Where the volume problem is

Most business websites are not short of ideas. They are short of hours.

A product catalogue with four thousand items needs four thousand descriptions. A services company knows it should be publishing regularly and does not. A site launch stalls because nobody has written the About page.

AI solves the volume problem. It does not solve the judgement problem, and website content is public — which raises the cost of getting it wrong.

What it does well

Four things, in order of how safely you can adopt them.

Product descriptions

The clearest win. From product attributes, at scale, in a consistent tone.

Give it the real attributes from your product records rather than letting it fill gaps. Most invented specifications come from missing information.

Check every specification. Dimensions, materials, compatibility, certifications. It will produce plausible ones.

Blog posts and articles

A first draft to react to, which for most people is faster than starting from nothing.

The catch: generated articles are competent and conventional. If everyone in your industry publishes the same way, everyone reads the same.

What makes an article yours is the specific example, the real customer situation, the opinion. That is what you add.

Page copy

Service pages, About pages, FAQ answers. Useful for a first pass.

Read carefully. These are the pages people judge you on.

Meta descriptions and titles

Genuinely tedious, needed on every page, and low risk.

A good use. Nobody enjoys writing four hundred meta descriptions and they matter for search.

The search engine question

A real concern, and the answer is more nuanced than either extreme.

Search engines do not penalise AI-generated content for being AI-generated. They penalise content that is unhelpful, thin or duplicated.

What that means practically:

Original and useful is fine, regardless of how it was produced.

Generic filler is not, regardless of how it was produced. And AI makes generic filler very easy to produce in bulk.

Duplicated content is the real risk. If your generated descriptions closely resemble everyone else’s generated descriptions of the same products, none of them rank.

The test: does this page contain something a reader could not get from ten other pages? If not, publishing it does not help you.

What makes generated content good enough

Four things, and the first two matter most.

Give it your facts

Real product attributes, real specifications, real details from your records. The gaps you leave are exactly where it invents things.

Give it examples

Three descriptions you like produces far better output than a paragraph describing what you want. This is the single highest-return technique available.

Add what only you have

A real customer situation. A specific number from your own experience. An opinion. This is what makes the page worth existing.

Review for truth, not style

The grammar will be fine. The question is whether the claims are correct, and that needs somebody who knows the products.

The mistakes that cost most

Publishing at volume without review. The pattern: four thousand descriptions generated, spot-checked at first, then trusted. One invented certification, across the whole catalogue.

Filling a content schedule. Twenty thin articles because somebody decided on weekly publishing. They do not rank and they dilute the pages that would have.

Copying competitor structure. Asking it to write about a topic produces something resembling everything already published on it.

Losing your voice. If your differentiation is how you talk to customers, generated text flattens it. Notice if your site starts sounding like everyone else’s.

No fact list. Decide what it must never claim — certifications, guarantees, compatibility — and check against that list.

The review habit that decays

Predictable, and worth planning against.

Week one, everything is read carefully. Month three, output has been consistently good and reading becomes skimming. Then one error publishes at scale.

Two safeguards:

Spot-check a fixed percentage, deliberately, permanently.

Never remove review from anything published in volume. The volume that makes generation worthwhile is the same volume that makes one error expensive.

A workable process

Six steps.

1. Start with the facts. Pull real attributes from your product records. Do not let it guess.

2. Give it three examples of what good looks like for you.

3. Generate a batch of ten, not four hundred.

4. Review those ten properly. Learn where it gets things wrong — usually specifications and claims.

5. Adjust the prompt based on what you found.

6. Then scale, with a permanent spot-check.

Step 3 is the one people skip. Generating everything at once means you discover the systematic error after it is everywhere.

Where Odoo helps

If your site runs on the same database as your products, generated descriptions can use the actual attributes — dimensions, materials, variants, real availability.

Less is invented, because less is missing. That reduces the error rate meaningfully.

It does not remove the need to read the output. It does mean there is less wrong to find.

What not to generate

Legal pages. Terms, privacy policies, disclaimers. These need a professional.

Anything with contractual weight. Warranty terms, service commitments, guarantees.

Case studies with invented details. If you are describing a real customer, use real facts and get their permission.

Anything you cannot verify. If nobody in your company can confirm a claim is true, do not publish it.

The short version

AI solves the volume problem in website content, and it does that well.

It does not decide what is worth publishing, and it will produce confident specifications it has no source for.

Give it your real data. Give it examples. Add what only you have. Review for truth.

And generate ten before you generate four hundred — that one habit prevents most of the expensive mistakes.