A useful content-quality audit starts with the page's purpose, evidence and actual search problem. Do not diagnose a Google penalty from an AI-detector percentage, a low word count or a site's total number of pages. Separate factual defects, Search policy issues, technical access and ordinary performance changes so that each repair addresses the right problem.
This guide covers AI-assisted articles, product pages, single-page sites, repetitive publishing, dated information, user contributions and structured data. It uses Google's public documentation to define the checks; it does not claim access to an internal quality score or a hidden list of penalty thresholds.
Identify the problem before assigning a quality label
| Observed problem | Evidence to inspect first | Appropriate next step |
|---|---|---|
| A page is not indexed | URL Inspection, response, robots directives, rendered content and selected canonical | Resolve demonstrated access or indexing issues, then assess the page's independent value |
| An indexed page receives few impressions | The relevant queries, market, topic and competing answers | Check whether it answers a real task and differs usefully from existing pages |
| Impressions remain but CTR falls | Comparable query/page groups, titles, positions and result-page features | Investigate the changed presentation and query mix |
| A group of pages loses traffic | Change history, demand, shared templates, technical state and search changes | Locate the common issue before applying a sitewide rewrite |
| A manual action is listed | The specific notification, examples and affected scope | Correct the stated issue and follow its review process |
Google's traffic-drop guide helps organize the investigation. The Manual Actions report is a separate source of evidence. No listed action does not mean every page is high quality, and a traffic decline does not establish that a manual action exists.
AI-assisted articles: verify the answer and its contribution
Google's AI-content explanation focuses on quality rather than how the content was produced. Its current generative-AI guidance calls for accuracy, quality and relevance. Neither provides a safe AI percentage or a minimum proportion of human rewriting.
Read the complete article and open the material behind its important claims. Check that the source exists, says what the article attributes to it and applies to the stated product, place, date or conditions. Repeated references to the same unsupported assertion do not become independent confirmation.
Hypothetical example: a draft attributes “this setting always prevents downtime” to a named engineer. The linked source is only a general introduction and contains neither the quotation nor that promise. Find an attributable original statement that supports the claim, narrow the explanation to what the documentation establishes, or remove the unsupported passage. Varying sentence length or lowering a detector score would not repair it.
Then ask what the reader gains. A careful comparison, a reproducible procedure, a clear interpretation of primary material or a real documented observation can add value. Reordering another article's paragraphs with new synonyms may not. The AI-rewriting review explains how to check meaning, attribution and contribution separately.
For a decline affecting an existing blog, use the AI-blog diagnostic. For new work, use the AI-assisted publishing workflow to prepare sources before drafting.
Product pages: answer the buyer's unresolved questions
A product page should help a buyer identify the item, understand its suitability and take the next appropriate step. The required detail depends on the product and transaction. A replacement component may need dimensions, compatibility and a drawing; a configurable machine may need operating conditions, options and a quotation process.
Review the page against the actual product record:
- Is the model or variant unambiguous, including differences that affect selection?
- Are dimensions, units, materials and operating limits supported by the correct revision?
- Are separately tested maxima presented as if they can all be achieved at once?
- Do images show the actual item or a clearly identified illustration?
- Are price, availability, delivery and included items accurate for the stated market and option?
- Can a buyer order, request the necessary quotation or obtain the relevant document?
Do not duplicate the same broad description across materially different models and leave the distinguishing information to a salesperson. Conversely, do not expand a straightforward product into a long generic essay merely to reach a word count.
Use structured data only for information the page supports. Google's product-snippet documentation and merchant-listing documentation describe different search experiences and requirements. A quote-only page should not invent a price, availability claim or customer rating to satisfy a test.
If customers can purchase directly, verify that the selected variant, displayed offer and applicable feed agree. If they must submit requirements first, make that process clear. Markup is a description of the real commercial information, not a replacement for it.
Single-page sites: distinguish a small website from an application architecture
A website with one page can serve a narrow purpose well. It becomes insufficient when several independent customer tasks need detail that the page does not provide. Assess the actual questions, rather than assuming that a particular page count makes a site low quality.
A single-page application is a different concept. It can have many routes and views. Inspect whether important content is available at usable URLs, whether Google can render it and whether navigation exposes the routes. Google's JavaScript SEO basics explain the crawling, rendering and indexing process.
Use URL Inspection to separate the indexed state from a live test, and inspect the rendered main content. An attractive application shell is not evidence that the answer was available to the crawler.
When deciding whether to split a page, ask whether the new URL will have a distinct purpose, sufficient supporting material and a useful navigation path. A fragment link to a section does not create an independently indexed page. Splitting every keyword variant into its own thin page does not solve the underlying content problem.
Repetitive publishing: inspect purpose and useful differences
Google's scaled-content abuse policy concerns many pages generated primarily to manipulate rankings rather than help users, typically containing little or no original value. It does not depend on whether AI, people or a combination produced them.
Compare related pages side by side. Do the titles describe different tasks? Do the answers actually differ? Are local or product details real? Does each page contain the information needed for its specific decision, or only the same text with names swapped?
Do not turn this into a universal daily publishing limit, similarity percentage or link quota. A team may publish several useful updates when the material and review capacity exist; a single unsupported article can still be poor work. Our Chinese guide to publication frequency and editorial capacity discusses that planning question.
When several URLs genuinely duplicate the same material, review their users, links and history before choosing a preferred destination. Google's canonicalization guidance addresses duplicate URLs. It does not make copied material original or authorize reuse of another publisher's work. Avoid unrelated homepage redirects as a substitute for an actual migration decision.
Dated information: update what changed, not merely the date
An old publication date does not by itself make an explanation wrong. A historical article can remain useful when its scope is explicit. A current buying guide, legal instruction or interface tutorial may need attention much sooner because the underlying facts changed.
Identify the time-sensitive claims: prices, available plans, version-specific steps, legal requirements, product availability and named reports. Open the current primary material, compare the relevant details and correct the affected explanation, examples and links.
Google's helpful-content questions warn against changing dates to make pages seem fresh when the content has not substantially changed. Its date guidance also calls for consistent visible and structured dates that describe the page's publication or update, rather than the date of an event mentioned in it.
Keep a record of the substantive change and its sources. For an archived rule, retain a clear historical scope when that serves readers. For a genuinely replaced resource, plan the appropriate destination and links. Do not promise a fixed ranking recovery merely because the date is newer.
User contributions: manage abuse without discarding legitimate information
Comments, forum discussions and reviews can answer questions that editorial pages miss. Their origin does not make them automatically low quality. The practical problems are spam, deception, unsafe instructions, irrelevant promotions and unsupported claims that the site presents as reliable.
Google's user-generated spam guidance recommends clear abuse policies, reporting paths, moderation and measures that help identify suspicious accounts or activity. Choose controls appropriate to the service and review their effects on legitimate contributors.
For a review, distinguish the contributor's experience from independently verified product facts. Do not fabricate reviews or convert a testimonial into a universal performance claim. If a discussion needs correction, preserve enough context for readers to understand the issue rather than silently changing its meaning.
Apply link attributes according to the actual relationship. Google's outbound-link guidance recommends ugc for user-generated links and sponsored for advertisements or paid placements; nofollow remains acceptable for the latter. These attributes do not moderate the contribution, verify its claims or prevent the destination from being discovered elsewhere.
Structured data: check truthfulness as well as syntax
Begin with the visible page and the supported search feature. Does the entity exist? Does the selected type describe it? Are the required facts present and consistent? Then test the rendered markup for that feature, including any version created by JavaScript.
Google's general structured-data guidelines require accurate, representative and visible information. Valid syntax does not guarantee a rich result. An invented rating, hidden answer or irrelevant type is not made legitimate by a green test result.
The same document distinguishes a structured-data manual action from general Web ranking: such an action removes eligibility for the affected rich-result appearance; it does not itself affect how the page ranks in Google Web Search. Diagnose the actual notification rather than calling every markup warning a sitewide penalty.
Keep page text, markup, product data and dates aligned through maintenance. If an offer expires or an article changes, update the relevant information consistently. Do not add an unsupported “AI schema” field as a substitute for useful content.
Reading experience: measure the actual obstacle
A correct answer still needs to be accessible. On mobile, check whether text, tables, controls and diagrams can be understood without unexpected horizontal scrolling. Open downloads and relevant links. Inspect overlays, advertisements and layout movement that obstruct the task.
Google's page-experience guidance describes several aspects rather than one score. It says there is no single page-experience signal; good Core Web Vitals do not guarantee a top ranking.
The current Core Web Vitals are LCP, INP and CLS. The recommended good targets are LCP at or below 2.5 seconds, INP at or below 200 milliseconds and CLS at or below 0.1, assessed at the 75th percentile with mobile and desktop considered separately. These are experience thresholds, not AI-content detection rules.
Keep field and lab evidence distinct. A Lighthouse run can locate a problem under its test conditions; it is not your full population of real visits. TBT is not a replacement measurement for field INP, and a missing field sample is not a passing result.
For WordPress, investigate the actual theme, plugins, assets, caching and hosting before buying or removing a component. GuangSuan's Chinese WordPress hosting and speed-optimization page describes its service scope; use the published terms and the diagnosed issue when deciding whether assistance is relevant. A hosting purchase is not proof that editorial quality or search performance has improved.
Turn the findings into repairs you can verify
For each issue, record the URL, observed defect, supporting material, proposed change and the check that will show whether the change was implemented. Name the claim or element involved. “Improve E-E-A-T” is too vague to execute; “replace the unsupported test claim with the documented model and conditions” is actionable.
Truthful author and company information can help readers understand responsibility. Use the E-E-A-T guide for practical examples, or the Chinese version when preparing materials with a Chinese-speaking team. A byline does not substitute for the article's evidence. Similarly, the NLP content-review guide can help examine topic coverage, but a tool's entity suggestions do not establish factual completeness.
Prioritize demonstrated errors, blocked important content and broken user actions. Then improve incomplete answers and repeated production defects. Preserve useful URLs, legitimate contributions and working materials rather than deleting content in bulk because a classifier or low-traffic filter selected it.
After publication, verify the database or CMS state, the served page, relevant metadata, links and assets. Save the change date for later comparisons. If there is a manual action, follow its specific review process; for ordinary fluctuations, use the evidence-led investigation instead.
Google's core-update guidance does not guarantee that improvements will produce a particular result or recover by a fixed date. A correction can be complete while indexing, rankings, AI citations and qualified inquiries remain unverified. Track each outcome separately so the next decision rests on what actually happened.