Generative AI did to writing what phone cameras did to photography. A good photo once required a dedicated camera and the skill to use it, until phones put both in every pocket.
Since late 2022, anyone can produce a fluent article in a minute, without the skill or the hours it used to take.
When a technology puts a capability in everyone's hands, having it stops setting anyone apart. What still stands out is the subject a photographer chooses or the point of view a writer defends.
Our position at Unveil SEO is that neither Google nor answer engines like ChatGPT care whether AI wrote your page. They care what the page adds that others don't, and that addition has to come from you.
Half of New Articles Are Now AI-Written, and Few of Them Rank or Get Cited
About half of new online articles are now written mainly by AI. Yet they make up only 14% of Google's top results, 7% of its first positions and 18% of the articles ChatGPT and Perplexity cite. The closer to the top, the rarer AI content gets.
Most of the AI content flooding the web repeats what already exists. Our view is that Google and answer engines stopped rewarding pages that only repeat, because they already have that answer elsewhere.
The numbers we found point the same way. They come from two studies by the same researchers, one on what gets published and one on what ranks or gets cited.
| Where the articles appear | Share written mainly by AI | Sample |
|---|---|---|
| Position one in Google | 7% | Articles ranking for 31,493 keywords |
| Google's first two result pages | 14% | Same Google sample |
| Cited by ChatGPT and Perplexity | 18% | 100 questions per topic category |
| Published on the web | About 50% | 55,400 random English articles, 2020 to 2026 |
One case escapes these figures, AI drafts that a human edits or rewrites. That is how we work, and the authors of the ranking study believe it may perform better.
At the extreme, Google's spam policy on scaled content abuse targets pages mass-produced to manipulate rankings, whoever or whatever produced them.
Neither Google nor ChatGPT Cares Who Wrote Your Page
Google and answer engines judge what a page adds, not who or what wrote it. Fully AI-written pages can reach the top three, but rarely. This looks less like a penalty on AI than a filter on generic content.
Put yourself in the place of Google or ChatGPT. Their job is to give the best answer to each question, and the tool behind the sentences says nothing about whether a source is reliable or expert.
Google has said so since February 2023, in its guidance on AI-generated content. What it rewards is original, high-quality content, however it is produced.
Where AI hurts is indirect. A page written with no human input rarely says anything Google or an answer engine can't find elsewhere, so there is no reason to rank it first or to cite it.
The rankings bear this out. A study of 150,000 pages from Google's top ten found fully AI-written pages in the top three, but only about one in twenty.
Its authors came to the same conclusion as us. Google targets bad content, and AI content is often bad.
Answer engines behave the same way. In a Princeton experiment presented at KDD 2024, pages enriched with statistics, quotations and sources became up to 40% more visible in AI answers, while keyword stuffing barely helped.
Google's rater guidelines seem to say the opposite at first. Updated in January 2025, they tell raters to give the lowest score to pages mostly generated by AI.
Read the full sentence and the condition appears. The lowest score goes to AI content made with little effort, little originality and little value, which is exactly our point.
What we take from it. Ask whether a page says something its competitors don't. A generic page fails that test whether a human or an AI wrote it.
The same filter works before ranking even starts. Heavily AI-written pages get indexed less often, and a page Google discovers but never indexes never gets a chance to rank.
AI Writes Smoothly but Stacks Too Many Arguments
AI writes smooth, correct prose. In technical content, its weakness is that it piles up more arguments than a reader can follow, when web readers look for one idea at a time.
We saw it in our first AI drafts for this blog. The sentences were correct and fluent, yet the pages were tiring to read.
The prose was never the problem. In technical content, AI tries to fit every argument in and forgets that each one needs room to be understood.
Readers don't read closely enough to sort it out. In Jakob Nielsen's 1997 study of how people read online, 79% of users skimmed new pages instead of reading them word by word.
His advice was one idea per paragraph, because readers skip any idea the first few words don't announce. We now write every article that way.
A typical generated paragraph. "Internal links play a crucial role in SEO. They help search engines discover pages, distribute authority, signal topical relevance, improve engagement and reduce bounce rates, making them essential to any content strategy."
The same topic, one argument. "Google finds new pages by following links from pages it already knows. A page that no other page links to depends on the sitemap alone to be found."
The first version lists five benefits and explains none. The second makes one claim and explains it, so a reader skimming the page keeps it.
Fluent writing no longer sets pages apart. Every competitor can produce it in a minute, so it cannot be what wins position one or a citation in an AI answer.
Unique Content Starts With What the Model Doesn't Know
AI pushes everyone who uses it toward the same ideas. Unique content therefore has to come from the author. We draw it from three sources, which are our positions, connections between fields and field experience.
Ask an AI model a question and you get the most likely answer, the same one every other model gives. An answer engine already knows it, so a page that only repeats it gives the engine nothing to cite.
Researchers have measured this on writers. In an experiment published in Science Advances, stories written with AI ideas were rated better, but they looked more alike than stories written without.
So we changed how we write. We bring the ideas, from three places a model cannot reach, and AI helps us shape and test them.
- A position you defend. A clear view on your subject, held against the strongest data that contradicts it.
- A link between fields that rarely meet. Research from another discipline applied to yours, such as findings from psychology used to explain an SEO pattern.
- Field experience. What you have seen on real projects, including what failed, which no model has read anywhere.
The Next Filter Will Judge the Whole Site
Looking two or three years ahead, we think this goes further. As pages look more alike, Google and answer engines will lean on the site behind them and on how much it can be trusted on its subject.
That trust is built in two ways. The brand earns mentions across the web, and internal links show how wide and deep the site's expertise runs.
The first half already shows up in the data. In a study of 75,000 brands, mentions across the web tracked visibility in AI Overviews far more closely than backlinks did.
Internal links play a different role. They don't create AI citations by themselves, but they show which topics a site covers in depth, the idea behind semantic clusters built with internal links.
Brand mentions bring authority from outside, and internal links decide which pages receive it. Unveil SEO crawls your site to show which pages already support a topic and which ones sit alone.
What this changes in your content plan. Each off-topic article pulls a site away from its core subject. Before writing a new page, check which existing pages it will link to and from.
How We Use AI Without Writing What Everyone Writes
Our blog runs on a pipeline where AI plays seven specialist roles and a human validates every stage. The author's positions come first, each one is tested against the strongest evidence against it, and every figure is traced back to its original study.
We built that pipeline around one rule. AI does the work a competitor could also do, and the author brings what nobody else can.
The work runs in six stages, from brief to publication, with a pause for human validation after each one. A correction at the brief costs a sentence, while the same correction after the draft costs a rewrite.
- 1
Start with three questions to the author
Before any research, the author answers three questions.
- What does the usual advice get wrong?
- What have you seen on client sites, your own sites or your product's data?
- How will the subject evolve in the coming years?
The broadest answer becomes the thesis, and every answer stays in the author's words.
- 2
Send the research after the counter-evidence
For each position, the research looks first for the strongest evidence against it.
A position comes back as holding, holding with a nuance, or contradicted. A contradicted one goes back to the author, who revises it, defends it or drops it.
- 3
Dig where competing articles don't
A role we call the Digger searches patents, official documentation, conference talks and research from neighbouring fields.
It also reads sources in other languages. French work on internal linking methodology, for example, rarely shows up in English articles.
- 4
Trace every figure to its origin
When a number circulates across several articles, we open the original study and check what was counted, on what sample and over what period.
Secondary sources often drop exactly that, and a figure read with the wrong unit produces a confident wrong claim.
- 5
Draw the conclusions no source states
Compiling sources only produces a summary. The value comes from conclusions that follow from the evidence without being written anywhere.
Each one is graded solid, plausible or speculative before it reaches the draft, because a forced conclusion does more damage than a missing one.
- 6
Write for a reader who scrolls
Paragraphs hold one or two sentences and stay under 40 words, with a list, table or callout at least every three paragraphs.
Each section opens on a 30 to 50 word answer that stands on its own, ready to be lifted into a featured snippet or an AI answer.
- 7
Ban sentence structures, not words
Lists of banned words go stale, because AI writing habits change with each model version. The em dash, long the best-known giveaway, has already become rare.
We ban structures instead, such as a short label followed by a colon that announces the point. It reads as generated whatever the words.
- 8
Review three times before anyone reads it
Three passes run before a draft is shown, on conformity, readability and substance.
The last pass names the weakest argument, which then gets strengthened, qualified or cut.
- 9
Link the new page into the site
A new article gets found faster when older pages on the same topic link to it, which is why internal links to new posts should come from old ones.
Where the argument touches a neighbouring subject, we name it in full in a sentence that stands on its own. The day that page exists, the sentence becomes a link.
Asking an AI to suggest these links works on a small site. Past a certain number of pages and topics, it can no longer keep the whole site in view.
Beyond that point, automating internal linking with AI needs a tool that maps every page before suggesting a single link.
Let AI handle what any competitor could also produce, and spend the time it saves on the part only you can write.
