Five GEO Strategies Tested on Live Pages: CORE Wins Top-1, but Engines Agree on Only Half the Brands

Five GEO Strategies Tested on Live Pages: CORE Wins Top-1, but Engines Agree on Only Half the Brands

Summary

  • LIVE-GEO tested five GEO strategies on 80 live product pages across ChatGPT, Gemini and Grok; CORE led with a Top-1 rate of 66.3%–76.3%.
  • Engines agreed on the same top product for only 48.8%–66.3% of brands, so single-engine citation reports measure one surface, not market-wide visibility.
  • Reversing candidate URL order moved ChatGPT's CORE Top-1 rate from 73.8% to 93.8%, making single-run, fixed-order wins unreliable.
  • Statistics Addition produced all 80 flagged descriptions in the numerical-faithfulness check, despite being a widely recommended tactic.
  • Demand per-engine reporting, reverse-order retests, verification for added statistics and a structural audit before claiming any GEO win.

A new preprint tests five GEO strategies on real, live webpages rather than text passed directly into an API. The results support one widely used strategy and complicate the way agencies report citation wins. Teams building a GEO measurement practice should read the caveats as carefully as the headline numbers.

The study is LIVE-GEO (arXiv:2611.01126), a 12-page preprint posted 31 October 2026. It has not been peer-reviewed. All figures in this article come from that preprint unless another source is stated.

What LIVE-GEO Measured and How

Most GEO evaluations test on raw text or controlled retrieval pipelines. LIVE-GEO published 80 GEO-rewritten product pages as live websites and then measured how generative engines treated them — a design that more closely reflects real deployment conditions.

The benchmark covered:

  • Brands: 80 fictional brands across 18 product families
  • Strategies: CORE, C-SEO, E-GEO, AgenticGEO and Statistics Addition
  • Engines: ChatGPT with GPT-6, Gemini 3.6 Thinking and Grok in Expert mode
  • Evaluators: Four direct-text API evaluators plus one local evaluator
  • Price controls: Frozen to ±0.2% variance, to isolate content effects from price effects

The central metric is the Top-1 rate: the share of brands for which a given strategy produced the engine's top-ranked product recommendation when the engine received candidate URLs.

Finding 1: CORE Leads, but the Margin Is Engine-Dependent

CORE produced the highest Top-1 rate of the five strategies across all three engines, ranging from 66.3% to 76.3%. The other four strategies varied in relative performance depending on which engine was tested.

That range matters. There is no single CORE result. The 10-point spread across engines already signals that performance is not uniform, and it is one reason why reporting a single aggregated figure understates the uncertainty in any real deployment.

The preprint describes CORE as improving page organisation and strengthening argumentative structure rather than adding supporting evidence. That design pattern — structure and argument over added content — is what the winning strategy has in common with structural enrichment tactics evaluated elsewhere. The SAGEO Arena preprint found that enriching structural fields such as schema, headings and meta descriptions lifted retrieval by approximately 22%, while body-text optimisation alone was frequently ineffective.

Finding 2: Engines Disagreed on the Top Product for up to Half of All Brands

This is the finding that most directly challenges the "we got you cited" claim.

The three engines selected the same top product for only 48.8% to 66.3% of the 80 brands. For between a third and a half of all brands, the top-ranked product changed depending on which engine was asked.

An independent study by Qwairy reached a similar conclusion from a different angle: on the same prompt, two AI surfaces cite the same domain only 4% to 19% of the time. Even two Google products — AI Overviews and AI Mode — overlapped on only 19% of citations.

The operational consequence is direct. A citation in ChatGPT is not evidence of citation in Gemini or Grok. Programmes using a single-engine report to demonstrate AI visibility are measuring one surface, not market-wide presence. Effective AI visibility measurement tracks each engine separately and treats cross-engine agreement as its own signal.

Flying Blind on AI Search?

Finding 3: URL Order Moved ChatGPT's Top-1 Rate by 20 Points

The most operationally significant finding concerns measurement stability, not strategy ranking.

ChatGPT's Top-1 rate for CORE was 73.8% under one fixed URL order. Under the reversed order, it was 93.8%. The benchmark's randomised order produced 82.5%, sitting between the two. The same strategy, the same content, a different order of candidate URLs — and a 20-point swing in outcome.

This is not an isolated property of the LIVE-GEO design. Research on LLM listwise ranking shows that reordering a candidate set can promote a bottom-ranked item into the top-5 for up to 57% of target items. Position sensitivity is a measurable, reproducible phenomenon in large language model ranking behaviour.

The direct implication for reporting: a Top-1 result from a single prompt run, with a fixed candidate order, is not a stable finding. It may reflect order bias as much as content quality.

Finding 4: Statistics Addition Produced Every Flagged Description in the Test

The LIVE-GEO numerical-faithfulness check assessed whether rewritten descriptions preserved the factual accuracy of the original product data. The verifier flagged 86 of 400 rewritten descriptions.

All 80 descriptions rewritten using the Statistics Addition strategy were flagged. No other strategy produced a comparable concentration of flags. The preprint's own verifier scored 90.7% agreement with reference labels on 194 numerical cases, so the finding is directional rather than exact — but the pattern is clear.

This matters because Statistics Addition is among the most commonly recommended GEO tactics. The original GEO paper by Aggarwal et al. (KDD 2024), covered in detail at Synscribe's GEO methodology overview, reported visibility uplifts of up to 40% when the tactic was applied. That result drove widespread adoption.

The replication picture is more complicated. A critical survey of 45 GEO studies found that Statistics Addition did not survive realistic retrieval conditions. The C-SEO Bench paper (NeurIPS 2025) went further: under realistic retrieval settings, Statistics Addition actively lowered ranking in 19 of 24 tested configurations.

Adding numbers to a page is easy to recommend and easy to deliver. The risk it carries — factual inaccuracy introduced at scale — is harder to detect without a verification step.

Four Rules That Follow From These Findings

The four findings above translate into concrete requirements for evaluating any GEO programme or tool.

1. Demand Per-Engine Reporting, Not a Combined Score

Cross-engine agreement in LIVE-GEO reached only 48.8% to 66.3%, and the Qwairy data puts overlap between 4% and 19%. Any report that presents citations as one number, without naming the engine, is aggregating results that should not be combined. Require mentions, share of voice and citation counts to be broken out by engine — ChatGPT, Gemini, Grok, Perplexity and any other surface being tracked.

Some AI visibility tools restrict entry-tier plans to a single engine, which limits the data available at lower price points.

2. Run the Same Prompts Twice, in Reverse Order, Before Claiming a Win

A 20-point swing in Top-1 rate from a URL order change is large enough to render a single-run result unreliable. Before accepting any reported gain, require the same prompt set to be re-run with the candidate order reversed. If the top-ranked result changes, the finding is not stable. LLM position sensitivity research confirms that permutation stability is a usable audit signal, not a theoretical concern.

3. Scrutinise Any Strategy Built Around Adding Statistics

Statistics Addition is the most-recommended tactic in published GEO guidance. It is also the only strategy in LIVE-GEO where every rewritten description was flagged for numerical-faithfulness issues.

Before accepting a deliverable built on this tactic, ask two questions: are the added numbers sourced, and is there a QA step before publication? A citation gain that introduces factual errors into a product page carries a compliance and reputational cost that outweighs the visibility gain.

4. Prioritise Structure and Argument Over Added Evidence

The strategy with the highest Top-1 rate in LIVE-GEO improved page organisation and strengthened argumentative structure rather than appending supporting evidence. A consistent pattern across recent GEO research points the same way: structural signals — schema markup, heading hierarchy, self-contained answer capsules under H2s — produce more durable retrieval gains than content additions. That research also finds that pages lacking a standalone, citable passage are excluded regardless of topic quality.

The structural work is not optional.

What LIVE-GEO Proves, and What It Leaves Open

LIVE-GEO is a preprint. It used fictional brands, frozen prices and a small product set. It has not been peer-reviewed. The figures it produces — CORE's Top-1 range, the engine agreement percentages, the order-sensitivity numbers — are findings from one controlled experiment, not constants that will hold across every category and engine update.

What the study does establish, directionally and with corroboration from other recent work, is this:

  • Engine agreement on the same top pick is low enough that single-surface reporting is structurally incomplete.
  • ChatGPT's sensitivity to candidate order is large enough to invalidate single-run, fixed-order measurement.
  • The most widely recommended GEO tactic — Statistics Addition — carries a documented accuracy risk that the industry has not consistently addressed.
  • The strategies that performed best improved structure and argumentation rather than adding volume or evidence.

None of these findings support the claim that CORE is proven at scale, and none of them should be read as a ceiling or a floor for what any specific programme can achieve. They are constraints on how results should be measured and reported.

Teams running GEO programmes should treat four things as minimum requirements for any reporting they receive or produce:

  • Per-engine breakdowns. Mentions, share of voice and citations split by ChatGPT, Gemini, Grok and any other surface being tracked.
  • Order-stability tests. The same prompt set re-run with the candidate order reversed before any gain is reported.
  • Statistics verification. A documented QA step for every number added to a page.
  • A structural audit. Schema, heading hierarchy and self-contained answer capsules checked before a citation win is claimed.

Ready to Win on Every Engine?

If your team needs reporting that survives this level of scrutiny — per-engine breakdowns, order-stability retests, structural audits — talk to Synscribe about building the measurement programme behind it.

Frequently Asked Questions About the LIVE-GEO Study

What is LIVE-GEO?

A 12-page preprint (arXiv:2611.01126), not yet peer-reviewed, that tested five GEO strategies on 80 live product websites across ChatGPT, Gemini and Grok. Because it used live pages and frozen prices rather than text piped into an API, it reflects real deployment more closely than earlier benchmarks. Treat its figures as directional.

What is the CORE strategy in GEO?

CORE is a GEO strategy that improves page organisation and strengthens argumentative structure instead of adding new supporting evidence. In LIVE-GEO it produced the highest Top-1 rate of the five strategies across all three engines, ranging from 66.3% to 76.3%. Structural work — clear headings, self-contained answer blocks, schema — is the closest thing to a proven tactic.

Does LIVE-GEO prove CORE works at scale?

No. It is a non-peer-reviewed preprint using fictional brands, frozen prices and a small 80-brand product set, and its 66.3% to 76.3% range across engines shows the result is not uniform. It offers directional evidence, corroborated elsewhere, that structure beats added evidence — not a guarantee for your category or the next model update.

Why do ChatGPT, Gemini and Grok recommend different products?

Because each engine retrieves, weights and ranks sources differently. In LIVE-GEO the three engines picked the same top product for only 48.8% to 66.3% of brands, and Qwairy found two AI surfaces cite the same domain just 4% to 19% of the time. A citation in one engine is not evidence of visibility in another.

Is Statistics Addition still a good GEO tactic?

Treat it with caution. All 80 descriptions rewritten with Statistics Addition in LIVE-GEO were flagged for numerical-faithfulness issues, and a critical survey of 45 GEO studies found the tactic failed under realistic retrieval conditions. If you use it, source every number, add a QA step before publishing, and accept the accuracy risk.

How do I check whether a GEO citation win is real?

Re-run the same prompt set with the candidate order reversed. In LIVE-GEO, ChatGPT's Top-1 rate for CORE moved from 73.8% to 93.8% purely from reversing URL order. If the top-ranked product changes, the win reflects order bias rather than content quality. Report stability alongside the gain.

How many AI engines should I track for GEO reporting?

At minimum three: ChatGPT, Gemini, and Grok or Perplexity where budget allows. Because cross-engine agreement sits between 48.8% and 66.3%, a single-engine report measures one surface, not the market. Break out mentions, share of voice and citations by engine, and treat cross-engine agreement as its own metric.

How should agencies report GEO results to clients?

Per engine, never as one combined score. Break out mentions, share of voice and citation counts by surface, disclose the prompt set and candidate order used, and show an order-reversal retest alongside any claimed gain. Aggregate scores hide cross-engine disagreement and order sensitivity — the two failure modes LIVE-GEO documents most clearly.

What should I optimise first for AI visibility?

Structure and argument, not added volume. The top-performing strategy in LIVE-GEO improved page organisation and argumentation rather than adding evidence. Structural signals — schema markup, heading hierarchy, self-contained answer capsules under each H2 — consistently outperform content additions, and pages lacking a standalone citable passage are often excluded.

Tags:
Published on November 19, 2026

Dominate ChatGPT and Google Search

Synscribe helps B2B companies with SEO & GEO using programmatic SEO approach. Book a call to find out how we help you win.