Different results within one batch of AI-generated articles do not, by themselves, establish an AI penalty. First distinguish undiscovered URLs, uncrawled pages, non-indexed pages, indexed pages with little visibility and pages that have lost established rankings. Each state calls for a different investigation.
Define the outcome before explaining it
Save the batch's URLs, publication dates, content versions, intended questions and observation times. A rank tracker's “outside the top 100” result is not an indexing report. One manual search is not a complete view of a page's performance across queries and audiences.
Search Console's URL Inspection tool also separates indexed information from a live test. A successful live test does not guarantee indexing, while the stored information may still describe an older version of the page.
| Observed state | What it establishes | Next investigation |
|---|---|---|
| URL unknown to Google | No recorded discovery of that URL | Public access, internal discovery paths and sitemap inclusion |
| Discovered, currently not indexed | The URL is known but not yet crawled | Serving capacity, crawl conditions and discovery, rather than an immediate full rewrite |
| Crawled, currently not indexed | A crawl occurred, but the page is not currently indexed | The fetched version, duplication, representative URL and usefulness |
| Another canonical was selected | The inspected URL is not the group's chosen representative | The current page, declared canonical and Google-selected page |
| Indexed with few impressions or clicks | Index eligibility and search visibility differ | Demand, intent, competing answers, time and measurement scope |
| Previously visible, now persistently lower | A change needs explaining | Deployment history, demand, updates, affected groups and actual notices |
Use the definitions in Google's Page indexing report. “Crawled, currently not indexed” does not identify a sentence pattern as AI-generated. A duplicate URL being represented by the correct canonical can be an intended outcome rather than a failure.
One generation process does not create identical opportunities
The same prompt can produce an answer to a genuine model-specific question and a generic article with little distinct purpose. One topic may have source diagrams and clear conditions; another may lack the information needed for a useful answer. Different query demand, competing results, discovery paths and observation periods also remain relevant.
SE Ranking's published experiment illustrates why outcomes need separating. It observed AI-assisted articles on its established blog and a separate group of 2,000 generated articles deployed across 20 new domains in November 2024. The new-site group used a common theme, received no further manual content adjustments and was tracked in the US. The report describes pages remaining indexed while their visibility declined.
That is not evidence of a universal three-month penalty for AI text. Site history, topic selection, editorial treatment and ongoing work differed between the groups. Their difference cannot be assigned entirely to one variable, and coverage of the same experiment by another publication is not an independent replication.
Look for shared technical faults, then inspect individual pages
Group the URLs by template, language, release period and content type. If one template consistently removes the main content, adds noindex or points canonical tags to the wrong page, fix that shared cause and verify the affected scope. When only a few pages differ, inspect their actual state rather than assuming the template explains everything.
- Access: inspect final responses, redirects, unintended login requirements, CDN blocking and failed resources.
- Rendering: compare the public page with available crawl and rendering information. Confirm that the answer is present, not just an application shell or error message.
- Representative URL: check canonical signals, sitemap entries and internal destinations for consistency.
- Discovery: follow relevant category and article links to the page. An article linking to others does not establish that it has an incoming route.
Google's linking guidance emphasizes crawlable links and useful context. When planning internal links, check the source passage, anchor and destination together. A fixed links-per-thousand-words quota cannot establish relevance.
Review the answer, not just its similarity score
Break the title's promise into the questions a reader needs answered. A comparison needs comparable conditions. Installation guidance needs prerequisites and usable steps. A rule may need a jurisdiction, version and applicability statement. An article with substantial background but no usable answer can still require major work.
- Facts: trace specifications, features, statistics and examples to actual sources. A rating for one manufacturer's model does not describe the entire product category.
- Added value: identify the explanation, comparison or material the reader gains beyond a paraphrase of another page.
- Page roles: determine whether neighboring articles answer distinct questions or merely substitute synonyms in their titles.
- Expression: check logic, omitted conditions and unsupported claims of personal testing.
The AI-rewrite editing guide provides a practical review process. Google's people-first questions do not prescribe a word count or percentage of manual rewriting. Its scaled-content policy is not a promise that publishing fewer than a particular daily quota is safe.
Revisit topic selection using customer questions, relevant Search Console queries and appropriately scoped keyword research. No impressions on a new page do not, on their own, show that demand is absent. Equally, a tool estimate does not prove that this particular answer deserves the traffic.
Repair real experience problems without inventing ranking thresholds
On mobile, test whether someone can read the comparison table, identify important image details, open the supporting material and complete the next step. For large images, compare formats, dimensions and visual quality. Converting to WebP does not guarantee a fixed size reduction, and a smaller file is not a full-page performance result.
GA4 engagement and bounce metrics describe sessions. A brief visit can complete a task; a long visit can reflect difficulty. Neither establishes satisfaction or a Google penalty on its own. There is no basis for diagnosing a page using a universal “47 seconds” or “68% bounce” cutoff.
For an established decline, build a change timeline
Compare equivalent periods and relevant page, query, country, device and search-type groups. Distinguish lower impressions, lower CTR and changed positions. Then examine deployments, migrations, seasonal demand, access failures and Google update dates.
Google's traffic-drop guidance treats these as different investigative paths. A matching date is a lead, not proof of causation. If there is an actual manual action, use the notice-specific remediation and review process.
For a core-update analysis, Google recommends confirming that rollout has finished and waiting at least a full week before making the before-and-after comparison. This is not an instruction to leave a broken site unrepaired, nor a one-week recovery guarantee.
Keep the repair and the observed result in separate fields
A useful per-URL record includes the current state, evidence date, hypothesis still to test, specific change, responsible person, release verification and subsequent search observation. “Added verified model-compatibility information,” “public page shows the revision” and “relevant clicks increased” describe three different things.
Consider a diagnostic exercise rather than measured site results: page A has an accidental noindex; page B duplicates an older answer; page C is indexed but lacks target-query impressions; page D has a sustained decline. The common batch label does not make “rewrite 20%” the right action for all four.
After a substantive repair, request recrawling where appropriate and monitor. Google says repeated requests for the same URL do not make crawling faster, and a request does not guarantee indexing. Preserve unresolved explanations instead of assigning every difference to supposed AI fingerprints.