A team starts publishing with AI. Output rises, then organic traffic falls. The obvious explanation is that Google detected the AI writing.
That explanation can send the team toward the wrong fix. Rewriting every sentence by hand will not resolve an indexing problem. Running the articles through a humanizer will not give fifty interchangeable pages a reason to exist.
Google does not prohibit content simply because AI helped create it. Its guidance on AI-generated content allows appropriate use of automation. Using that permission well requires understanding what the policy covers, and what a decline in traffic can actually tell you.
Key takeaways
- AI involvement alone is not a Google spam violation. A human rewrite does not automatically make a page useful or compliant.
- Scaled content abuse concerns pages produced primarily to manipulate rankings without helping users. There is no universal safe publishing quota.
- A manual action, an algorithmic ranking change and a traffic decline are different findings. Investigate before assigning a cause.
- Research based on AI detectors describes classified samples. It cannot establish exactly how each page was written or why Google ranked it.
- Keep this decision practical: what does the page contribute, can its claims be supported, and does it deserve a separate URL?
What Google's policy allows, and where the boundary sits
Google's generative AI content guidance recognizes uses such as research and structuring original material. It also warns that generating many pages without adding value may violate its spam policies.
The relevant policy is scaled content abuse. It covers large amounts of content created primarily to manipulate search rankings rather than help users, regardless of the production method.
There is no published number of articles that makes a program compliant. Nor does an editor's approval, by itself, settle the question of whether the output adds value.
Consider two hypothetical programs. One turns a company's documented support cases into troubleshooting pages, each resolving a distinct problem. Another produces hundreds of variations of “best software for [industry],” changing the industry name while keeping the recommendations and reasoning almost identical.
Both could use AI. The second deserves closer scrutiny because the new URLs may contribute little beyond targeting more queries.
A useful question for the commissioning editor is: If we removed the target keyword from this brief, would there still be a clear reason to publish the page?
A ranking decline is not automatically a penalty
“Google penalized us” often collapses several possible explanations into one. That makes it harder to decide what to do next.
A manual action in Search Console means a human reviewer at Google has identified a policy issue. Start with the affected pages and the instructions in the report so you can address the specific violation.
If positions fall without a manual action, the cause remains open. Review the queries that lost ground and the pages now ranking above yours. Compare the timing with ranking updates and changes to your own content before deciding what needs fixing.
Pages disappearing from the index call for a different investigation. Check the Page indexing report and inspect affected URLs, paying attention to recent technical changes that could have altered their eligibility for indexing.
A drop in clicks while positions remain broadly stable needs another explanation. Examine impressions and click-through rate, then look at whether the query mix or search demand has changed. Fewer visits alone do not establish that Google lowered your rankings.
Google explains how to identify and address a confirmed manual action in its Manual Actions report documentation. A clean report rules out a reported manual action, not every possible algorithmic problem.
A core update can change which pages perform best without establishing that the pages moving down violated a policy. Google also uses site-wide signals alongside page-level systems. A page is not assessed entirely in isolation from the site that publishes it.
For a marketing team, the immediate task is to name the observed change precisely. “Our comparison pages lost positions for these queries” is a useful starting point. “Google dislikes our AI content” is a hypothesis.
What the ranking studies actually show
Two widely cited studies can appear to tell opposite stories. Their methods explain why neither should become a blanket publishing rule.
In July 2025, Ahrefs analyzed approximately 600,000 pages drawn from top-20 results for a keyword sample. Its detector classified 81.9% as a mixture of human and AI writing. The reported correlation between estimated AI content share and ranking position was 0.011, effectively zero in that dataset.
That is evidence against a simple relationship within the sampled ranking pages. It is not a controlled test of what happens when a publisher switches to AI. Pages outside the collected ranking sample were not represented in the same way, and detector estimates are not verified production histories.
Graphite's October 2025 study used a different detector and classification method. Across 31,493 keywords, it classified 86% of sampled Google-ranking articles as human-written. Articles classified as AI-generated tended to rank lower. The researchers explicitly said they had not evaluated heavily human-edited AI-assisted content.
The percentages are not directly interchangeable. The studies used different samples and different definitions of what counted as AI-generated.
For an editor, neither study answers the question that matters most: whether the particular page awaiting approval is worth publishing. Use the research to challenge sweeping claims about AI authorship. Use the actual draft and its evidence to make the publishing decision.
The problems a human rewrite will not solve
A draft can sound natural and still fail to give a reader anything useful. These are the problems to look for before spending time smoothing out the prose.
A page with no distinct purpose
Imagine commissioning separate articles for “AI content and SEO,” “is AI writing bad for SEO” and “does Google penalize AI content,” then giving each the same policy explanation.
Changing the introductions does not create three different reader needs. Depending on what is already published, one comprehensive page may serve the questions better. Other pages can refer to it briefly and link to the full explanation.
For this article, the purpose is Google policy and publishing risk. The detailed production process belongs in How to Create AI Content That Ranks and Gets Cited.
A summary presented as expertise
A comparison article says one tool is “best for enterprise teams,” but never explains which enterprise requirement it meets. A tutorial implies the writer tested a workflow when nobody did.
Ask the author to supply the basis for the statement. If it came from product documentation, say so. If the team performed a test, explain its scope. If neither happened, revise the claim.
Google's people-first content guidance asks publishers to consider originality and evidence of expertise. E-E-A-T is not a single ranking factor or a score that a byline can satisfy. Author information helps readers understand accountability; it cannot substitute for the work.
A precise claim without precise evidence
Small errors can change a buyer's decision. A price might require annual billing. A feature might exist only on a higher plan. A customer result might cover a different period from the one the article describes.
Google specifically calls for manual review of AI-generated content for accuracy and trustworthiness. Its guidance includes metadata and structured data, so checking the body copy alone is insufficient.
For your team, keep the source beside the claim during review. An editor should be able to see what supports “available on every plan” without repeating the writer's entire research process.
A publishing target that outruns review
Suppose the team can properly review six articles a week but generates twenty. The extra fourteen do not become ready merely because the software finishes them.
A workable response is to adjust the queue: defer pages with unresolved evidence, combine briefs serving the same need, and give subject experts the passages that require their judgment. Treat the review capacity as part of production planning.
This is an operating recommendation, not a Google-approved publication threshold.
If traffic fell after you started using AI, investigate the affected pages
Start with the URLs and dates. Was the decline concentrated in newly generated articles, or did older pages fall too? Did a template change, migration or indexing rule go live at the same time?
Google's traffic-drop troubleshooting guide covers several possible causes, including technical problems and changes in search demand. Use Search Console to compare the affected pages and queries across equivalent periods. Separate changes in position from changes in impressions and click-through rate.
If the timing overlaps a core update, Google recommends waiting at least a full week after the rollout finishes before assessing the impact. Timing can guide the investigation; it does not prove the cause.
Then review a manageable sample of affected pages. Look for a shared problem: unsupported comparisons, near-duplicate coverage, outdated answers or a mismatch with what the query asks.
Keep a change log as you respond. Record the problem found and the edit made, then follow the relevant pages and queries. A rewrite followed by recovery is encouraging, but several changes may have happened together.
Avoid deleting a whole library because a detector labels it AI-generated. Decide whether each affected page should be improved, consolidated or removed based on its purpose and evidence.
A practical decision before publication
Before approving the page, ask the editor to complete this sentence:
“A reader should use this page because it helps them ______, using ______.”
The first blank names the task. The second names the contribution: a tested procedure, an explanation grounded in primary documentation, a useful calculation or evidence from your own work.
“Understand the topic, using information from other articles” may reveal that the brief needs more thought. It is not automatically worthless, but the editor should be able to explain what makes this treatment useful.
Publish when the page serves a distinct need, its material claims have support, and review is complete. If the purpose is sound but evidence or an important explanation is missing, return it for revision with that gap clearly identified.
When an existing page already serves the same need, consider consolidating the useful material there. A second URL should earn its place through a different purpose, rather than a slightly different keyword.
Hold the draft if its central claim cannot yet be supported or the team cannot explain why the page should exist. Waiting for evidence is a better editorial decision than filling the gap with confident prose.
These are editorial decisions. They do not guarantee a ranking outcome.
Where Pixis Visibility fits
Pixis Visibility brings content planning and execution together with search and AI visibility monitoring. It can help teams move from an identified gap to a brief and draft, then assess visibility after publishing.
The publishing decision still needs an accountable reviewer. A platform can surface an opportunity; the team has to establish whether the proposed page supplies a useful answer.
Use Google Search Console to investigate Google's indexing and search-performance data. Use AI visibility tracking to examine the responses in your monitored prompt set. Neither result alone is a certificate of content quality.
For the full writing and review process, see how Pixis creates AI-assisted content that ranks and gets cited.
