Who holds the pen? AI is already in the chain, the question is elsewhere

The public debate about AI and content still revolves around a question that no longer exists: "should we use it?" In the field, that was settled a while ago. Agencies use it for first drafts, SEO consultants for outlines and tag variants, copywriters for research and rephrasing, merchants so their catalogue is not left mute. Claiming otherwise is a sales posture.

The useful question is a different one, and far more demanding: who holds the pen? In other words, who decides what gets published, under whose responsibility, and with what knowledge of the subject.

Accusing the trade of laziness is a bad case to make

There is an easy line of argument, and we refuse to use it: the one that paints agencies and SEO experts as people who produce a text in three minutes and bill three hundred euros for it. It is false, and above all it misses the point.

In a content engagement, typing has always been the smallest part of the work. The useful time goes into choosing the topic, identifying the search intent, knowing the catalogue and the market, structuring, fact-checking, internal linking, publishing, tracking positions, and making calls when results do not come. AI has reduced a fraction of that time. It has touched none of the rest.

What has really changed is that the marginal cost of mediocre content has fallen to almost zero. That is why the web is filling up with smooth text. And it is exactly why the value of someone who knows how to decide has gone up, not down.

Three things a model cannot decide for you

What is true of your product. A model describes what it is given. It does not know the reference has been out of stock for six weeks, that the supplier changed the composition, or that this particular model mostly sells as a gift. Verification is not a formality: an invented feature on a product page commits the merchant, not the tool.

What deserves to be written. A model will happily produce an article on any keyword you give it. It will not tell you the query has no buying intent, that it is held by three domains with unreachable authority, or that it is already covered by one of your pages, which would then compete with itself. That call is craft — and it is where SEO is won or lost.

What must be refused. Content can be correct and still be bad for the brand: a promise you cannot keep, a tone that does not sound like the house, a comparison that runs a competitor down. Someone has to be able to say no. A tool never says no.

The real subject: sustained production versus the one-off question

Much of the confusion comes from mixing two very different uses.

The assistant you query. You ask a question, you get a text, you take it away. Excellent for getting unstuck, rephrasing, generating variants. It is also what a native assistant in a back office does: available, immediate, with no memory of previous work and no tracking of what has been published. PrestaShop now offers its own, and that is a good thing.

Sustained production. This is no longer a conversation, it is a chain: you start from the real catalogue, you choose topics from what measurement tells you, you frame with templates, you review against a checklist, you plan, you publish, you watch what moves, you start again. What counts here is not the quality of one answer but the regularity and traceability of a hundred answers. The distinction is developed here.

The two coexist perfectly well. They simply do not solve the same problem, and confusing them leads to the classic disappointment: "we tried AI, we published fifteen articles, nothing happened".

Why smooth content does not rank

Google does not penalise generated content; it demotes useless content. That nuance has practical consequences. Mass-produced text is recognisable by traits that are precisely the ones the algorithm judges poorly:

• it adds no information that is not already in the ten existing results;
• it takes no risk, therefore settles nothing, therefore is cited by nobody;
• it contains no verifiable example, no figure from its own data, no acknowledged limit;
• it is interchangeable — swap the brand name and nothing changes.

Conversely, what no competitor can copy comes down to three materials: your data (what you observe on your catalogue, your returns, your sales), your judgement calls (why you recommend this rather than that), your limits (what you do not sell, what does not suit). That is the subject of humanising generated content, and it is also what makes content eligible for citation by answer engines — the subject of GEO.

What it implies for an expert or an agency

Three concrete consequences, if you take the reasoning seriously:

1. Do not bill by production time. It is the surest way to lower your own price as you become more efficient. Bill the result.
2. Own your tooling. An explained chain — brief, anchored generation, human review, planned publishing, measurement — reassures more than maintained vagueness. And it protects you the day the client discovers the tool on their own.
3. Move your time upstream and downstream. Framing and measurement have become the rare part. Production no longer is.

In short

AI is already in the content production chain, at almost everyone's desk. What separates good content from bad has not changed: did someone decide the angle, check the facts, accept responsibility for what is published? That is the only question, and it is a question of craft.

Publium is built around that line: the tool produces and publishes, the person decides and signs. If that is how you work too, let's talk — or just ask your question.