TL;DR
- Ask which expert inputs a page needs, not whether a machine typed the first draft.
- Two large ranking studies ran top results through AI detectors and got very different pictures. Neither shows that authorship causes rankings.
- Keep numbers, fit conditions, examples, advice and limits under named human ownership. Let AI help with structure, drafts and consistency checks.
- Publish when every claim has a source or an owner’s sign-off. Hold the page when a key fact has neither.
The real choice is about inputs, not authorship.
Many B2B teams now use AI somewhere in content work. So the useful question is no longer “AI or human?” It is “Which parts of this page must come from someone who answers for them?”
A buyer acts on specific things: a performance figure, a compatibility statement, a price basis, a warning about when a service is unsuitable. If one of these is wrong, the buyer pays for it later and your sales team inherits the problem. Those inputs need an owner, whatever tool produced the sentence around them.
Google’s guidance draws a similar line. It asks publishers to fact-check and review AI-generated content before publishing. It also warns that generating many pages without adding value may breach its policy on scaled content abuse. The test is value and accuracy, not the tool.
What do the AI content ranking studies measure?
Both studies measured how an AI detector labels pages that already rank in Google. A detector is a model that estimates how likely text is to be machine-written. Neither study tested what causes a ranking.
Ahrefs (July 2025) took the top 20 results for 100,000 keywords, about 600,000 pages, and used its own detector. It labelled 4.6% of pages pure AI, 13.5% pure human and 81.9% a mix. The correlation between a page’s AI share and its position was 0.011, effectively zero.
Semrush (April 2026) collected the top 10 results for 20,000 keywords in November 2025 and kept about 42,000 blog pages. Using the GPTZero detector, it found a position-1 page had an 80.5% probability of being labelled human-written and 10% of being labelled AI-generated.
Model estimates for one page set, not verified authorship. Ahrefs, July 2025.
Why do the two studies disagree so sharply?
Read side by side, the results clash. One detector called roughly one ranking page in seven purely human. The other called most top-ranking blog posts human-written. Different page sets may explain part of that gap.
The tools are another likely reason. The two detectors are different models with different thresholds. One used a broad mixed label for most pages. Detector labels are estimates, not a record of who wrote what.
The studies also share a limit. They looked only at pages that already rank, at one point in time. Nobody assigned AI or human drafting to similar pages and watched what happened. A human-written page may rank first because of its brand, links or expertise. This kind of data cannot separate those causes.
| Feature | Ahrefs (July 2025) | Semrush (April 2026) |
|---|---|---|
| Pages analysed | About 600,000 from the top 20 for 100,000 keywords | About 42,000 blog posts from the top 10 for 20,000 keywords |
| Classifier | Ahrefs’ own AI detector | GPTZero |
| Labelled human-written | 13.5% of all pages | 80.5% at position 1; 63.9% at position 10 |
| Labelled AI-generated | 4.6% of all pages | 10.0% at position 1; 17.5% at position 10 |
| Relationship reported | 0.011 correlation between AI share and position | Human-labelled share is highest at position 1 |
| What it cannot show | Whether AI use changes rankings | The same; there is no controlled comparison |
Figures come from the Ahrefs and Semrush studies linked above.
Use the table to set expectations, not to pick a winner. Neither study supports a rule such as “AI pages cannot rank” or “human pages always win.” Both fit a narrower point: detector labels are noisy, and rankings depend on far more than how the text was produced.
Which expert inputs must survive the AI draft?
Split the work into inputs a person must own and tasks AI can speed up. The split below is our editorial recommendation, not a finding from either study.
| Input | Why it needs a named owner | Typical owner | What AI can safely do |
|---|---|---|---|
| Numbers and specifications | Buyers compare and quote them. One wrong value travels into proposals. | Product or engineering lead | Flag every number in the draft for checking |
| Fit and exclusion conditions | A delivery expert needs to check where the offer fits and where it does not. | Delivery or applications lead | Turn approved notes into a clear checklist |
| Examples and cases | They need permission, accurate details and honest results. | Account owner, with client approval | Summarize the approved case record |
| Recommendations | A judgment call carries professional responsibility. | Subject-matter expert | Draft options for the expert to accept or reject |
| Limits and risks | Leaving them out misleads buyers and causes disputes. | Expert and service owner | Check the draft for unsupported promises |
| Dates and versions | Products, standards and prices change. | Page owner | List every claim that carries a date or version |
If a page needs none of these inputs, ask whether it should exist. A page that only restates what buyers can read elsewhere adds little, however it was written.
Worked example: a claim review that changed the page
Worked example — a fictional compressed-air services firm. An AI tool drafts a service page for leak detection audits from an outline and old brochure text. The draft reads well. Before editing the prose, the marketing lead pulls every factual claim into a review table for the applications engineer.
The engineer checks each claim against the approved audit procedure and recent reports. Two claims fail. One passes with a recorded source.
| Draft claim | Check | Decision | Published wording |
|---|---|---|---|
| “A leak audit typically cuts compressor energy use by 30%.” | No approved data supports a typical figure. Savings depend on each site. | Removed | “We estimate savings from measured leak flow and your run hours, and report them after the survey.” |
| “Our ultrasonic survey finds every leak.” | Incorrect. Detection depends on access and background noise. | Corrected | “The report lists any areas we could not survey, and why.” |
| “Each leak is tagged and logged with an estimated flow rate.” | Matches the current audit procedure and two recent reports. | Kept; source recorded | Unchanged. The procedure version is noted in the review file. |
The acceptance test was simple. Every remaining claim had a source or the engineer’s sign-off, and the engineer could defend the page on a sales call. The review does not prove the page will rank or bring leads. It removed two statements that could have cost the firm credibility with an informed buyer.
How to keep the speed without lowering the bar
Use AI where its output is easy to check: outlines, first drafts from approved notes, summaries and consistency passes. In the same Semrush report, a survey of 224 SEO professionals, 70% named faster, lower-cost production among AI’s biggest benefits, but only 19% picked better quality. Speed is the gain. Quality still needs a reviewer.
Then add three habits. Pull every factual claim into a review list before polishing prose. Give each claim an owner and a source, or remove it. Record who approved the page and when, so the next update starts from known facts.
Do not judge the program by page count. More pages only help when each one answers a real buyer question with approved evidence. Our guide to evidence-ready content briefs covers the fields to agree before anyone drafts.
AI Content and Expertise: Common Questions
Does Google penalize AI-generated content?
Not for the method alone. Google’s 2023 guidance says appropriate use of AI or automation is not against its guidelines. Its spam policies still apply to pages produced at scale without adding value, however they were made.
Should we run our drafts through an AI detector?
A detector will not tell you whether a page is accurate or useful. It gives a probability, and the two studies above show how differently two tools can label ranking pages. A claim review answers the question that matters: can each statement be defended?
Should we tell readers that AI helped write a page?
Google suggests explaining how content was created where that helps your audience. For most B2B pages, naming the accountable expert and the review date tells buyers more than a tool label.
Do these ranking studies apply to AI search answers?
No. Both measured Google’s organic results, not the sources AI assistants cite. Citation behaviour needs its own checks; see how AI search cites sources.
Who should review technical content if we have no in-house expert?
Ask the person who would answer the buyer’s follow-up question: an engineer, a delivery lead or a qualified partner. If nobody can sign off a claim, narrow it or leave it out until someone can.
Decide what your pages need before you scale them.
If your team can draft faster than your experts can review, the review step is the real constraint. SemShift’s GEO service connects buyer questions to pages and improves the evidence on them. Bring a few current pages and the questions your sales team hears most, and we can discuss where expert review should sit in that work.