AI Content Creation Without the Generic Sludge

AI content creation is the use of generative models to produce written, visual, or video content, and the honest summary of where it stands is this: the tools work, the average output does not, and the gap between those two facts is where the entire opportunity sits. Google does not penalise content for being AI-assisted. It penalises content for being worthless, which most unedited AI output is.

The evidence cuts both ways, which is why the debate keeps going in circles. Around 86.5% of pages ranking in the top 20 contain some AI-generated or AI-assisted content, and one analysis found 81.9% of top-20 content is a human and AI mix. A Semrush study found 57% of AI-generated content and 58% of human-generated content reached the top 10, which is as close to parity as a study can get. So AI content ranks.

And then the other half. A 16-month study of 4,200 articles found content published with no editorial enhancement ranked 23% lower on average than human-written pieces targeting the same keywords, and after the March 2026 core update the split widened sharply: human-written up 11%, AI-assisted up 4%, pure AI down 18%, with deindexation spiking in the three weeks afterwards (Digital Applied, 2026).

Read together, those findings say something precise. AI-assisted content is now the norm at the top of search. Unedited AI content is a liability. The variable is not the tool. It is what a human added.

What Google actually said

Google’s position has been consistent and is narrower than either camp claims. AI content is allowed. What violates policy is scaled content abuse, and the definition requires two things to be true at once: volume, and little value added. Volume alone is not the offence, which is why legitimate programmatic pages, product catalogues, and multi-location service pages can exist in the thousands without issue, provided each page serves a genuinely distinct need.

That distinction matters because most firms worry about the wrong half. They ask whether using AI is safe. The question that predicts the outcome is whether the page adds something that did not exist before it was published.

The four failure modes, and why they are structural

Generic AI content is not the result of a weak prompt. It is the predictable output of a system asked to produce text without the inputs that make text specific. Four failures follow from that, and each has a different fix.

Sameness. A model trained on the consensus produces the consensus. Run the same prompt across twenty pages and you get twenty pages that argue the same way, structure the same way, and reach for the same transitions.

Self-cannibalisation. Sameness has a search consequence most firms never diagnose: multiple pages on your own site competing for the same query, because the model kept writing the same article with different headings. You do not lose to a competitor. You lose to yourself, and it takes a site crawl to see it.

Voice erosion. A model does not know your voice and is particularly poor at reproducing a distinctive one. Publish enough unvetted output and the thing that made your firm sound like your firm quietly disappears, one adequate paragraph at a time.

Fabrication. Models invent statistics, misattribute quotes, and cite sources that do not exist, fluently and confidently. On any topic where trust is the product, one fabricated figure costs more than the content saved.

Three of those four are invisible to the person producing the content and obvious to the reader, which is why internal quality checks so rarely catch them.

What a model cannot supply

The fix is not better prompting. It is supplying the four inputs a model structurally cannot generate.

• Proprietary data. What your clients actually said, what your engagements taught you, the numbers only you hold. A model can write about customer acquisition cost; it cannot tell the reader what happened to yours.
• Lived experience. The specific detail that comes from having done the thing, including the parts that went badly. This is the single strongest signal that a human was involved, and it cannot be faked at scale.
• A position. A model is trained to produce the middle of the distribution. A position is a departure from it. Content that refuses to take one reads as thorough and forgettable.
• Accountability. A named person who answers for the claim. This is why a byline and a real author profile do more for credibility than another thousand words.

Every one of those is a reason the scarcity moved rather than disappeared. When production costs collapse, the premium shifts to whatever production cannot manufacture.

A workflow that survives contact with a core update

The teams getting this right have converged on roughly the same shape, and it is not “write with AI, then proofread.”

Decide the position first, before any generation, because it is the thing the model cannot supply and everything downstream inherits it. Give the model your context, not just your topic: the transcripts, the real objections, the numbers. Let it draft, which is genuinely what it is best at. Then do the part that creates the value: cut the consensus paragraphs, insert what only you know, verify every figure against a primary source, and put a real name on it. Finally, check the new piece against your existing library so you are not publishing your fourth article on the same query.

The order matters more than the tools. A team that adds its position and its data at the end is editing sludge. A team that supplies them at the start is directing a draft.

This is the same seam running through what AI commoditises and what it does not: the input gets cheap, the judgment does not, and the firms that win are the ones that stop selling the input. It is also why AI content and AI search reward the same thing, and why content only compounds when it runs on a context layer you own rather than a subscription anyone can rent.

The test worth applying

Before publishing anything AI-assisted, one question settles it: could a competitor with the same tools and the same prompt have produced this? If yes, publishing it adds a page to the internet and nothing to the business. If no, name the reason it could not be copied, because that reason is the only part doing any work.

Most firms do not need to publish less or publish more. They need the version of their content that only they could have written, which is a different problem with a different fix.

If your content no longer sounds like only you could have written it, tell Daniel what you publish, what you have built to produce it, and what it needs to do for the business. Book 30 minutes with Daniel.

Frequently Asked Questions

Does Google penalise AI-generated content?

No. Google penalises scaled content abuse, which requires two conditions at once: high volume and little value added. AI-assisted content is explicitly permitted, and around 86.5% of pages in the top 20 contain some. What fails is unedited, undifferentiated output published at scale. The tool is not the violation; the absence of added value is.

Can AI content rank on Google?

Yes, and it already does at scale. A Semrush study found 57% of AI-generated and 58% of human-generated content reached the top 10, effectively parity. But a 16-month study of 4,200 articles found content published without editorial enhancement ranked 23% lower on average, and pure AI content fell 18% after the March 2026 core update while human-written content rose 11% (Digital Applied, 2026). Editing is the variable.

Why does AI content sound generic?

Because a model produces the centre of its training distribution, and the centre is what every competitor’s model produces too. Without your proprietary data, lived experience, and an actual position, the output is the consensus rendered fluently. That is a structural property of how the models work, not a prompting mistake you can fix with better instructions.

What is keyword cannibalisation and how does AI cause it?

Cannibalisation is several pages on your own site competing for the same query, so none of them ranks as well as one strong page would. AI causes it because the same prompt produces the same article repeatedly under different headings. The firm ends up competing against itself, and because each page looks fine in isolation, it usually goes undetected until someone crawls the site.

How do you edit AI content so it actually performs?

Supply what the model cannot: decide the position before generating, feed it your real context rather than just a topic, then cut the consensus paragraphs, insert proprietary data and specific experience, verify every figure against a primary source, and publish under a real byline. Finally, check it against your existing content to avoid competing with yourself.

Is it worth using AI for content at all?

Yes, for drafting, restructuring, repurposing, and volume work, which is where it genuinely outperforms. The mistake is treating it as the author rather than the drafter. Used to produce content nobody could distinguish from a competitor’s, it adds cost and risk. Used to remove the production bottleneck so your team spends its hours on the parts that cannot be copied, it pays.