LLM Citation Optimization: What the Studies Actually Show

99Visibility Research · Updated October 21, 2026 · 8 min read

Short answer

Most LLM citation optimization, as it is sold, has weak evidence behind it. C-SEO Bench, a 252,000-trial factorial and a 2026 survey of 45 studies all point the same way: topical relevance and position in the model's context drive citation. Citation-oriented rewrites are mostly ineffective and sometimes harmful. The one positive on-page signal is stating a specific fact, such as a price, on a page AI crawlers can fetch.

Key takeaways

Nearly every published GEO effect was measured on pages already inside the model's context, so it says nothing about how a page gets retrieved in the first place.
C-SEO Bench found most conversational SEO rewrites were largely ineffective and frequently lowered a document's ranking.
The 252,000-trial factorial and the 2026 survey both name topical relevance and context position as the main drivers of citation.
The only positive on-page signal is a specific, quotable fact (a price, a number, a one-line definition) on a page crawlers can actually read.
Your real levers are access levers: crawler permission, firewall reachability, server-side rendering and snippet eligibility.

What citation studies actually measure

An AI answer passes through four stages, and it helps to know which one a study tested.

Parametric: what the model learned before its knowledge cutoff. Nothing you publish this quarter changes it.
Access: whether a crawler is allowed to fetch and store your page.
Retrieval: whether your page enters the model's context for a given query.
Generation: which retrieved documents the model uses and cites.

Nearly every published GEO effect size was measured at the generation stage. Researchers placed documents directly into the model's context, with no live search call, then edited them and counted citations. A finding like "adding quotations raises citations" therefore means one thing: once your page is already in the context window, this edit may change whether the model cites it. It says nothing about getting into that window.

Citation is also a noisy signal of influence. A 2026 mechanistic study (arXiv:2606.28358) repaired more than 90% of missed citations by amplifying specific attention heads, without changing the sources at all. The attribution mechanism study explains why an AI citation and actual use of your content are not the same thing.

The next move: read any citation claim by asking which stage it tested.

C-SEO Bench: most rewrites fail or backfire

C-SEO Bench (NeurIPS Datasets and Benchmarks, 2025) tested conversational SEO methods across several model families, including GPT-4o-mini, Claude-3.5-Haiku, o3 and o4-mini. Its headline result is blunt: "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking, which is opposite to what is expected."

The authors found that "traditional SEO strategies" were "significantly more effective." Read the method before quoting that line. The SEO arm was implemented by moving the target document to position 1 in the model's context. No index was queried and no ranking was earned. It is a position shortcut that no publisher can trigger.

The benchmark also found that as more sites adopt the same tactics, the gains shrink. The authors describe the problem as "congested and zero-sum." A rewrite trick that works for one site stops working when competitors copy it.

What it means for you: rewriting pages to win citations is a crowded bet with a negative expected result in the best-designed test available. The full breakdown is in our C-SEO Bench review.

The 252,000-trial factorial: relevance and position

The largest controlled test, "What Gets Cited" (arXiv:2605.25517), ran 252,000 trials across multiple models. Its mixed-effects models show that "topical relevance and list position are the biggest drivers of being cited first." It also found that "formatting-only edits have little impact." Headers, bullets and bolding did not move citation in any meaningful way.

One result points toward content that helps. The authors report that "including explicit price information and a recent timestamp also helps consistently." The price finding fits the one positive signal covered below. Treat the timestamp finding with care: publishers set their own date fields, and nobody has published how many uncited pages carry recent dates too.

Know the limits before you lean on it:

Each trial had only two candidate documents, which likely overstates position effects compared with real answer engines.
The corpus was a synthetic set of 100 e-commerce blogs with brands anonymised by design.
No live search engine was ever called.
The authors work at Sprinklr, a company that ships a GEO checklist.

Our factorial study review covers the design in detail.

The 2026 survey: what reproduces across 45 studies

A July 2026 critical survey of 45 GEO studies (arXiv:2607.14035) pulled the field together. Its summary: "topical relevance and context position are the most reproducible levers, generic heuristics transfer poorly, competition can erode individual gains, and citation-oriented rewrites can impair retrieval."

Its broader conclusion matters more: "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." Most of the market sells the second thing. The evidence supports only the first.

The survey has its own weaknesses. It is single-author, unaffiliated and not peer reviewed, with no primary data. Its agreement with the factorial is corroboration from two interested sources, not independent replication.

The widely quoted "up to 40% more visibility" figure comes from the original GEO paper (arXiv:2311.09735). That study ran on an engine the authors built, with sources already supplied in context. The same paper found keyword stuffing lowered its own visibility metric, scoring 17.8 against a 19.5 baseline. Before trusting a vendor's uplift claim, read our guide on whether to trust GEO case studies.

The one positive signal: specific facts on a fetchable page

Strip away the rewrite tactics and one on-page signal survives. Stating a specific fact (a price, a number, a one-line definition) gives an engine a sentence it can quote. It is the only positive experimental signal for on-page content in the evidence, and it fits the factorial's price result.

Two conditions make it work.

The fact has to be readable without JavaScript

Vercel's crawler study found that ChatGPT's and Claude's crawlers fetch JavaScript files but "don't execute them. They can't read client-side rendered content." If your pricing table loads only after hydration, those crawlers see an empty box. Googlebot renders JavaScript, so this gap mostly affects non-Google engines.

The page has to be relevant to the real question

Topical relevance is the factor every study agrees on. That means fit with what the buyer asks, not keyword coverage. Google's AI optimization guide says you don't need "to write in a specific way just for generative AI search."

A hypothetical: say you sell scheduling software. "Plans start at $29 per user per month" in plain server-rendered HTML is quotable. A price slider built in client-side JavaScript is not. Our guide to writing a blog post for AI citations applies this to editorial content.

What you control and what you don't

Be clear about where your effort lands. Most of what decides a citation sits outside your site.

Stage	What decides it	Your control
Parametric	Model training data and cutoff	None this quarter
Access	robots.txt per bot, firewall or CDN rules, rendering	High
Retrieval	The engine's own rewritten queries and ranking	Partial: relevance and index eligibility
Generation	Context position and the model's choices	Low: relevance and quotable facts

The access stage is where checkable fixes live, and they work by removing blocks. There is no markup, phrasing or structure with credible evidence that it makes a page more likely to be chosen.

Retrieval is also per query and discretionary. As of October 2026, ChatGPT decides when to search, Google AI Overviews often do not appear, and engines rewrite the user's prompt into their own search queries. No one outside the engine sees those queries in full. Nobody can promise you a citation, and any vendor who does is overstating what anyone controls.

A practical order of work

Work from the gates inward. Each step only matters if the one before it passes.

Check crawler permission per bot. robots.txt rules are independent for each crawler. OpenAI's docs say OAI-SearchBot controls ChatGPT search, while GPTBot covers training only. Also check PerplexityBot and Claude-SearchBot.
Check the firewall. A CDN or firewall can block a crawler that robots.txt allows, often with a 403. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
Check rendering. Compare raw HTML with the rendered page. Move prices, product names and key claims into the server-rendered HTML.
Remove accidental suppression. Look for noindex, nosnippet, max-snippet:0 and data-nosnippet left on templates. For Google AI Overviews and AI Mode, a page must be indexed and eligible for a snippet.
State facts once, plainly and consistently. Put the same price and the same product names on the pricing page, about page and footer.
Measure each engine separately. The free checker on this site gives a snapshot of whether engines name you for three buyer questions.

For engine-specific detail, see what you control in ChatGPT citations.

Common mistakes

These tactics are widely sold. The evidence does not support them.

Rewriting for citation. Appended quotation blocks, injected statistics and "according to authoritative sources" boilerplate match the tactic family that C-SEO Bench found ineffective or harmful. Treat them as risks, not wins.
Adding schema to get cited. Ahrefs' matched-control study of 1,885 pages found AI Overview citations fell 4.6% after schema was added. Existing markup does no harm, and whether schema helps a page with no citations is untested. It is not a lever for AI answers.
Chunking content for LLMs. Google rejects chunking as a requirement and says there is no ideal page length. Engines segment pages inside their own systems.
Publishing an llms.txt. Google says Search ignores it, and no other engine documents reading it.
Treating citation counts as influence. Engines retrieve more sources than they cite, and citation is mechanically separable from use.
Hidden text or prompt injection. Research on LLM search agents (arXiv:2606.16821) shows manipulated content can sway answers once it is retrieved. It is illegitimate, it breaks engine and search policies, and it puts your brand at reputational risk. Never use it.

Frequently asked questions

Does adding statistics and quotations get my page cited by LLMs?

Not reliably. Those edits come from lab studies where the page was already inside the model's context. C-SEO Bench found most such rewrites were largely ineffective and often lowered ranking, and the 2026 survey warns that citation-oriented rewrites can impair retrieval. A real fact you state because it is true, such as your price, is different from a statistic inserted to look authoritative.

Does ranking in Google's top 10 still matter for AI citations?

It is one input, not a guarantee. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the first 10 results, in a March 2026 study. The rest came from lower positions. Ahrefs changed its method between studies, so earlier and later figures are not a measured trend. Index and snippet eligibility is the hard requirement for Google's AI features.

Is schema markup useful for LLM citation optimization?

The evidence says no for pages already cited. Google states structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after schema was added. Whether schema helps a page with no citations enter the candidate set is untested. Existing markup does no harm, so there is no need to remove it either.

Why do different AI engines cite different pages for the same question?

Each engine uses its own index, its own query rewriting and its own models. As of October 2026, Google says AI Overviews and AI Mode may use different models and techniques, and ChatGPT rewrites prompts into its own searches. A page can be retrieved by one engine and missing from another, which is why per-engine measurement beats a single blended score.

Can I make content more likely to be picked once it is retrieved?

Only in narrow ways. Topical relevance to the question is the most reproducible factor across studies, and a specific quotable fact gives the model a sentence to use. Context position also matters a lot, but no publisher can set it. Formatting-only edits had little impact in the 252,000-trial factorial, so restructuring pages for AI is not worth the effort.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.