Does AI content rank? Read the study's denominator first

What the published study measured

Semrush describes collecting 20,000 keywords in November 2025, taking their top ten Google results and retaining URLs containing a blog path, leaving 42,000 pages. It then used a detector to classify the text. The article was published on 1 April 2026; we read it on 21 October 2026. Study and methodology.

That is a compact description of the published design, not our replication. The source page also discusses a separate survey. Survey responses about a writing process and classifications of ranking pages are different evidence streams. Do not treat an opinion in one as a measured explanation for the other.

The first useful habit is to write the unit of observation before interpreting the chart. Here the reported analysis concerns selected ranking pages, not every page a business might publish. A claim about the success rate of future drafts requires a different denominator.

Keep the conditional probability facing the right way

Suppose a chart describes the mix of content labels among pages at a particular position. Its question is: given that a page is in this observed position, how often does it carry this label? That is a statement about the composition of an observed group.

The business question is often the reverse: given that I publish a page using a particular process, how likely is it to reach that position? Reversing the condition changes the question. The first quantity does not, by itself, answer the second.

In notation, P(label given position) is not generally equal to P(position given label). To relate them, you need additional information, including how common the label is in the relevant population. You also need a defensible definition of that population and comparable observation conditions.

This is not a technical objection for its own sake. If one writing process produces far more pages, it can contribute more ranking pages even when the success rate of each individual page is different. Counting winners without knowing the relevant starting population cannot resolve that possibility.

Selection changes what the chart can tell you

A study that starts with ranking pages has already selected on the outcome people care about. It can describe those selected pages well under its stated collection rules. It does not automatically describe the pages that failed to rank or were never eligible for the sampled queries.

A blog-path filter adds another boundary. Many useful articles do not live under that path, while some pages under it serve different purposes. The filter may be practical and reproducible, but its convenience does not make it a complete definition of all editorial content.

Keyword selection matters too. A set of questions can differ in competition, topic, intent, geography and freshness needs. Without matching those conditions, a difference in observed labels could reflect the types of pages and queries included rather than an isolated effect of the writing process.

These are questions to ask of the design, not allegations that the underlying work is worthless. A narrow descriptive study can be valuable. Trouble starts when a narrow finding becomes a universal causal headline or an individual forecast.

A detector label is not a production diary

A text detector classifies the finished text under its own method. It does not watch the author interview a specialist, draft an outline, revise a passage or verify a claim. A page's actual history may contain several human and automated stages that a single label cannot fully represent.

Classification error can affect comparisons. To understand how much, you would need validation evidence appropriate to the content types, languages and editing patterns being studied. This article does not have that dataset and therefore does not estimate an error rate or “correct” the chart.

Avoid treating a detector's label as a moral assessment of the page. A human-written page can be inaccurate, derivative or unhelpful. An assisted draft can contain valuable first-hand material after accountable review. Neither observation establishes a measured ranking advantage for the process.

For your own editorial operation, keep a production record directly. Name the source facts, expert contribution, reviewer and material revisions. That record is more useful for accountability than asking a detector to infer what your team already knows about its own work.

Look for alternative explanations before claiming cause

Ranking can be associated with many page and site properties. Established sites may have different editorial resources, link profiles and topic choices from new sites. Content age and update practices can also differ. A writing label may travel alongside those properties rather than explain the outcome independently.

A causal question needs a design capable of separating the effect of interest from those alternatives. That might involve a carefully controlled experiment or another justified analytical approach. It also requires a defined outcome, a comparable observation period and enough evidence to support the inference.

This article does not propose an informal race between two newly published pages as a reliable experiment. Different subjects, intent or competition can make that comparison misleading. Publishing duplicate versions merely to create a test can introduce additional problems instead of resolving the original question.

The practical conclusion is modest: read the chart as descriptive evidence within its design. Use it to ask better editorial questions. Do not use it as a multiplier for promised traffic, a detector score target or a guarantee that a draft will rank.

Turn the discussion into a stronger draft review

For fictional Peak Plumbing, imagine an owner-supplied article draft about preparing a repair enquiry. The review should ask whether the content answers the customer's question accurately and uses verified business facts. It should not begin by trying to make the prose “pass” a detector.

Use Claude or ChatGPT to organize that review with supplied sources. The following prompt is proposed and unexecuted:

Review this draft using only the business facts and sources supplied.
List each factual claim and the evidence supporting it.
Flag generic passages that add no useful answer.
Identify where first-hand owner knowledge is needed.
Do not invent examples, credentials, results or citations.
Return a revision plan for a human reviewer, not a ranking forecast.

Illustrative expected review record, not actual host output:

Claim: the business offers a stated service in a stated area
Source: owner confirmation still needed
Decision: leave the claim unresolved until confirmed
Useful addition: explain what details an enquiry should include
Ranking prediction: not assessed
Detector target: not used as an editorial acceptance condition

What should the editor do with that record? Obtain the missing confirmation, remove unsupported statements and add useful first-hand detail where appropriate. Then reread the actual revised text. A checklist completed by an assistant is not a substitute for inspecting the finished page.

Review originality as contribution, not unusual wording

Ask what the page contributes that the reader could not obtain from a generic summary. It might organize a difficult decision, explain a real process or provide a verified example. Unusual adjectives and paraphrased definitions do not create that contribution on their own.

Keep the source of practical knowledge visible. If the owner explains how their enquiry process works, identify that as supplied operational information. If an external source supports a technical point, link it and verify that it actually supports the sentence. Never invent a source to make a generated passage look researched.

Editing should also preserve limits. Removing a cautious sentence can make a page read more smoothly while making its claim false. A useful reviewer checks both clarity and truth, especially when the draft discusses safety, availability, qualifications or outcomes.

Finally, read for the person with the question. Can they understand the answer, recognize the conditions and identify the next appropriate step? That is an inspectable editorial standard. This article recommends it because it improves the usefulness of the page, without claiming the study measured its effect on rankings.

Measure your own work without rewriting the evidence

Keep dates for publication, substantial revision and measurement changes. If you later study performance, retain unsuccessful pages and missing observations in the relevant population. A portfolio made only of successful examples cannot estimate the success rate of your whole process.

Record the actual mix of drafting and review practices rather than assigning a binary label after the outcome is known. Define the outcome before looking for a flattering comparison. Separate ranking observations from visits, enquiries and verified business results.

When sharing a study, state the question it answered and the question it could not answer. That gives readers a useful conclusion without making the evidence carry more than it can support. For this Semrush article, the right response is careful reading and accountable editorial work, not a universal claim that one origin of text must win.