MixedEXP-03iones labUpdated on 20 September 2026
SEO & LLMs: ten weeks of assisted content on a test site
The questionHow far can we speed up technical content production with an LLM without losing indexing and visibility?
Verdict
Fully generated content is poorly indexed and does not get clicks. Content structured by AI and fed with human expertise performs as well as fully human content, in a third of the time.
Decision“AI skeleton + human expertise” protocol adopted for our own content. Generation without review ruled out.
- Status
- Mixed
- Duration
- 10 weeks
- Team
- 1 SEO consultant, 1 editor, 1 developer
- Stack
- Search Console
- Claude
- Screaming Frog
- Next.js
- Looker Studio
58%pages indexed, fully generatedafter 10 weeks
92%pages indexed, AI + expertsame timeframe
×3.4clicks AI + expert vs fully generatedat comparable impressions
−65%production timeAI + expert vs fully human
01
Context
Everyone has an opinion on AI-generated content and search ranking. Few people have measurements. Between clients who wanted to “produce 200 articles in a month” and those who refused any assistance for fear of a penalty, we needed clean data to advise honestly.
So we set up a test site on a new domain, on a technical topic we master (hosting and operating web applications), with no artificial inbound links, and published for ten weeks under four strictly separated protocols.
02
What we tested
48 articles, 12 per protocol, on queries of comparable difficulty chosen in advance. Same page template, same internal linking, same publishing pace. Only the writing method changes:
- A · Fully generated: one-line brief, article produced by the model, published without review.
- B · Generated then reviewed: same generation, review and factual corrections by the editor (20 minutes per article).
- C · AI skeleton + expertise: outline and structure produced by the model, technical sections written by an engineer with real examples, figures and screenshots, finishing by the editor.
- D · Fully human: complete writing by the engineer and the editor, without assistance.
Each week we tracked indexing, impressions, clicks and average position in Search Console, as well as the actual production time of each article.
03
Results
- 58%
- indexing at 10 weeks · fully generated
- 75%
- indexing · generated then reviewed
- 92%
- indexing · AI skeleton + expertise
- 92%
- indexing · fully human
Protocol A drops off from the fourth week. Five articles out of twelve are never indexed, and those that are plateau beyond position 30. At comparable impressions, they generate 3.4 times fewer clicks than protocol C: titles are generic, snippets interchangeable.
Protocols C and D are indistinguishable in Search Console. Same indexing rate, average positions within one point of each other, equivalent clicks. The difference lies elsewhere: a C article takes on average 2 hours 10 minutes of human work, a D article 6 hours 15 minutes.
- Review alone (B) corrects errors but not emptiness: the articles remain plausible and hollow, and Google can see it.
- The best-performing sections in C and D are those containing a concrete example, a command, a measured figure: exactly what a model cannot invent without getting it wrong.
- No manual penalty or “automated content” signal on any protocol. It is not the method that is sanctioned, it is the result.
- Ten weeks is still short; the gap between A and C was still widening at the end of the test.
04
What we conclude
Google does not need to detect AI. It only needs to detect the absence of value: no example, no experience, no point of view. A model is very good at structure, transitions and rephrasing; it is very bad at evidence. The question is therefore not “AI or not”, but “who provides the expertise, and where”.
For our clients, the recommendation has become simple: AI as a production accelerator, never as the author. An engineer who spends two hours on an article structured by a model is worth six hours of traditional writing, and infinitely more than ten minutes of generation.
05
What’s next
- Protocol C is our reference method for the iones blog and for the content we produce as part of SEO support.
- The test site stays online so that we can extend the measurement to six months and see whether the gap stabilises or keeps widening.
- Next test: the same protocol on a non-technical topic, to check that the conclusion holds outside our area of expertise.