AI Content Without Human Review: How a Site Burns Through Its Authority

When I took over this site, I expected to find work to do. I didn’t expect to have to start by deleting pages.

This isn’t some tiny site getting a few hundred clicks a month. It’s a domain with strong authority in its niche, in a sector where Google raises the quality bar because a badly written page can do real harm to the people reading it. Some time ago they had started experimenting with RankPill, one of the many tools that generate and publish articles automatically.

The tool itself isn’t the culprit, and it does exactly what it promises on its site: keyword research, a three-thousand-word article, direct publishing, every day, with no manual intervention. The culprit is taking that promise literally. Keyword in, article online, next one please, for months, with nobody reading anything.

The examples below have been moved to a different industry so the client can’t be recognized. The flaws are identical; only the topic changes.

What I found when I opened the sitemap

Keywords that nobody actually says

The generator optimized for strings pulled from a tool, not for things a person would really say. One article hammered away at a phrase like “boiler help” as if it were a technical term, repeated in the opening, in the subheadings, inside sentences: “if you’re looking for boiler help, here you’ll find practical advice”, “boiler help can prevent serious breakdowns”.

Nobody searches like that. People type “boiler won’t turn on” or “who to call if my boiler is leaking”. Another piece was built around a keyword with the preposition baked in, with results that hurt to read: “understanding the nature of boiler”, “over 50% of Italian households regularly own of boiler”. There was even a slug that read /of-boiler, and another with an agreement error that sat there for months.

The same article written six times

The cannibalization was systematic, because the generator treated every synonym variant as new content. For a single search intent there were six different pages, with titles that only changed the word order: fix the boiler, how to fix the boiler, boiler repair, mend the boiler, sort out the boiler, boiler help. On another branch, five pages were competing for the same commercial query and three for yet another.

Six pages on the same intent don’t give you six chances to rank. They give you six pages that sap each other’s strength, none of them authoritative enough to win.

City pages where there was no city

Then there was the run of local doorway pages: the exact same page replicated for Naples, Turin, Bologna, Florence, Puglia, Switzerland. Places where the client has no office at all. Content identical to the letter, with only the place name swapped in.

Every article was the same article

Opened side by side, they followed a template recognizable at a glance: an opening with an alarming statistic and the current year, a promise of a complete guide, a list of what you’ll “discover”, a closing that urges you to take the first step. Same definitions, same risk factors, same advice, same conclusions. Not one page said anything the others didn’t already say.

This uniformity is the signature of mass-produced content, and it can be read by an algorithm and by a person alike.

The damage you can see and the damage you can’t

What you can see is the SEO side. Google stopped rewarding quantity a long time ago, and where the subject is health, money or important decisions, the threshold is higher still. A mass of formulaic, redundant content doesn’t lift you up: it drags you down, including on the good pages you already had.

But the worst damage is the other kind, and it’s the same mechanism I wrote about when discussing when it makes sense to use Lovable: the bill doesn’t arrive right away, and it doesn’t arrive as an error. No page breaks. The site simply starts to be worth less, and the traffic you never got doesn’t show up on any chart.

Then there’s the part no tool measures. People who land on those pages are there because they have a real problem and are deciding whether they can trust you. They read three sentences in which the same nonsensical phrase comes back eight times, and they’ve already made up their minds. Not in the way the site was hoping.

Commodity content isn’t neutral. It isn’t “better than nothing while we wait to write something good”. It costs reputation, and in some industries reputation is the only thing you sell.

What I did

I opened the sitemap and sorted every single page under four labels: keep, fix, merge, delete.

Delete the doorway pages and the unsalvageable content. Merge the duplicates into a single page, with 301 redirects from the removed slugs, because some had accumulated traffic and links anyway and that value shouldn’t be thrown away. Fix the pages with a good angle but poor writing. Keep, and they were the minority, those with genuinely original content, almost always written before the generator went into operation.

On top of all this I put a structure back in place: a pillar page for each core topic, and around it the deeper pieces, each with an intent of its own that nobody else covers.

The amount I deleted and consolidated is embarrassing, and the work isn’t finished yet. The upside is that we stepped in early: the site went through the three Google updates of 2026 without losing rankings, and that’s the most defensible result of the whole year.

What AI can’t do in your place

The lesson I take home isn’t “AI shouldn’t be used for writing”. I use it every day and it saves me hours.

The lesson is that AI can assist writing, not replace ideation.

Ideation is the part that decides whether a piece of content works: which page answers which search, what angle you have that nobody else has, what you know that others don’t, where the reader lands once they’ve finished reading. It’s exactly the reasoning I describe in when a blog is worth it for SEO, and it’s not work you delegate to a machine, because it requires two things a machine doesn’t have: knowledge of the market and the direct experience of the person who actually delivers the service.

If you skip that step, what you get is commodity material: the average of what was already online, regurgitated in different words. It adds nothing for the reader, nothing for Google, and it takes something away from you.

How I use it

The line, the way I find it useful, runs between the “what” and the “how”.

The what stays with me and the client: the content map, the choice of queries, the angle of each page, the technical or professional expertise we put into it, the check that what’s written is true.

The how I’m happy to share with AI: writing a first draft from an outline I made myself, rewriting a paragraph that doesn’t flow, picking apart a structure, lightening a text that’s too technical, coming up with variants of a headline.

There’s only one operational difference: a person always reads everything before it’s published, and has the authority to say no. The moment that step is skipped, you’re no longer using a tool. You’re letting something write in your name, and you sign without reading.

If you have a generator running

If you’re producing content automatically and you’re not sure what’s ending up online, there are three things you can check today in half an hour: how many pages you have that answer the exact same intent, whether there are slugs no person would ever have written, and whether reading three articles in a row you find the same structure and the same information.

If the answer makes you uncomfortable, the good news is that this kind of damage can be recovered from, as long as you catch it early. The bad news is that you recover by deleting, and watching two hundred pages vanish from your own site is never a pleasant conversation.

If you have a site in this situation, tell me about it.