Why is AI-generated SEO advice usually wrong?

Reason one: the corpus is mostly noise

SEO is one of the most heavily written-about topics online, and the great majority of that writing is content marketing repeating the same twelve tips. A model trained on the open web has read that material thousands of times and genuinely novel technical guidance perhaps once. The statistical centre of "SEO advice" is therefore the listicle, and that is what you get by default.

This is why the advice feels familiar. You are receiving the average of everything ever written on the subject, which is a reasonable starting point and never a diagnosis.

Reason two: it ages badly and the model cannot tell

Practices that were correct at some point remain in the corpus permanently, and nothing marks them as expired.

• Keyword density targets. A number to hit was never real and has been actively counterproductive for over a decade. Still confidently quoted.
• The meta keywords tag. Ignored by every major engine for many years. Still suggested.
• Exact-match anchor text everywhere. Now a spam signal.
• Submit your site to search engines. Replaced by sitemaps and ordinary crawling.
• Word-count minimums. A correlation mistaken for a rule; padding a page to reach a number makes it worse.

Reason three: it cannot see your site

This is the largest factor and the easiest to fix. Asked what to improve, a model with no observation of your site suggests the things most sites need. If your titles are already good and your problem is that your entire content renders client-side and no crawler can read it, you will get advice about titles.

Worse, the advice will be correct in general, which makes it hard to reject. You will spend a weekend improving something that was not the constraint.

Reason four: no sense of proportion

Even given a list of real problems, a model with no effort or impact data will present them flatly. Twenty suggestions, all phrased as important. The one that would move the needle sits between two cosmetic ones. Prioritisation requires knowing both what a fix costs and what it recovers, and neither is inferable from the description of a problem.

The fix, in one step

Measure the site, then hand the measurement to the model.

1. Run a real audit and take the structured result — findings, evidence, effort, and what was not covered.
2. Paste it into the conversation, or connect a measurement tool so the assistant fetches it itself.
3. Ask the question you actually have: what should I do first, and why does it matter for this site?
4. Make one change.
5. Re-measure and confirm the specific finding cleared before moving on.

The difference in answer quality is not subtle. The model stops generating plausible advice and starts reasoning about your evidence, which is the task it is genuinely good at.