# 45 GEO Studies Reviewed: No Technique Showed Stable Gains

> A survey of 45 GEO papers found no technique with a stable, cross-platform causal effect. What the evidence supports and how to prioritize anyway.

URL: https://llmagnet.com/blog/45-geo-studies-reviewed  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-09-19  
Updated: 2026-11-10  
Category: Research

A July 2026 survey of 45 GEO studies found no technique with a stable, long-term, cross-platform causal effect on discoverability, clicks or conversions. What it does support is narrower: content that has already been retrieved can be made more likely to be cited.

This doesn’t mean GEO is fiction. It means the advice has run ahead of the evidence, and that a good plan starts with what has been shown and treats the rest as tests.

## Contents

1. What the Survey Found
2. The Two Problems GEO Advice Conflates
3. What Has Reproducible Evidence
4. What Is Weak or Contradicted
5. What the Survey Doesn’t Cover
6. Why Controls Change the Answer
7. How to Prioritize
8. Measuring What Can Be Measured
9. How to Read the Next GEO Claim
10. The Practical Bottom Line
11. FAQ

## What the Survey Found

The paper is [a critical survey by Olivier Martinez](https://arxiv.org/abs/2607.14035), posted to arXiv on July 15, 2026. It covers 45 studies from November 2023 to July 2026. They aren’t all peer-reviewed: the set mixes peer-reviewed papers, accepted work, workshop papers and preprints, and the survey weights them accordingly. The survey is itself a preprint and a scoping review, so it maps the field and doesn’t pool effect sizes. Its cutoff was July 14, 2026.

Its central point is that the famous gains were measured at one stage only. The [Princeton GEO study](https://arxiv.org/abs/2311.09735) (Aggarwal et al., KDD 2024) placed a source in a fixed set of documents and measured how much of the answer drew on it. On Position-Adjusted Word Count, against baseline, Quotation Addition gained about 41% and Statistics Addition about 31%. These are benchmark tests, not proof for live engines.

The survey rejects “GEO increases visibility by 40%” as a general claim. The gain comes from one metric in one setup, and it says nothing about whether a page gets found in the first place.

The survey grades each major claim. Its ratings, paraphrased, without one row on security attacks:

| Claim | Confidence |
| --- | --- |
| A document already in the context can change its rank, citation or use | High |
| Relevance and context position are major determinants | High |
| Commercial engines differ from each other and change over time | High |
| Extractable evidence and suitable structure often help | Moderate |
| Optimized or learned methods often beat fixed heuristics | Moderate |
| Competitive adoption can erode individual gains | Moderate |
| A white-hat change durably improves discoverability across engines | Low |
| Citation scores predict clicks, conversions or revenue | Very low |
| “GEO increases visibility by 40%” as a general claim | Rejected |

## The Two Problems GEO Advice Conflates

Most GEO content treats AI visibility as one problem. The survey separates seven stages.

![Seven stages of generative search: activation, crawling and indexing, retrieval, reranking and context, generation and citation, absorption and fidelity, and attention, click and conversion. Most GEO studies optimize stages 4 and 5.](https://llmagnet.com/images/blog/charts/geo45-pipeline.png)

*Source: Martinez, arXiv, July 2026. Preprint. Stage descriptions are our paraphrase.*

Most studies work on the stages from context allocation to citation. Far fewer observe crawling, organic retrieval or what users do next. For planning, we’d focus on two of the stages.

**Discoverability** is whether a page enters the system’s candidate pool at all. Evidence here is thin, and the survey rates durable improvement across engines as low confidence.

**Citation selection** is whether a page already retrieved gets used in the answer. This is where controlled evidence is strongest, and the survey gives it high confidence.

The two can pull against each other. In SAGEO Arena, a test the survey reviews, across 171,003 documents and 2,700 queries, rewriting the body alone cut top-20 presence by about 9%, top-10 presence after reranking by 16% and final citation by 6%. A rewrite that looks good once a page is in the context can make it harder to retrieve.

![Change after body-only optimization in SAGEO Arena: top-20 presence down about 9 percent, top-10 presence after reranking down 16 percent, final citation down 6 percent.](https://llmagnet.com/images/blog/charts/geo45-sageo-body-only.png)

*Source: SAGEO Arena (Kim et al., 2026), as reported in the July 2026 survey. Across 171,003 documents and 2,700 queries.*

## What Has Reproducible Evidence

The survey names relevance and context position as “the most reproducible levers.” Relevance means the content directly answers the question. Context position means where a source sits among the documents given to the model, which isn’t the same as where a sentence sits on your page, and you can’t set it directly.

It gives moderate confidence to extractable evidence and suitable structure: relevant, verifiable, dated and attributed facts. One factorial experiment it reviews, with 252,000 trials, found effects for explicit prices and recent dates, and weak effects from formatting alone.

Outside the survey, some observational data points the same way, with limits:

- **Third-party sources.** [Muck Rack’s May 2026 report](https://media.muckrack.com/documents/What_Is_AI_Reading__May_2026.pdf) (vendor data from more than 25 million cited links) puts about 84% of citations in its “earned” group. That group is broad: it includes journalism, academic and government sources, Wikipedia, social platforms and third-party company content. The survey notes that this kind of finding doesn’t show that coverage causes a recommendation.
- **Freshness.** [Ahrefs’ study of about 17 million citations](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/) (vendor data) found AI assistants cited URLs about 25.7% newer than organic results on average, with Google AI Overviews as the exception. The study doesn’t establish a 90-day window, and the survey reports that recency depends on query intent.

## What Is Weak or Contradicted

The survey says “generic heuristics transfer poorly.” In C-SEO Bench, only 3 of 54 method and domain combinations were significantly positive in the main experiment. It also found that gains shrink as more sites adopt the same method, approaching a zero-sum game. A test by one site can’t predict what happens once everyone does it.

**llms.txt for citations.** [SE Ranking’s study of nearly 300,000 domains](https://seranking.com/blog/llms-txt/) (vendor data) found no measurable link to citations. Google’s [AI optimization guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) says you don’t need AI text files for Google Search. Don’t expect citations from it, although it may still give agents a useful map.

**Keyword stuffing.** It scored about 8% below baseline in the Princeton study.

**Schema as a universal lift.** The evidence doesn’t support it. [An Ahrefs test of 1,885 pages](https://ahrefs.com/blog/schema-ai-citations/) (vendor data) found no major uplift, though it pooled all types and covered only pages already heavily cited.

**AI-specific tags and files.** Google says you don’t need new machine-readable files or markup for its AI features.

## What the Survey Doesn’t Cover

The survey studies generative answers. It doesn’t test AI agents that act for a user, or protocols like the Model Context Protocol (MCP), which gives agents a standard way to reach tools and data.

MCP has real developer adoption. [Anthropic reported](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) more than 10,000 active public servers and 97 million monthly SDK downloads when it donated the protocol to the Linux Foundation’s Agentic AI Foundation in December 2025. That doesn’t show that having a server earns customers, and agents can also read ordinary pages. Its effect on business results is unmeasured, so test it only if agents could complete a real task on your site.

## Why Controls Change the Answer

One study in the survey shows how easily a good story forms. Watanabe and Nakayashiki analyzed logs from a website where some pages received an answer-engine optimization change. Total ChatGPT referrals rose by a factor of 5.7, which would make a fine case study.

![Growth in ChatGPT referrals on one website: all referrals rose 5.7 times, while untreated pages alone rose 3.5 times.](https://llmagnet.com/images/blog/charts/geo45-referral-growth.png)

*Source: Watanabe and Nakayashiki (2026), as reported in the July 2026 survey. One website.*

But pages that received no change had already grown 3.5 times, because ChatGPT itself was growing. A controlled time-series analysis put the additional effect at 1.82 times, with a 95% interval of 1.31 to 2.54. A conservative placebo test came back at p = 0.16, so the survey calls the effect suggestive and not causally established. A real gain may exist, and it is far smaller than the raw number.

## How to Prioritize

Ranked by the strength of the evidence:

**Do first.** Make key pages technically retrievable, and make each one answer a specific question directly. The survey’s best-supported recommendation is a relevant, comprehensive, verifiable, clearly structured and retrievable page.

**Do next.** Add relevant, sourced and dated facts. Update pages for time-sensitive queries when the facts change. Work on being accurately covered by the third-party sites AI answers already cite, knowing that this evidence is observational.

**Test on a few pages.** Schema types, question headers, restructured answers and an MCP server. Keep an untreated group of pages to compare against.

**Don’t expect results from.** llms.txt as a citation tactic, AI-specific tags, keyword stuffing, or any rule taken from one platform and applied to all.

## Measuring What Can Be Measured

Results vary from run to run, so a single check tells you little. The survey reports three useful numbers:

- Across four engines and 45 days, daily source overlap was only about 0.34 to 0.42 on a Jaccard scale, where 1 means identical.
- Page overlap in AI Overviews over two months was 18%, against 45% for organic Google, across 4,706 queries in the US and Germany.
- In one setup, 57.8% of ChatGPT repetitions didn’t use web search at all, so a share of citations that ignores those runs overstates visibility.

For each page you change, keep comparable pages unchanged. Run each prompt seven to eight times as a starting point, across paraphrases, dates and engines, and report the runs that returned no citation.

The survey also says to split visibility into activation, retrieval and citation given retrieval, so a gain at one stage can’t hide a loss at another. Plan the workload before you start. Twenty buyer questions with three paraphrases, eight runs and four engines come to 1,920 runs per measurement. Ten questions on two engines come to 480. If that is still too many, cut questions and keep the repeats, since a single run can’t show you the noise.

Add a placebo as well: an edit of similar length that doesn’t change the substance. If the placebo pages move as much as the treated ones, you can’t credit the content change. Keep the prompts, raw responses, dates and locale, so you can check the result later.

## How to Read the Next GEO Claim

1. **Which stage was measured?** A citation gain inside a fixed context says nothing about retrieval.
2. **Which engine, mode and date?** Commercial engines differ and change, so a result carries a timestamp.
3. **Was there a control?** Without unchanged pages, you can’t separate the edit from platform growth.
4. **How many runs, and were empty runs kept?** Dropping runs with no citation inflates the result.
5. **Who ran it, and is it reviewed?** Vendor data can be right, but you should know who sells what.
6. **Absolute or relative?** A 40% relative gain on a small base can be a small absolute change.

## The Practical Bottom Line

The evidence supports less than the industry claims, and most of what it supports looks like good information work: answer the question, cite real sources, stay current and make the page easy to retrieve. Be wary of anyone who promises a percentage.

LLMagnet measures AI visibility across seven engines and has an agent that prepares fixes for your approval. To see where you stand today, start with the [free LLMagnet audit](https://ai-visibility.llmagnet.com), and then repeat your measurements before you trust a change.

## FAQ

### What did the survey of GEO research find?

Reviewing 45 studies from November 2023 to July 2026, it found no technique with a stable, long-term, cross-platform causal effect on organic discoverability, clicks or conversions.

### Were all 45 studies peer-reviewed?

No. The set mixes peer-reviewed papers, accepted work, workshop papers and preprints, and the survey gives them different weight. The survey itself was posted as an arXiv preprint in July 2026.

### Does that mean GEO doesn’t work?

No. It found good evidence that content already retrieved can be made more likely to be cited. What’s missing is evidence that those gains last and carry across engines, or lead to traffic.

### What does the 40% GEO boost apply to?

One metric, Position-Adjusted Word Count, in a fixed benchmark where the source was already in the model’s context. The survey rejects it as a general claim about visibility.

### Why are AI visibility results so unstable?

Engines cite different sources and vary between runs. One audit in the survey found AI Overview pages overlapped 18% over two months, against 45% for organic Google.

### Why do control pages matter in GEO tests?

Because AI platforms are growing on their own. In one study, ChatGPT referrals rose 5.7 times overall, but untreated pages had already grown 3.5 times. The controlled estimate of the added effect was 1.82 times, and the survey calls it suggestive.

### How should I prioritize GEO work given this evidence?

Start with technically retrievable pages that answer specific questions, then add sourced, dated facts. Test schema and restructuring on a few pages against unchanged ones, and don’t expect citations from llms.txt.