Citable Content: What the Research Actually Shows

Version note: This post is based on the GEO (Generative Engine Optimization) paper by Aggarwal and colleagues (KDD 2024), the C-SEO Bench study by Puerto and colleagues (NeurIPS 2025 Datasets and Benchmarks) and the July 2026 survey "A Critical Survey of Generative Engine Optimization (2023–2026)".

Guess First

Bookstore's marketing team reads a blog post: "Research shows adding statistics and quotations to content boosts visibility in AI (Artificial Intelligence) answers by up to 40%." The proposal is to rewrite every book review within a week, adding "statistics and quotations". Where, and under what conditions, do you think that 40% was measured? Does the same effect hold in a real AI answer engine while competitors do the same thing? And think about this: what does adding statistics to a book review that has no statistics even mean?

The GEO Paper: What Did They Measure?

The study that coined the term GEO is the paper by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, published on arXiv in November 2023 and presented at KDD (Knowledge Discovery and Data Mining) 2024. The sentence from its abstract is quoted often: GEO can boost visibility in generative engine responses by up to 40%. To understand what that number means, you have to look at the experimental setup.

The authors built a benchmark called GEO-bench with 10,000 queries collected from nine sources. As the "generative engine" they used a setup that takes the top 5 Google results for each query and writes a cited answer based on them; the answer was written by GPT (Generative Pre-trained Transformer) family model GPT-3.5-turbo. They then rewrote the text of one of these 5 sources with nine different methods and measured how that source's visibility in the answer changed. There are two visibility metrics: the share of words in the answer attributed to that source, weighted by its position in the answer (Position-Adjusted Word Count), and a subjective impression score judged by an LLM (Large Language Model) (Subjective Impression).

The nine methods are: Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms. Quotation Addition, Statistics Addition and Cite Sources gave the best results; the best methods improved Position-Adjusted Word Count by about 41% and Subjective Impression by about 28%. The classic SEO (Search Engine Optimization) tactic, keyword stuffing, barely worked at all and even fell below the unoptimized text on Position-Adjusted Word Count. The authors also tried a 200-query subset on Perplexity.ai: Quotation Addition improved Position-Adjusted Word Count by 22%, Statistics Addition improved Subjective Impression by up to 37%, and keyword stuffing did about 10% worse. They also note that effectiveness varies by domain.

Its Limits: A Fixed Context and a Single Benchmark

These results hold within their own setup, and it's a carefully built study. But when it turns into a marketing sentence, three details get lost.

First, the source is already in the context. In the experiment, the optimized page is in the prompt from the start as one of the 5 sources the answer relies on. What's measured is not "is this page found?" but "how much space does an already found page take up in the answer?". In the chain from the first post of the series, that's the step after retrieval. Adding quotations to a page that doesn't get into retrieval changes nothing.

Second, a single benchmark and 2023's models. GPT-3.5-turbo writes the answer; today's engines use different models, query fan-out and different retrieval layers.

Third, a single actor. In the experiment only one source is optimized; the other four stay as they are. In the real world, competitors read the same blog post too.

C-SEO Bench: Most Methods Don't Work

The study that tests the third point directly is C-SEO Bench by Puerto, Gubri, Green, Oh and Yun, published in June 2025 and accepted at NeurIPS (Neural Information Processing Systems) 2025 Datasets and Benchmarks. It tests "conversational SEO" methods on two tasks, question answering and product recommendation, each with three domains, and most importantly it models multiple competing actors by varying the number of actors who adopt a method.

The findings are clear. Most current C-SEO methods are largely ineffective and can even hurt a source's ranking. Traditional SEO strategies are clearly more effective at improving a source's ranking in the LLM context. And as more actors adopt C-SEO methods, the gains shrink; the authors explain this by the congested, zero-sum nature of the problem. Here's the answer to the second opening question: once everyone adds statistics, adding statistics stops being an advantage.

The 2026 Critical Survey: No Durable Causal Effect

The survey "Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)", published on arXiv in July 2026, reviews 45 studies from November 2023 to July 2026. Note that it's a single-author preprint, not a peer-reviewed publication. The core of its findings:

• GEO is not a single ranking problem but a multi-stage, uncertain pipeline; each stage should be measured separately.
• The first GEO paper's gains hold within its own setup, with the source in a fixed context; they don't show organic discoverability or lasting traffic effects.
• The most reproducible levers are topical relevance and position within the context. Generic heuristics transfer poorly across platforms, competition can erode gains, and citation-oriented rewrites can hurt retrieval.
• Commercial audits find low source overlap across platforms and large run-to-run variability.

And the main sentence: none of the reviewed techniques shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream user behavior. The survey proposes a reproducible protocol: repeated measurements, paraphrase variants, control groups, human validation and accounting for multiple competing actors.

The Three Studies Side by Side

Practices Consistent with the Evidence

The conclusion from this table isn't "nothing works"; it's that the evidence supports classic good content practices. Both the research and Google's guide point in the same direction:

First-hand experience and original data. Google's emphasis on non-commodity content and the research's "topical relevance" finding end up in the same place: information that exists nowhere else is a real reason for an engine to pick you.

Clear definitions, verifiable claims and sources. The situation in which the GEO paper's "Cite Sources" and "Quotation Addition" methods worked is one where claims are genuinely sourced. A fabricated statistic or a quotation without context is the opposite; it's also a risk under Google's helpful content questions and spam policies.

Author and date information. Google's Who and How questions recommend making clear who the author is and how the content was made. A version note and an update date matter both for the reader and for an engine judging how current the information is.

A structure that answers the question directly. Headings that correspond to a question and a first sentence that answers it help the reader find what they're looking for fast. Google's passage ranking system also evaluates individual sections of a page. That's structuring content well, not fragmenting it.

As you may have noticed, nothing on this list is specific to AI. It's all already good SEO.

Reviewing a Post: "What's New in C# 14"

Let's read the post "What's New in C# 14: Extension Members, field and Null-Conditional Assignment" from this site's C# series against these criteria.

Strengths: the post opens with a version note saying the code was built and run with the .NET 10 SDK (Software Development Kit). Every example has its real output; for instance, the two-line output showing that the right-hand side is skipped in null-conditional assignment is measured evidence that can't be copied from anywhere else. The headings match the feature names exactly; someone searching for "the field keyword" can go straight to that section.

Room for improvement: the site's HTML (HyperText Markup Language) rules don't allow links in the post body; sources are named but not clickable. That's a deliberate editorial decision, but a reader or an engine can't reach a claim's source in one step. The post's summary gives a list of features rather than a clear answer in its first sentence; the site's own GEO notes recommend that a summary start with an answer-like sentence. Author and date information comes on the page through JSON (JavaScript Object Notation) based JSON-LD (Linked Data) markup and the visible meta line; nothing is missing there.

What this review doesn't say matters too: we don't know how much these changes would change how often the post is cited in any AI answer, and the research doesn't allow a reliable estimate of that.

Production Notes

Don't do mass rewrites. Rewriting all content to a "GEO template" can, as the survey found, hurt retrieval, and on the Google side it gets close to the risk of producing content at scale without adding value.

Make the change measurable. If you change something on a page, record the date and do the before/after comparison through the same channels. We'll see which channels give data in the last post of the series.

Label the level of evidence. When a GEO proposal is discussed in the team, separate "official documentation", "peer-reviewed research", "preprint" and "blog claim".

Common Pitfalls

Taking a benchmark result for a guarantee. The 40% is a share increase measured in a fixed context, with a single actor and 2023 models.

Adding data that doesn't exist. Adding a fabricated statistic or an out-of-context quotation to a review is lying to the reader and also a risk under Google's helpful content and spam rules.

Expecting an advantage from a tactic everyone uses. C-SEO Bench's zero-sum finding shows that every tactic loses value as it spreads.

Skipping retrieval and focusing on citation. A page that isn't found has zero citability.

A Small Experiment

★☆☆ Evaluate a post on your site against the four practices in this post: first-hand data, sourced claims, author and date, a structure that answers the question. Point to evidence on the page for each.

★★☆ Read the GEO paper's abstract and experiments section yourself. Write in your own words in which metric and in which setup the result quoted as "40%" in marketing posts was obtained.

★★★ Design the protocol the survey proposes, at small scale, for your own site: which questions you'll measure, on which engines, with how many repetitions, what your control group will be, and in which case you'll interpret the result as "no effect". Write only the plan; don't send bulk automated queries to an engine.

In the next post we look at a file Google says it doesn't need but other tools can use: the llms.txt spec, who uses it, and how to generate it in ASP.NET Core together with Markdown versions.