AI

The Trick That Didn't Replicate

Everyone is selling hacks to get AI to quote your writing. We went through the research. The shortcuts fail, and what works is simpler and harder.

Every week there’s a new hack for getting AI to quote your writing. Add statistics. Add quotes. Add FAQ blocks. Add schema markup. Rewrite everything for the machines.

We wanted to know which of these actually work, because this network lives or dies on being worth quoting. In The Machines Don’t Read Page One, we showed that AI tools quote pages far beyond Google’s front page. So the next question was obvious: what makes them pick one page over another?

We went through the evidence: the lab experiment behind the hacks, the study that tried to repeat it, hundreds of thousands of real AI citations, and the research on how people actually read. The short version: the famous shortcuts don’t hold up, and what does work is what readers have always wanted.

The famous +41%

Most “write for AI” advice traces back to one paper. In late 2023, a team of academic researchers tested how small edits to a web page changed its visibility inside AI-generated answers. Their study, GEO: Generative Engine Optimization, later published at the KDD 2024 conference, reported that adding quotations raised a page’s visibility by 41% and adding statistics by 31% on its main visibility score. Keyword stuffing made that score 8% worse.

Those numbers launched an industry. “GEO” became the new SEO.

But look at how it was measured: mainly one AI model (GPT-3.5, with a smaller check on Perplexity), one lab benchmark, and visibility scores the authors designed themselves.

Then someone checked

In 2025, a separate team built a benchmark called C-SEO Bench and tested ten of these rewrite methods on newer models, including GPT-4o-mini, Claude 3.5 Haiku, o3 and o4-mini. Across 54 test cases, only 3 showed a significant improvement. The “add statistics” rewrite actually lowered rankings in 19 of 24 settings. (The study was peer-reviewed and accepted at NeurIPS 2025’s Datasets & Benchmarks track.)

What mattered far more than any rewrite was simply where a page landed in the list of results the model was given to read. And the small gains that did exist shrank as more competitors used the same method, which is what eventually happens to every trick.

What 468,000 real citations show

The most convincing evidence we found was posted last month. Two researchers, Chen and Koudas, analysed 468,037 citations in 158,099 answers to commercial questions from ChatGPT, Gemini and Perplexity, collected over two 90-day periods. Crucially, they compared pages within the same website, which strips out the head start that big, popular sites have.

Two things helped. A paragraph that closely matches the question was linked to 15.5% more citations per standard deviation of closeness, and a title that matches the question to 14.1% more. FAQ blocks, schema markup and author boxes showed no significant effect once pages were compared within the same website. Neither did length, headings, lists or tables.

Then the biggest number of all. In ChatGPT, how often a page was retrieved by the engine’s own searches explained 32% of the differences in citations between pages of the same site. Everything on the page combined explained 2.8%. (This one is a preprint, not yet peer-reviewed.)

Google says much the same thing in plainer words. Its official guidance says “specific optimization isn’t required” for AI Overviews and AI Mode, and its advice to site owners is to “focus on making unique, non-commodity content”.

The people side

Here’s where it gets interesting. The research on human readers points in exactly the same direction.

People skim. Nielsen Norman Group found that on an average web page, users read about 20% of the words. Chartbeat data published by Slate in 2013 showed 38% of visitors leaving without engaging with the page at all, and most of those who stayed scrolling only about halfway down. If your answer is at the bottom, most people never see it.

And effort beats tools. Orbit Media’s 2026 survey of 1,042 content marketers found that 92.4% now use AI for blogging, yet AI use “correlates with nothing” when it comes to results. What did help was original research, which improved the likelihood of performance by 50%. (It’s a self-reported survey, so treat it as a direction, not proof.)

So the machines and the people want the same three things: the answer early, real evidence, and something you can’t get anywhere else.

Why the machines still need checking

There’s one more finding we couldn’t ignore, because it’s about the tools we use ourselves. AI is a superb research assistant and an unreliable witness.

When the Tow Center for Digital Journalism tested eight AI search tools on 1,600 queries in 2025, asking them to identify the source of news excerpts, more than 60% of the answers were wrong. A larger study by the European Broadcasting Union and the BBC, covering 22 public broadcasters in 18 countries, found that 45% of AI answers about the news had at least one significant issue.

Even Anthropic, describing how it built its own research agents, noted that its early agents kept choosing SEO-optimised content farms over academic papers until human testers caught it.

That’s why our rule is simple: AI finds, humans verify. Every figure in a Dynamic Nexus post is checked against its original page before it goes out, and that check is done separately from the writing.

What this changes for the network

We’ve turned all of this into an internal playbook that every Dynamic Nexus post now follows. Without giving away the whole recipe, the spine of it is:

  • Answer first. The core finding goes in the opening paragraphs, and every section starts by answering its own question.
  • Bring something new. Original analysis, a connection across our nodes, a figure we recomputed ourselves. No rehashed roundups.
  • Show the evidence. Specific numbers, named primary sources, short quotes, every link checked.
  • Say how it was made. Research shows that specific disclosure of how AI was used is received better than a vague label, so we’ll tell you exactly.

And the point that matters most for anyone who writes: there is no trick. Write for the person asking the question, and the machines follow.

Have you tried a trick to get AI to quote your work? Did it do anything? Hit reply and tell us.

How we made this: AI tools helped us research and draft this post. We checked every figure, quote and link against its original source before publishing. Figures are as of 2 November 2026.

Signal from the Dynamic Nexus AI node.