In his July 2025 account of GetInvoice, now Tailride, founder Miki Palet described publishing more than 22,000 AI-generated pages. The site reached roughly 130 organic clicks a day in February, he said. On February 28, that traffic fell to zero.
For anyone using AI to build a blog, that story deserves attention. Humanization matters because someone must turn generated material into a useful, credible answer. An AI detector can support that review, but changing a score cannot repair a weak page.
Palet described a pipeline that scraped source material, turned it into structured posts and published without manual review. Return-policy articles repeated instructions; later financial pages reused substantial content and often had little connection to the product. He reported no manual action and did not know the precise trigger. The drop also preceded the March 2025 core update, so it should not be presented as a confirmed consequence of that update.
The uncomfortable part is how familiar the publishing logic is: find a keyword, produce an answer, repeat. A rising traffic chart can make that process look validated while nobody is checking whether the next article deserves the reader's time.
The losses extend beyond one publishing experiment
SEO analyst Lily Ray's May 2026 analysis of AI-content platform customers gives the concern a wider context. Among more than 220 selected sites and subfolders, she reported that 39% had lost at least half their estimated peak organic traffic. She also identified repeated blog formats, including large sets of comparison articles and self-promotional lists without clear evidence of hands-on testing.
Ray selected domains from publicly available customer stories for AI-content platforms, sometimes narrowing the analysis to the relevant subfolder. Her measurements came from Ahrefs, checked against Sistrix visibility data, rather than the publishers' own analytics. This is a warning from a selected group, not a failure rate for every website using AI; Ray explicitly says the tools themselves are not established causes of the declines.
A later example appears in Glenn Gabe's August 2026 spam-update case studies. He reported that an unnamed site combining programmatic pages with content he identified as AI-generated lost rankings for more than 200,000 queries. That is a count of search queries, not lost visitors, and the case concerns a broader programmatic site rather than a verified blog-only loss. Still, it adds a recent example of scaled content suffering a sharp loss of visibility.
The shared concern is a publishing system that can multiply pages faster than anyone can establish their value. A blog may look substantial because it covers every keyword variation, yet offer almost the same answer throughout. Smooth prose can conceal missing evidence, weak comparisons and advice that stops just before the reader needs a specific answer.
Google's scaled-content abuse policy addresses pages produced primarily to manipulate rankings while adding little value, whatever method created them. Its examples include generating many low-value pages with AI and transforming scraped content through synonym changes. Google's guidance on generative AI content also recognizes useful roles for AI in research and structuring original material. The responsibility sits with what the publisher puts on the page.
Humanization has to improve what the article offers
Useful humanization begins with decisions a rewrite cannot make for you. Which question is worth answering? What do you know that helps someone answer it? Which claims can you support, and which should come out? Those decisions give an AI-assisted article substance before an editor works on its voice.
Consider a blog comparing software for small businesses. A paragraph calling every option powerful, intuitive and suitable for growing teams gives the buyer little help. Rephrasing those adjectives more casually leaves the decision just as difficult. A stronger comparison explains the task each product handles, the limitation that could rule it out, and where the supporting information came from.
When a writer has actually tested a product, the article can describe what they tried and what happened. When the comparison relies on documentation, it should say so and link to the relevant details. A customer interview can add another perspective, with permission and attribution. None of that requires inventing a personal story or pretending an AI-generated sentence came from hands-on experience.
This also changes how you edit a weak answer. Suppose a draft advises a reader to “choose the right tool for your needs.” The useful revision identifies the choice: whether the tool supports the reader's file format, fits the required workflow, or has a limitation that blocks the task. Verify those details first, then explain them plainly. The value comes from resolving the uncertainty that brought someone to the article.
Accuracy needs the same attention. Open the cited source and check that it supports the sentence, including its date and limits. Keep a reported traffic estimate distinct from a site's own analytics; keep an analyst's explanation distinct from a confirmed cause. A fluent paragraph with the wrong claim becomes more convincing after a cosmetic rewrite, which makes it more harmful.
Original wording deserves a separate check. As explained in AI Detection and Plagiarism Checks Answer Different Questions, text can avoid matching a source while still adding little of its own. Preserve attribution when you paraphrase, and add something the source does not already provide: a relevant comparison, a checked example, an explanation of a difficult step or a useful limitation.
A blog's internal links should carry that usefulness forward. A comparison can lead to a detailed setup guide; an explanation can lead to the tool that performs the task. Use anchors that tell readers what they will find, and keep the destination relevant. Linking a collection of interchangeable posts together does not make their answers more complete.
Give AI detectors and humanizers a defined job
For a publisher accepting drafts from several writers, an AI detector can add a consistent place to begin a conversation about the writing. Detecting AI returns an overall score and highlighted passages that can be read in context. Use those passages to ask about sources, drafting choices and the evidence behind a claim, while reviewing the rest of the article too.
The detector's methodology describes the report as a review signal and says it cannot establish who wrote a document. Human writing can be flagged, and generated writing can receive a low signal. Understanding false positives in AI detection helps prevent a score from becoming an accusation or an automatic rejection of a useful contribution.
Once the substance is sound, the AI humanizer can help revise stiff or repetitive wording. Its changed-wording highlights make the proposed edits easier to compare with the original. Give it a passage whose meaning you can judge, then read the result against the sources before using it. Keep a qualification such as “may” when the evidence does not justify “will,” and reject changes that lose a condition, attribution or exception.
That combination has a practical purpose: detection supplies another review signal, humanization assists with expression, and a person remains responsible for the answer. It can be useful whether the draft started with a writer, a model or both. The work is complete when the article communicates something accurate and useful, rather than when a number looks reassuring.
There is no supported detector-score target that guarantees Google rankings. Different systems can score the same passage differently, as the guide to why AI detectors give different results explains. Repeatedly rewriting to reach a preferred percentage may change the writing without fixing its missing evidence or poor fit for the reader.
If your blog has already lost traffic, look at the affected pages before commissioning a site-wide rewrite. Google's traffic-drop guidance can help distinguish ranking changes from indexing problems, seasonal demand and other causes. Then examine whether those pages contain useful material worth improving, overlap with a stronger answer elsewhere on your site, or exist mainly because a keyword was easy to generate content for. A better article is worth building, but recovery still has to be measured rather than promised.
The next AI-assisted post should contain something a reader can point to and use: a verified answer, a revealing comparison, a tested instruction or an honestly explained limitation. Make that contribution visible before you press Publish.