Two and a half years have passed since the AI-content gold rush kicked off. Every agency and every in-house marketing team has experimented with at least one workflow that involves a large language model generating draft copy. The early signal — late 2023 through mid-2024 — was that the content ranked just fine and the productivity gains were real.
The signal in 2025 is different. Google’s helpful content updates through 2024 and the SpamBrain rollouts this spring have changed the landscape meaningfully. Here is what we are seeing in client traffic data and how the playbook has shifted.
What the data shows
We watch traffic on roughly 40 client sites across SEO retainers. The sites span e-commerce, B2B SaaS, agency portfolios, and content marketing. The data through the last 18 months shows three distinct patterns:
Pattern 1: pure AI content. Sites that published large volumes of AI-generated content (typically 50+ posts per month) with minimal human editing saw initial traffic growth through early 2024, followed by sharp drops between September 2024 and March 2025. The typical pattern was 30 to 70 percent organic traffic loss over two to four months. Some sites lost rankings on existing pages, not just the new AI content.
Pattern 2: AI-assisted content. Sites that used AI for first drafts, outlines, or research, with substantial human rewriting and editing on top, saw mostly stable rankings through the same period. Some saw modest growth. The signal here is harder to read because the workflow is harder to detect.
Pattern 3: human-written content with AI tooling. Sites that did not use AI for content generation at all but used AI tools for research, fact-checking, and minor copy edits performed roughly the same as they did in 2022, adjusted for the general SERP volatility.
What SpamBrain seems to be looking at
We do not know what is inside Google’s classifier. The behavior of the algorithm against client sites suggests it is looking at multiple signals, not just text patterns:
- Pages that read like AI output (certain syntactic patterns, average sentence length, specific transition word frequencies) get downgraded
- Sites with high volumes of templated, near-duplicate content structures get downgraded site-wide, not just per page
- Sites without genuine author profiles, contact information, or other E-E-A-T signals get downgraded
- Sites whose content does not match real user search intent — answering tangentially rather than directly — get downgraded
- Sites with thin content disguised as long content (lots of restating the same idea) get downgraded
The pattern is that SpamBrain is not specifically detecting AI. It is detecting low-effort content. AI-generated content is a high-volume source of low-effort content, but AI is not the underlying signal. Low-effort writing produced by a human can fail the same way.
The workflow that holds up
Across the client sites that have continued to grow organic traffic through 2024 and into 2025, the content workflow has a consistent shape:
- Topic selection is driven by either a clear customer question or a real expert observation. Not by keyword research alone.
- An outline is drafted by a human with subject matter knowledge.
- AI tools — if used — assist with research, finding sources, structuring data, or generating section drafts that get heavy human rewriting.
- The final copy is written or substantively rewritten by a person who actually knows the topic.
- The piece includes specifics — numbers, named cases, original observations — that an AI cannot have generated without them being in the prompt.
- The piece is reviewed by someone with publishing authority on the site, not just published from a draft queue.
The cost per piece in this workflow is meaningfully higher than the cost per piece in a full AI-generated workflow. The output is also lower volume. The math on per-piece traffic — and on long-term traffic stability — has flipped in favor of the high-cost workflow for almost every client we watch.
What the failed sites have in common
The sites in pattern 1 — the ones that lost traffic — share consistent traits:
- They published more than the editorial team could have realistically read
- The author bylines were generic or AI-generated
- The pieces had little to no original data, examples, or first-person observation
- The internal linking was algorithmic — auto-generated keyword anchors rather than editorial decisions
- The site structure was templated to a degree that suggested the content was the only artifact, with no real product or service behind it
The classifier is reading the site, not just the page. A site that has all five of these traits is going to struggle even on pages that were not AI-generated.
What we tell clients now
The recommendation has changed in three ways since 2023.
First, content volume is no longer a goal. We used to set monthly post quotas with clients. We no longer do. The quota is a quality threshold, and the velocity is whatever quality permits.
Second, AI is fine in the workflow but cannot be the workflow. Using AI to research, outline, summarize, fact-check, or generate first drafts that get heavily rewritten is fine. Publishing AI output with light editing is no longer safe.
Third, the author byline and the E-E-A-T signals matter more than they did. Generic agency bylines are weaker than named author bylines with real LinkedIn profiles, real publication histories, and topic-specific authority signals.
The forward look
Google has not announced whether its position on AI content will harden further. The public statements continue to be that AI is fine if the content is helpful and human-reviewed. The algorithm behavior continues to suggest that the boundary line is closer to ‘meaningful human authorship’ than the public statements suggest.
For 2025 and the foreseeable future, the working assumption on our client sites is that any content the team would not be proud to put a real person’s name on is content the algorithm will eventually penalize. That is the bar we are now using on every content engagement, and it has not failed us yet.