Goliath and David: Predictable Prose and AI Citation (arXiv 2509.14436)

The paper is "When Content is Goliath and Algorithm is David: The Style and Semantic Effects of Generative Search Engine", posted to arXiv as 2509.14436 (version 1, 17 September 2025). It asks a narrow question: when a generative engine picks which retrieved sources to cite, does the style of those sources matter?

The style property it focuses on is perplexity. Perplexity measures how surprised a language model is by a piece of text. Low perplexity means the model finds the text predictable. High perplexity means it finds the text unexpected.

The study was run on Google's generative surfaces and on a retrieval engine the authors built themselves. A robustness appendix covers Bing and Copilot. It is a single preprint, and our evidence base records it as an emerging finding that is not actionable.

It also sits in a specific place in the pipeline. AI answers pass through four stages: what the model already knows (parametric), whether a crawler can fetch your page (access), whether your page enters the model's context for a query (retrieval), and which retrieved pages get cited (generation). This paper is about the last stage only. It says nothing about how a page gets retrieved in the first place.

The paper has two arms, and they need to be read separately.

The observational arm

The authors looked at what generative engines actually cited and compared the perplexity of cited content with content that was not cited. This tells you about correlation. It cannot tell you why the cited content was chosen.

The causal arm

To test cause, the authors rewrote pages with an LLM and checked whether citation changed. The rewrite was whole-page. It changed register, structure, hedging, terminology and length all at the same time. Perplexity dropped as a side effect of all those changes, not as a single isolated variable.

That design choice decides how far the result can be pushed. Our explainer on lab GEO studies versus live engines covers why within-context experiments like this rarely transfer cleanly to production search.

The headline result, in the authors' words: "Our empirical analysis reveals that GEs exhibit preferences for citing content characterized by significantly higher predictability for underlying LLMs and greater semantic similarity among selected sources."

Two findings sit in that sentence:

• Predictability. Cited content was more predictable to the model doing the citing.
• Similarity among sources. The sources an engine selected together tended to resemble each other semantically.

The paper also reports a contrast that matters for anyone tempted to rewrite their site: "perplexity measurements demonstrate minimal influence on ranking within conventional Google search results." So the effect, where it shows up, is a generation-stage preference. It is not a ranking signal in classic search.

This fits a broader pattern. The critical survey of 45 GEO studies (arXiv 2607.14035) concludes that "already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." The perplexity paper is one more result about content already in context.

It does not show that writing simply works

The causal arm rewrote pages with an LLM. The fair reading is that having an LLM rewrite your page changed citation in that setup. You cannot separate perplexity from the other changes the rewrite made. "Write simply" is a different intervention that nobody tested here.

The observational arm is confounded

Predictable text correlates with other things that plausibly drive citation. Publishers with larger editorial budgets produce polished, conventional prose. Text that resembles the model's training data is predictable to that model by definition. Either could explain the correlation without perplexity doing any work.

Perplexity is model-specific

Perplexity is always relative to one model. Text that is predictable to one model can be less predictable to another. The study predates every 2026 frontier release, and as of October 2026 the consumer apps route between several models per surface. A number tuned to one model has no guaranteed meaning for the one answering your buyer.

Other evidence points the other way on rewriting

C-SEO Bench (arXiv 2506.11097) found that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." The 252,000-trial factorial (arXiv 2605.25517) found "formatting-only edits have little impact." The 45-study survey warns that "citation-oriented rewrites can impair retrieval." Google's AI optimization guide says: "You don't need to write in a specific way just for generative AI search." We cover the wider argument in whether an AI-friendly writing style exists.

To act on this finding, you would need the perplexity of your page under the model that cites it. That model is closed. Its token probabilities on your text are not exposed to you.

You can compute perplexity under an open-weight model. That number is a proxy, and an uncalibrated one. It measures how predictable your text is to a different model than the one producing the answer. Any tool that reports a "perplexity score" for AI visibility is reporting that proxy.

There is a second problem. If whole-page LLM rewriting is the thing that moved citation, chasing the effect means rewriting your content with a model. That raises separate risks. Mass-generated pages without added value are what Google's spam policies call scaled content abuse. Our page on whether AI engines prefer AI-written content looks at that question directly.

Treat this paper as research context. Do not let it drive editorial changes. Here is where effort has evidence behind it instead.

1. Check access first. Most AI crawlers do not execute JavaScript. Vercel's crawler study found ChatGPT and Claude crawlers "don't execute them. They can't read client-side rendered content." Make sure the facts that matter are in the raw HTML.
2. Check robots.txt per bot. Rules are independent. OAI-SearchBot controls ChatGPT search, while GPTBot controls training use. Check PerplexityBot and Claude-SearchBot too.
3. Check the firewall. A CDN or firewall can block a crawler that robots.txt allows. That failure is silent.
4. State specific facts plainly. A price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote. This is the only positive experimental signal for on-page content.
5. Stay topically relevant. The survey and the factorial both name topical relevance as one of the most reproducible factors. That means fitting the question, not stuffing keywords.

None of this guarantees a citation. Which retrieved pages get cited is the model's decision. If you want a baseline before changing anything, the free checker on this site shows whether AI engines name you for up to three buyer questions. For writing guidance that stays inside the evidence, see how to write a blog post for AI citations.