Two GEO levers survive controlled testing: being retrieved high up by the engine, and publishing a page whose title and one paragraph answer the question asked. Schema, llms.txt and added statistics have been tested, with no gain under those test conditions. Reddit, Wikipedia, YouTube and PR remain correlations: as of 6 November 2026, no published controlled test could be found that measures what a brand's action changes there.
That has a budget consequence. Fund the first group, stop the last, and treat everything else as a bet with a cap, a control group and an end date.
Even the two levers that hold come with a caveat: they concern citations, not revenue. Olivier Martinez's review (preprint, July 2026, 45 studies) finds no technique with a stable, longitudinal, cross-platform causal effect on organic discoverability, and rates confidence in the link from citation scores to clicks, conversions or revenue as "very low".
What is the evidence behind each GEO tactic?
Four grades, set by the type of test rather than the size of the sample:
- Strong evidence. The effect was reproduced under controlled conditions on several models. That does not guarantee a lasting effect on a live engine.
- Correlation only. Observed at scale, never isolated by an intervention with a control.
- No evidence. Widely repeated advice, with no test found that isolates it.
- Failed. Tested with no gain under the test's conditions, or dismissed by the engine's own publisher for its own system.
| Tactic | Grade | Best evidence available | What remains unknown |
|---|---|---|---|
| Being retrieved high by the engine | Strong evidence | Leading the context outweighs any rewrite (C-SEO Bench, 2025). ChatGPT cites a source in 60% of answers when read first, 43% when read last; Gemini 57% to 59% throughout (Chen and Koudas, Oct 2026) | How to earn that rank: the signals in the internal ranking are not observed |
| Title and paragraph matched to the question | Strong evidence | Off-topic is a near-knockout on 6 models, 252,000 trials (Sprinklr, May 2026). Observed within a site: +15.5% citations per SD for the best paragraph, +14.1% for the title (Chen and Koudas) | The effect of an isolated rewrite on a real site |
| Google ranking | Correlation only | Cited pages found in the top 10: 0.1% to 0.6% (Chen and Koudas), 21% to 28% (Minuttia, Aug 2026, excluding ChatGPT), 37.9% on AI Overviews (Ahrefs, Mar 2026) | Which figure holds for your market and engine |
| Brand mentions, PR | Correlation only | Correlation with ChatGPT visibility: 0.66 across 75,000 brands, 0.15 across the top 50 domains (Ahrefs, Dec 2025 and June 2025). 84% "earned" citations (Muck Rack, a PR vendor, May 2026) | What a campaign does: no controlled test found |
| YouTube | Correlation only | YouTube mentions: about 0.737, the strongest correlate (Ahrefs, Dec 2025). Share: 24.5% on AI Overviews, 2.3% on ChatGPT (top 50, US, Nov 2026) | What publishing does: one public test, no control |
| Wikipedia | Correlation only | 10.1% share on ChatGPT, 4.4% on AI Overviews (Ahrefs, Nov 2026). About 55%, then under 20%, of ChatGPT answers in 2025 (Semrush) | Whether a brand page changes answers |
| Correlation only | Close to 60%, then about 10%, of ChatGPT answers from August to mid-September 2025 (Semrush). At equal rank, ChatGPT rarely cites threads it retrieved (Chen and Koudas) | What active participation does | |
| Backlinks | Correlation only | Domain Rating: 0.27 to 0.33 (Ahrefs, Dec 2025). Referring sites: +22.7% citations per SD within a site, mostly on Gemini (Chen and Koudas) | Direction of cause: links were measured after citations |
| Freshness | Correlation only | Cited content 25.7% fresher than organic results (Ahrefs, Dec 2025, 17M citations). Age: -5.1% per SD, not selected consistently (Chen and Koudas) | What a new date does without new content |
| Listicles | Correlation only | 34% to 47% of cited "alternatives to X" listicles are self-promotional, by engine (Minuttia, software, Aug 2026) | Whether a citation becomes a recommendation of the brand |
| Answer at the top of the page | No evidence | The best-matching paragraph sits at a median 45% of the way down the page (Chen and Koudas) | No test isolates placement |
| Chunking | No evidence | Google: not required. No independent test found | The effect outside Google |
| Added statistics and quotations | Failed | 3 significant cases out of 54 in C-SEO Bench (2025), none for these two methods; statistics lowered rank in 19 of 24 settings | The effect of truly original data |
| Schema | Failed | 1,885 pages, about 4,000 controls, 30 days: -4.6% AI Overviews, +2.4% AI Mode, +2.2% ChatGPT, the last two indistinguishable from zero (Ahrefs, May 2026) | First citation, longer windows, individual types, schema injected with JavaScript |
| llms.txt | Failed | Google Search does not use it. 97% of files got no requests in May 2026 (Ahrefs). No correlation with citations across nearly 300,000 domains (SE Ranking, Oct 2026) | Use by other services or agents |
Full sources at the end of the article. Domain shares do not compare across publishers: Ahrefs reports against the top 50 sources, Semrush against answers.
Why does the "+40%" GEO figure not apply to your site?
Because it measures a share of text inside an imposed context, not the chance of being found. In the founding paper by Aggarwal and co-authors (KDD 2024), adding quotations moved a source from 19.3 to 27.2 on position-adjusted word count, which works out to +40.9%, on a simulated engine built on GPT-3.5 turbo with five sources already selected. Its "real world" test on Perplexity does not test retrieval: the source text was supplied as file uploads, answers came from those files alone, and the sample was 200 queries.
The same paper holds a less convenient result. Citing sources gained +115.1% for a site ranked fifth, while the top-ranked site lost 30.3% on average. If you already lead, this recipe works against you in the very study that proposes it.
C-SEO Bench (NeurIPS 2025) retested these methods on citation rank, across more than 1,900 English queries: 3 significant cases out of 54, with gains shrinking as competitors adopted the method. The authors call it a zero-sum game. A rewrite your rivals can copy carries an expected gain trending to zero. In SAGEO Arena's reconstructed search environment, rewriting body text alone even pushed documents out of retrieval, notably when common words gave way to rare vocabulary. Martinez's review rejects the 40% as a general claim.
The closest test to real conditions appeared on 31 October 2026. LIVE-GEO published 80 fictional brands on the live web and queried ChatGPT, Gemini and Grok: part of the aggregate gain survives, most brand-level gains do not, and the engines agreed on the top product for only 48.8% to 66.3% of brands. Its verifier flagged all 80 descriptions rewritten with statistics, because the method allows hypothetical data. Asking a tool to add a number so you get cited can end with a false number on your site.
The retouch is what failed. Martinez rates the evidence on extractable material as "moderate to strong", and in Sprinklr's test a missing price can be enough to rule out a product review. Original data has not been tested on live engines.
Why are schema and llms.txt graded as failed?
Because they are among the few pieces of GEO advice that have been tested, and the tests show no gain. The Ahrefs study has one major limit: its pages were already heavily cited, so it says nothing about a first citation.
Chen and Koudas (Toronto, October 2026, preprint, data from a monitoring platform) explain why so many studies claim otherwise. Across sites, schema goes with +7.5% citations and an FAQ block with +18.1%. Between pages of the same site, those gaps become -5.3% and -3.4%, neither significant. Sites that take care of their markup take care of everything else too. Any claim that cited pages are several times likelier to carry JSON-LD falls into this trap.
One study points the other way: SAGEO Arena (February 2026) finds that structural fields taken together (titles, metadata, markup) improve retrieval in a pipeline the authors built themselves. It does not isolate schema. Keep whatever serves rich results, and stop counting it as a GEO win.
On llms.txt, Google's guide says the file neither helps nor hurts in Google Search and that other services may use it. It costs little, so there is no need to delete it. Just stop counting it in your reporting.
Should the answer go first, and should pages be chunked?
For readers, probably. For engines, nothing proves it yet. In Chen and Koudas, the paragraph that best matches the questions sits in the top quarter of the page for 32% of cited pages and in the bottom quarter for 28%. Their conclusion: engines find it wherever it is. First-paragraph match carries +5.9% per SD, against +15.5% for the best paragraph. An answer needs to exist somewhere on the page more than it needs to sit at the top.
The nearest live test comes from Watanabe and Nakayashiki (preprint, June 2026), affiliated with the company behind the site they studied. Pages moved to question-form titles and two-to-three-sentence summaries. ChatGPT referrals to the domain grew 5.7x, but untouched pages grew 3.5x. The estimated effect is 1.82x (95% interval 1.31 to 2.54), with a placebo test at p = 0.16: the authors call it suggestive, not conclusive. The bundle also included URL consolidation and new pages, so structure alone is not isolated.
On chunking, Google writes that there is no requirement to break content into small pieces. Headings, lists and tables show no significant within-site effect in Chen and Koudas, with one nuance: at equal rank, pages with more headings are cited more often. This article opens each section with its answer and ends with an FAQ because that reads better, not because a study shows it earns citations.
Does organic ranking still matter?
Yes, as long as you watch the right one: the rank in your SEO tool is not the engine's. When C-SEO Bench concludes that traditional SEO remains decisive, it is measuring a document's position in the model's context, not a Google rank. The GEO and SEO guide covers what the two share.
The rank that matters most is the engine's own. In Chen and Koudas's replay, ChatGPT cites 43.8% of the pages it ranks first and 3.5% of those ranked 11 to 20, and its retrieval explains 32% of the variation between pages of a site, against 2.8% for page properties. It rewrites the question on every answer: between two runs on the same day, query overlap is only 2.7% (Jaccard index).
The table's figures on the link with Google collide because they measure different things: US commercial questions in Chen and Koudas, software questions in Minuttia (August 2026), 863,000 SERPs in Ahrefs (March 2026), which halves its earlier "76%" without confirming why: Ahrefs points to better parsing, the move to Gemini 3 and fan-out. Google writes that its AI features are rooted in its core ranking systems, which says nothing about ChatGPT. SEO remains the foundation on Google's surfaces. Elsewhere, your Google rank is not a reliable proxy until you have checked it engine by engine.
What are the correlations in the table worth?
They describe where engines read today, not what your own action would change there.
- Mentions and PR. Ahrefs' 0.66 applies to established brands (Domain Rating above 40), and the 0.15 dates from June 2025. The "84% earned" figure comes from a PR software vendor and covers far more than press: journalism accounts for 27%. Chen and Koudas found the reverse on US home internet, with 64% of ChatGPT's citations going to the providers' own sites. Sector changes everything.
- Your size changes the route. In the same study, 46% of ChatGPT's mentions of a brand absent from the question have no cited page behind them: a known brand gets named from memory. A niche B2B vendor gets named mainly when a cited third-party page names it (on ChatGPT, odds up 39.7x, against 4.2x in the consumer market). PR therefore has a better case for a niche player, as a hypothesis: the B2B sample is 55 questions.
- YouTube. One public test exists: 25 AI-generated videos on a new German-language channel, with share of voice clearly up on five of six engines in about ten days (published by Otterly, a tool vendor, April 2026). One brand, no control, and titles copied from the tracked prompts. Read it as a lead.
- Wikipedia. No causal test found, and this lever cannot be bought cleanly. English Wikipedia strongly discourages conflict-of-interest editors from editing affected articles directly and requires paid editors to disclose who pays them.
- Reddit. A further, provisional fall came in August 2026: from 3.8% to 0.5% of ChatGPT citations according to Promptwatch, which Semrush reports as saying it cannot rule out a problem in its own data. In Chen and Koudas, Reddit accounts for 2.2% of ChatGPT's citations and 13.1% of Gemini's.
- Backlinks. High-authority sites have more cited pages, not more citations per page (Chen and Koudas). The per-page effect is carried by Gemini (+33.7% per SD, against +3.7% for ChatGPT).
- Freshness. ChatGPT cites URLs 393 to 458 days newer than organic results (Ahrefs) and adds the year to 23% of its queries, against 2% for Gemini, which offers a plausible mechanism. Martinez notes that a recent date helps on a time-sensitive query, not necessarily on a stable definition. No test found supports the advice to refresh on a fixed schedule.
- Listicles. Quoted by Search Engine Roundtable (March 2026), Lily Ray describes a site with more than 200 self-promotional lists falling since January 2026, and she tracks about 30 similar cases. The piece cites no confirmation from Google. A citation gain may cost organic traffic.
What has nobody tested yet?
Almost everything a GEO budget pays for in practice. These gaps reflect public research as of 6 November 2026, and a private test may exist.
- An off-site intervention. No published controlled study of a PR campaign, a Reddit presence, a Wikipedia page or a YouTube program.
- The first citation. The schema test covers pages already cited. The question for an absent brand stays open.
- Markets outside the US. C-SEO Bench and Chen and Koudas use English, and Ahrefs' domain rankings are US data. No study in the table isolates France.
- Time. Controlled tests mostly run from a single snapshot to 30 days. In Chen and Koudas, ten weeks on, only 14% (ChatGPT) to 29% (Gemini) of the variation between pages remains shared.
- The link to revenue. The weakest link in the chain, covered in the piece on incrementality.
What should you do on Monday?
Take out of your reporting what does not hold, then isolate one serious test.
- Remove failed tactics from your KPIs. llms.txt and schema "for AI" no longer count as GEO wins. Stop inserting numbers into copy and building a page per fan-out variation: when made mainly to manipulate Google Search's AI responses, those pages violate its scaled content abuse policy.
- Rewrite one title and one paragraph on twenty pages. Pick the pages behind your priority prompts and use the question's own terms. No need to move the paragraph up the page.
- Check whether engines already know you. Ask your category questions several times without web search, where the engine allows it. If you get named, check that what is said is accurate. If not, third-party pages covering your category become the priority.
- Book every remaining line as a bet. Reddit, YouTube, PR: one hypothesis, one cap, one end date (budget frame here). Ask your vendors which grade in this table each recommendation reaches, and which control group measured it.
- Test against a control. Form pairs of comparable topics, draw at random which side you change, compare over the same period, and set the decision threshold before you start. A control prompt inside a worked topic would benefit from the work.
As an illustration, detecting a move from 30% to 40% presence takes about 350 observations per group. Twenty prompts observed twenty times each, across engines and runs, give 400 observations, yet repeats of one prompt are not independent, so treat that as a floor, and the unit of analysis remains the topic. The detail is in the piece on prompt counts.
Illustrative calculation: 95% confidence, 80% power, observations assumed independent.
A test like this means repeating the same prompts across engines and over time, which a spreadsheet tracks poorly. Nanga's GEO module tracks your prompts run after run, engine by engine (weekly by default, daily as an option), filters them by tag, and places completed actions as markers on the trend chart. A marker shows a coincidence; only the control group lets you talk about cause.
To see this protocol applied to your own prompts: request a demo.
Frequently asked questions
What counts as a proven GEO tactic?
One whose effect was measured in a controlled test, not inferred from a correlation. According to Olivier Martinez's July 2026 review of 45 studies, no technique shows a stable causal effect across engines. Relevance and position in the model's context, by contrast, are rated as well established.
Does schema markup improve citations in AI engines?
No, on the best test available. Ahrefs (1,885 pages, about 4,000 controls) measured -4.6% on AI Overviews and no detectable change on AI Mode or ChatGPT after 30 days. Those pages were already heavily cited, so the effect on a first citation remains untested.
Should a brand publish an llms.txt file?
Not to gain visibility. Google says Google Search does not use it, SE Ranking found no correlation with citations across nearly 300,000 domains, and Ahrefs found that 97% of files received no requests in May 2026. Other services may still read it.
Does adding statistics help content get cited by AI?
Not as a rewrite. In C-SEO Bench (NeurIPS 2025) the method lowered rank in 19 of 24 settings. Verifiable, original data still serves readers, but no test has isolated its effect on live engines.
How can a team test a GEO tactic itself?
Use a control group: form pairs of comparable topics, draw at random which one you change, leave the other alone, and compare over the same period. Before-and-after is not enough: on the site Watanabe and Nakayashiki studied, ChatGPT referrals to untouched pages still grew 3.5x.
Sources:
- Olivier Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), preprint, July 2026.
- Aggarwal et al., GEO: Generative Engine Optimization, KDD, August 2024.
- Puerto et al., C-SEO Bench: Does Conversational SEO Work?, NeurIPS 2025.
- Kong et al., Do GEO Gains Survive the Live Web?, preprint, October 2026.
- Vishwakarma, Kumar and Jamidar (Sprinklr), What Gets Cited: Competitive GEO in AI Answer Engines, preprint, May 2026.
- Chen and Koudas, What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), preprint, October 2026.
- Kim et al., SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization, preprint, February 2026.
- Watanabe and Nakayashiki, Disentangling Answer Engine Optimization from Platform Growth, preprint, June 2026.
- Google Search Central, Optimizing your website for generative AI features on Google Search, accessed 6 November 2026.
- Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved., May 2026.
- Ahrefs, We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read, June 2026.
- SE Ranking, Does LLMs.txt impact your AI visibility and citations? No, according to research, October 2026.
- Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied), December 2025.
- Ahrefs, Google Seems More Biased Towards Big Brands Than ChatGPT and Perplexity, July 2025.
- Ahrefs, Update: 38% of AI Overview Citations Pull From The Top 10, March 2026.
- Ahrefs, Fresh Content: Why Publish Dates Make or Break Rankings and AI Visibility, December 2025.
- Ahrefs, The 50 Most-Cited Websites in Google AI Overviews (November 2026), November 2026.
- Ahrefs, The 50 Most-Cited Websites in ChatGPT (November 2026), November 2026.
- Minuttia, Self-Promotional Content Data Study, October 2026.
- Semrush, The Most-Cited Domains in AI: A 3-Month Study, November 2025.
- Semrush, Reddit's citations in ChatGPT fall from 3.8% to 0.5%, August 2026.
- Muck Rack, Earned media still drives 84% of AI citations. Here's what that means for PR., May 2026.
- Otterly, 25 AI-Generated YouTube Videos, One Week, and What They Did to AI Search Visibility, April 2026.
- Search Engine Roundtable, Is Google Hitting Self-Promotional Listicles Hard Again?, March 2026.
- Wikipedia, Conflict of interest, accessed 6 November 2026.