There is a number circulating that deserves more careful reading than it usually gets. In a study of 42,000 blog pages across 20,000 keywords, Semrush found that content classified as human-written occupied the number one position roughly 80% of the time, while content classified as purely AI-generated took that spot around 9% of the time. At position one, human-written pages were about eight times more likely to appear.
Read quickly, that sounds like a verdict on AI writing. Read carefully, it is something more interesting and considerably more useful.
The finding is narrower than the headline
Two details change the interpretation entirely.
First, the gap is concentrated almost entirely at the top. From roughly position five onward, the difference between human-written and AI-generated content narrows sharply. AI-generated content actually appeared more frequently in the lower half of page one. If your benchmark is "we rank on page one," AI-assisted content is competing perfectly well. The separation only becomes dramatic when you are fighting for the single most valuable position.
Second, the classification came from an AI detector. Detection tools are widely known to be inconsistent, and Semrush flagged this themselves. Some pages labelled AI-generated were probably written by people with an unusual style; some labelled human were probably drafted by a model and edited well. The finding is directionally credible, not forensically precise.
The same research included a survey result that seems to contradict the ranking data: 72% of SEO practitioners said AI content ranks at least as well as human-written content. It does not actually contradict it. Both things are true at different altitudes. Most teams measure themselves against page one, where AI content holds up. The ranking data measures position one, where it does not.
What the detector is actually measuring
A detector does not know your process. It cannot see whether you used a model for the outline, the first draft, or nothing at all. It reads the finished artefact and judges the text.
That means the study is not really a finding about tools. It is a finding about output characteristics. Pages that read as generic, structurally predictable, and free of specific first-hand detail underperform at the top. Pages carrying original observation, concrete examples, real numbers, and a point of view do better. Whether a model was involved in producing them is invisible to the reader and, ultimately, to the ranking system.
This maps closely onto what search quality guidelines have described for years through experience, expertise, authoritativeness, and trustworthiness. The signals that separate top results are the ones a model cannot generate from a prompt alone: what actually happened when you did the thing, what the numbers were in your case, what you would do differently.
The workflow question
The practical version of all this is not "should we use AI." Most teams already have. Semrush's survey found 87% of SEO teams describe their content as either fully human-created or heavily human-led, which suggests the industry has already settled on assistance rather than automation.
The more useful question is which parts of the process benefit from a model and which parts are precisely the parts you cannot outsource.
Where AI earns its place:
- Research synthesis and competitive scanning before you write
- Outline generation and structural alternatives
- Rewriting for clarity, length, or reading level
- Producing variants of something already good — subject lines, meta descriptions, social versions
- First-pass editing for consistency and repetition
Where it consistently costs you the top position:
- The core argument or point of view
- Specific numbers, results, and dates from your own work
- Anything requiring first-hand observation
- Examples drawn from actual client situations
- The judgement calls about what to leave out
That split is roughly what the data describes. The survey found 70% of teams cite speed as the main benefit of AI, but only 19% say it improves quality. Those two figures together tell you exactly what the tool is for: it compresses the time between having something to say and having it written down. It does not supply the thing worth saying.
A disciplined content brief becomes the control point in this workflow. If the brief specifies the argument, the required first-hand evidence, and the examples that must appear, a model can draft against it without hollowing it out. If the brief is just a keyword and a word count, you get exactly the generic output the study penalises.
Building the input layer
The uncomfortable implication is that better AI output depends on inputs most organisations do not collect systematically.
If your team ships work every week, someone knows what went wrong on the last three projects, what the client pushed back on, what the actual conversion lift was, and which approach quietly stopped working. That material is the raw substance of content that ranks at position one, and it typically lives in people's heads, in Slack, or in call recordings nobody revisits.
The teams producing genuinely distinctive content have usually built a small habit around capture: a running document of observations, a monthly fifteen-minute conversation with delivery staff, a standing note of client questions that recur. None of it is sophisticated. It just means that when the brief asks for a specific example, there is one available rather than a placeholder that gets filled with something plausible and unmemorable.
This is the part of connecting search work with content marketing that no tool solves for you, and it is increasingly the whole competitive advantage.
The AI search complication
There is a second reason to care about all this, and it points in the same direction.
Generative engines synthesise answers from sources, and the sources that get cited tend to be the ones offering specific, extractable, verifiable claims. A page of competent generalities gives a model nothing distinctive to quote. A page with a concrete figure, a clear definition, or a stated methodology gives it something to reach for.
So the qualities that win position one in traditional search are largely the same qualities that earn citations in AI-generated answers. That convergence is convenient. It means you are not running two content strategies. Applying practical standards for content AI engines can understand and cite reinforces the same specificity that the ranking data rewards.
What to change on Monday
If you are producing AI-assisted content at volume and wondering why nothing reaches the top, audit five recent pieces against one test: how many sentences could only have been written by someone who did this work?
If the answer is close to zero, the problem is not the tool. It is that the brief never asked for anything the tool could not supply.
Originally published on the Mustard Seed blog.
Top comments (0)