When we think about optimizing for AI (GEO / LLM SEO), the conversation often revolves around structure, citations, extractability, and freshness. All true. But a new paper, When Content is Goliath and Algorithm is David, introduces a nuance that’s underappreciated: the act of polishing your content – tightening language, refining structure – doesn’t just boost clarity. Paradoxically, it increases citation diversity by generative models. (arXiv)
That’s a strategic lever worth understanding.
TL;DR
Polishing your content – tightening language, improving structure, and clarifying phrasing – doesn’t just make it easier to read. It increases how often and where AI models cite you. A new study (“When Content is Goliath and Algorithm is David”) found that when websites refined their content, generative search systems (like AI summaries) pulled from more diverse sources and included those polished pages more frequently.
Key takeaways:
- AI visibility isn’t winner-takes-all – it’s relational.
- Over-optimized or formulaic writing can make you invisible to LLMs.
- Polished, expressive, semantically coherent content helps both users (clarity) and models (predictability).
- Track “citation diversity” – if the model cites more sources after polishing, you’re moving into a richer ecosystem.
What the authors did, and what surprised them
The research team collected data from ~10,000 websites across traditional search and generative (AI) search channels. They measured which sources AI summaries cited, and analyzed traits like phrasing predictability and semantic similarity. Their findings:
- Generative engines favor sources with high predictability – phrasing and structure models are more familiar with.
- They also choose sources that are semantically similar to each other, reinforcing cohesion.
- In controlled experiments, when content was polished (rephrased, cleaned up), the AI’s citations expanded – it cited more sources, not fewer.
- On the user side, polished content reduced task completion time and increased information density (especially for less expert users). (arXiv)
These insights push back on the idea that “optimize until you’re a narrow fit and the model picks only you.” Instead, polish can help you sit more comfortably in the model’s reference ecosystem.
Why this matters in practice
Here’s why this nuance can shift your approach:
- Visibility is relational, not solo.
AI doesn’t always pick a “winner page” and only that – it often draws from multiple sources. Polishing helps your content be part of that relational set. - Avoid overfitting.
If your content is too formulaic or “obvious,” the model might skip it as redundant. Polished but still expressive content helps avoid that. - User + model alignment.
Polished content helps users (clearer path, less noise) and helps models (predictability). It’s one of the rare win-win areas. - Citation density as a metric.
Instead of fixating on “how many times I get cited,” measure how many total sources the model includes. After polishing, if citation diversity increases, that’s a signal your content is entering a richer context.
Tactical playbook for “smart polishing”
- Incremental polishing: don’t rewrite everything at once. Start with high-traffic, high-opportunity pages, refine passages, and monitor citation changes (via AI trace tools or sampling).
- Preserve variation: keep some sentences that are more speculative, illustrative, or contrastive, so the model can link to different frames.
- Use anchor snippets + narrative text: fact‑dense, bulletable “liftables” alongside narrative sections gives the model room to connect.
- Monitor citation map shifts: before and after, see which sources AI picks. If it narrows too much (only you), reconsider structural diversity.
- Refresh over time: periodic light polishing keeps your signals aligned with model expectations without overcooking.
Bottom line: Don’t just optimize – refine. Smart polishing helps your content become part of the model’s conversational context, not just an isolated page.
