and changed the Chinese content loader to fail when a translation is missing instead of silently showing English. Article pages now explain how the content was prepared, identify the person responsible for final publication, show the last review date, and provide a correction path. These changes make existing pages more accountable; they do not turn generic prose into original evidence by themselves.
Turn editorial expectations into executable checks
The repository now runs one publisher-readiness command before a review or release. It checks unique slugs, titles and excerpts; minimum substantive sections; English and Chinese completeness; article and section size; source coverage; HTTPS source URLs; publication and review dates; missing images; placeholder language; crawl configuration; and publisher trust pages. News and research require at least two traceable sources. The audit also confirms that localized pages fail closed, navigation exposes key policy pages, and utility pages do not compete for indexing. A successful check means the declared rules were met, not that the article is useful.
Keep the build and rendered pages in the evidence loop
Static content can pass a text audit and still fail as a website. The baseline quality run therefore paired the editorial audit with a production build, which generated the expected English and Chinese article routes. A local HTTP check then opened one English URL and its /zh/ counterpart and inspected status, visible language, canonical URL, and reciprocal hreflang links. This catches integration failures that word counts cannot see. Production deployment, domain routing, indexing, user behavior, and the advertising account remain separate checks after the repository build succeeds.
Put AI on the research side of the publication boundary
The recurring workflow has four stages: discover candidate questions from primary sources and repository evidence, score their likely value, draft only an approved candidate in an isolated worktree, and verify both languages before a human decides whether to publish. The queue explicitly disables automatic publication. A candidate needs at least 8 of 10 points, including at least 2 of 3 for original evidence. Summary-only rewrites, keyword variants, unverifiable dates, simulated first-hand experience, and individualized financial recommendations are hard failures regardless of the score.
Measure what automation cannot decide
A script can confirm that sources exist; it cannot prove that they support every nearby claim. It can compare section counts; it cannot guarantee that a Chinese translation preserves all uncertainty and limitations. It can enforce a minimum amount of text; it cannot determine whether a reader learned something unavailable elsewhere. Those judgments remain part of final review. The strongest candidates use artifacts the site can actually produce: a repository diff, a failing and passing test, repeated model runs, calculator boundary cases, a real incident log, or owner-supplied device observations.
What changed, and what remains unknown
The immediate result is a more complete existing library, a bilingual fail-closed rule, a stricter SEO audit, a separate candidate-scoring audit, and durable prompts for discovery and maintenance. The next 90-day target is deliberately small: six evidence-led new articles and meaningful updates to 12 to 15 existing pages, not daily autonomous publishing. None of this proves that Google will approve the next application. Search demand, actual reader usefulness, crawl and index state, account history, policy interpretation, and the timing of a new review remain outside this repository's control.
A reproducible sequence for another small site
Start by exporting a page inventory and classifying each URL by purpose, language, freshness, sources, and original evidence. Repair incomplete or misleading pages before adding more. Define a narrow editorial focus, a measurable candidate rubric, and hard reasons to reject a draft. Keep generated material outside published routes until a human verifies claims and translations. Run content checks and a production build, inspect representative rendered pages, deploy, and then use Search Console and reader feedback to decide what deserves maintenance. Reapply only after the public site—not merely the source code—reflects the completed work.
Frequently asked questions
How many articles does a site need before applying again?
Google does not provide a universal number in the cited guidance. A smaller coherent library with original evidence and complete user journeys can be stronger than a large collection of repetitive pages.
Can AI-generated articles be published on an AdSense site?
The relevant question is whether the content helps users and adds value. Generating many pages without added value can violate spam policies, so this workflow uses AI for bounded assistance and keeps evidence and publication under human review.
Does passing the repository audit guarantee approval?
No. The audit checks this site's declared editorial and technical rules. It cannot guarantee policy interpretation, indexing, traffic quality, account eligibility, or the outcome of a future review.