I was researching article topics recently and when scanning the SERPs and LinkedIn for ideas… I found the same piece over and over and over again. Not just the same topic, but the same angle, narrative, and in some cases, an almost word for word introduction.
When did we stop brainstorming content ourselves? When did we hand creativity over to ChatGPT?
For example, go and search “schema markup in 2026”. Do it now, in another tab, before you read the rest of this. Because the reason every article and blog post sounds the same is sitting right there on page one, and it is more obvious than any argument I could make about it.
Here is what came back when I ran it. “Schema Markup: What It Is and Why It Matters in 2026.” “Schema Markup: The Complete Guide 2026.” “Schema Markup Guide 2026: SEO, AEO and GEO Explained.” “Schema Markup: A Practical Guide for 2026.” “Schema Markup in 2026: The Complete Guide to Structured Data.”
Seven results, the same two-word opener, the same year bolted on the end, four of them calling themselves a guide.
Underneath the titles it gets worse. At least four open with a near-identical sentence explaining that schema markup, also called structured data, is code you add to a page. Three run the same argument that machines cannot rank what they cannot understand. Two independently reach for the same metaphor about speaking a language search engines understand.
Something has gone wrong upstream of the writing.
The AI writing theory doesn’t survive the data
AI-written prose is not what makes search results feel identical. Ahrefs analysed 600,000 pages across 100,000 keywords in 2025 and found that 86.5% of top-ranking pages contained some AI-generated content, with a correlation between AI content percentage and ranking position of 0.011. That is statistical noise. Google is not sorting by authorship.
Which is inconvenient, because “AI slop” has become the industry’s favourite explanation for why the web feels flat, and it is the wrong one. The machine can write. Give it a decent brief and it will produce clean, readable, factually reasonable copy that a human editor can tidy up in twenty minutes.
So if the prose isn’t the problem, and the ranking data says Google doesn’t mind, why does the first page of any commercial query read like one article that has been photocopied seven times?
Why does AI give everyone the same content ideas?
AI tools give everyone the same content ideas because most marketers now use AI at the ideation stage, and most of them use the same model. In a survey of 879 content marketers published in May 2026, Ahrefs found that 76% use AI to brainstorm topic ideas and 73% use it to build content outlines. Fewer use it to improve the writing.
Read those numbers again, because they invert the story everyone tells about AI content. The tool is being used more heavily to decide what to say, rather than just how to say it. Ideation homogenisation is what happens when a large group of people independently ask the same model the same question and receive answers drawn from the same narrow band of probable text.
Picture the actual workflow.
A content lead in Leeds opens ChatGPT on a Monday morning and types “give me ten blog ideas about schema markup for 2026”. So does an agency in Manchester. So does a freelancer in Bristol and an in-house team in Reading.
The model is well-behaved and consistent, so it gives all four of them broadly the same ten ideas, in broadly the same order, with broadly the same eight headings underneath.
Four weeks later those four articles publish. Nobody plagiarised anyone. Everyone used a slightly different tone of voice and their own examples. And the SERP still looks like a hall of mirrors.
Our own human-in-the-loop marketing piece put the underlying issue plainly: everyone has access to the same models and broadly the same training data, so baseline output now sits at parity across the industry. Ship the raw version and you ship what your competitors are shipping.
Does using AI for brainstorming reduce originality?
Yes, and there is now a decent body of research saying so. A randomised experiment presented at CHI 2025 put 460 participants through creative tasks with and without AI assistance and found that LLM exposure did not improve originality, and in some cases reduced the diversity of ideas people produced afterwards.
The detail that should worry every content team is the study design. Participants were split into three groups: one working alone, one given AI-generated ideas, and one given only AI-generated brainstorming strategies. Frameworks, not answers. The homogenising effect showed up in both AI conditions.
“I only use it to structure my thinking” is not the escape hatch people assume it is.
A separate analysis of 2,200 college admissions essays, published in 2025, measured how many genuinely new ideas each additional essay contributed to the pile. Every extra human-written essay added more novelty than every extra GPT-4 one, and the gap widened as the sample grew. Researchers tried to prompt their way out of it by adjusting instructions and model parameters. The effect held.
The pattern across all of this work is consistent and slightly uncomfortable. Individual output improves. Collective variety collapses. One person using AI to brainstorm gets more ideas than they would have had alone, and ten thousand people using AI to brainstorm get fewer ideas between them than they would have had alone.
When did we stop writing content for people?
Somewhere in the last few years the reader stopped being the first consideration in most content briefs. The brief now starts with a keyword, inherits a structure from whatever currently ranks, gets shaped to be extractable by an answer engine, and arrives at the reader last, if at all. Nobody decided this. It accumulated.
The scale of it is documented. Salesforce’s State of Marketing 2026 found near-total adoption, near-total sameness, running side by side.
Where we put the human
We should be straight about our own position here, because Roar is a digital agency and we use these tools every day. There are prompt templates, skill files, a documented banned-vocabulary list, and a house style that exists partly to strip out the tells we know models produce. Pretending otherwise would be daft, and you would spot it anyway.
The question worth arguing about is where the human sits in the workflow. Our model puts human judgement at three points: the brief, the editorial review, and the final approval. AI handles volume in between, and you can read the full version of that argument in our piece on how AI models pick up verbal habits nobody asked for.
Most teams have that loop the wrong way round. They hand the model the brief and keep the human for the tidying up, which is the one arrangement guaranteed to produce work that reads well and says nothing new. The brief is where the idea lives. Give that part away and no amount of careful editing at the other end puts it back.
What sameness actually costs you
Publishing a page that adds nothing to what already ranks is a visibility problem, not a taste problem. Google holds a patent on information gain scoring, written to address the situation where documents sharing a topic all contain similar information. Search Engine Journal’s July 2026 breakdown of the patent notes that documents scoring low on novelty can be reranked, demoted, or dropped from the results a user sees next.
Information gain is the measure of how much genuinely new information a page adds beyond what the reader has already seen on that topic. Score near zero and there is no reason for any system to show your page to someone who has read the one above it.
The same analysis points to two systems surfaced in Google’s leaked documentation: OriginalContentScore, and contentEffort, described as an LLM-based estimate of how much effort went into producing an article page. Whatever their exact role in live ranking, the direction of travel is not subtle.
The AI search side is starker still. Kevin Indig’s research, cited in that same breakdown, found first-party research is rare in AI citations but earns 3.3 times more of them when it appears, and that original data is the strongest single predictor of a page’s originality. Given how AI systems actually retrieve and compare sources, that makes sense. A model running several searches behind one prompt has no reason to cite the fourth page telling it the same thing, and every reason to cite the one telling it something new. This is the mechanism behind being cited by AI rather than merely being crawled by it.
Ten near-identical articles compete for one slot. The one with something nobody else has takes it.
How to come up with blog ideas that aren’t a regurgitation of the first page of Google
Use your brain.
That is the whole answer, but “be original” is advice nobody can action on a Tuesday afternoon with three briefs due.
Start somewhere other than a keyword. The first page of Google is a record of what has already been written, so using it as your idea source guarantees you arrive at the same place as everyone else who did that. Look instead at what only you have:
- Your client accounts. What did the data do last quarter that surprised you?
- Your inbox. Which question have three different clients asked in the last month?
- Your failures. The campaign that underperformed is a better article than the one that worked, and almost nobody writes it.
- Your arguments. If two people in your team disagree about something professionally, that disagreement is a piece nobody can copy.
Then bring AI in afterwards. Once you have an angle that is genuinely yours, the model becomes useful again: sharpening structure, finding the gaps, checking the counterargument, tightening the draft.
One test before you commission anything. Would this piece still be worth publishing if it ranked nowhere? If the honest answer is no, you have not had an idea. You have had a keyword.
Conclusion
Go back to that schema markup search in a year and it will look the same, because the incentives that produced it have not changed and the tools have got faster. Volume is cheap now. Having something to say is the only part that stayed expensive.
Two things are worth taking from this. The sameness starts at the brief, not the draft, so that is where the intervention has to happen. And the commercial cost is real rather than aesthetic, because search systems and answer engines are both built to skip the eighth version of a page they have already shown someone.
Generative engine optimisation is the practice of structuring content so AI-powered search platforms can read, extract, and cite it in their answers. Structure gets you eligible. Substance gets you picked.
If you want your content to be the one that gets cited rather than the one that gets skipped, take a look at our GEO solutions.
Frequently asked questions
Does Google penalise AI-written content?
No. Ahrefs analysed 600,000 pages across 100,000 keywords in 2025 and found 86.5% of top-ranking pages contained some AI-generated content, with a correlation between AI content percentage and ranking position of just 0.011. Google assesses quality rather than authorship. What does get penalised is scaled content abuse, meaning mass-produced low-value pages, however they were made.
Why do AI tools give everyone the same content ideas?
Large language models produce the most statistically probable response to a prompt, so different people asking similar questions receive similar answers. When a large share of an industry uses the same model for ideation, the effect compounds across every SERP. Ahrefs found in May 2026 that 76% of 879 surveyed content marketers use AI to brainstorm topic ideas.
What is information gain in SEO?
Information gain is a measure of how much new information a page adds beyond what a reader has already seen on the same topic. It comes from a Google patent designed to stop users being shown several documents that say the same thing. Pages scoring low can be reranked, demoted, or excluded from the results shown next.
Is it bad to use AI for content ideation?
Using AI for ideation reduces the variety of ideas produced across a group, even when the AI only supplies brainstorming frameworks rather than the ideas themselves. A CHI 2025 experiment with 460 participants found LLM assistance did not improve originality and sometimes reduced idea diversity. AI is more useful once you already have an angle, applied to structure, gaps, and drafting.
How do you make content stand out in an AI-saturated SERP?
Start from something the competition does not have: client data, first-hand results, original research, or a position you are willing to defend. Research cited by Search Engine Journal in July 2026 found first-party research earns 3.3 times more AI citations than content without it. Structure and schema help a page get read by machines, but only original substance gives them a reason to pick it.
Why does every blog post have the same structure?
Most content teams build outlines from what already ranks, and 73% of content marketers surveyed by Ahrefs in May 2026 use AI to generate those outlines. Both methods work backwards from existing pages, so the same H2s reappear across every result. Structural sameness is a symptom of shared source material at the planning stage.
