In short. The GEO score paper supports using a deterministic content score as a manipulation-resistant quality filter, with citation outcomes evaluated separately.
Definition: The GEO score paper studies a deterministic content score for generative engine optimization (GEO), the study of website visibility in answer engines. It validates the score as a manipulation-resistant quality filter and examines the limits of citation prediction.
Key facts
- The arXiv record identifies Elisha Bajemon and André-Louis Rochet as the authors and records submission on 7 September 2026. Both authors list TW3 Partners (Citead) as their affiliation in the paper.
- The earlier GEO paper, first submitted in November 2023, introduces GEO-Bench, a benchmark pairing questions with relevant web sources. The score study uses sources from this benchmark.
- The authors' reproducibility package provides stored feature captures, citation outcome records and analysis scripts for rerunning the reported analyses offline.
How does the score work?
In the paper's scoring specification, the input is page text in Markdown. Eleven components use regular expressions and statistics to measure properties such as statistic density and repetition. Scoring requires no network access or language model.
The content components pass through a square-root transformation before weighted aggregation. This reduces the marginal reward from repeatedly increasing a single feature. The page-level formula then blends content with freshness, meaning signals of publication and update recency. In plain-text notation, the formula is:
GEO(x) = 0.92 * GEO_content(x) + 0.08 * F(x)
Here, GEO_content(x) is the content score and F(x) is the freshness component. Word-count caps follow this blend: a page under 100 words has a ceiling of 35 points. These are properties of the scoring formula, with the citation experiments evaluated separately.
What do the engine experiments find?
The outcome experiment gives an engine a question and five sources, then compares edits to the target source against neutral filler of matching length. Its visibility measure allocates position-weighted answer words to the sources cited by those sentences.
Across ten engine families, the study reports no positive citation effect in the published direction from the tested quotation, statistics and cited-source interventions associated with the earlier GEO work. Recalibrating against the newly measured effects removes the score's components responsive to those interventions.
The experiment evaluates a controlled setting with supplied sources and a fixed citation prompt; its scope excludes open-web retrieval and competition. For a website audit, the paper therefore supports treating the score as an editorial quality filter, while assessing actual citations from engine answers.
How is resistance to manipulation tested?
The validation protocol selects scoring formulas using five tests, called falsification gates. They check the response to keyword stuffing, increasing edit doses, single-feature amplification, duplication and added neutral length. Selection uses training data, with confirmation on held-out sources.
On the 500-source adversarial benchmark, amplifying calibrated features produces a reported gain of at most six score points, decreasing as the edit dose grows. That numerical cap is an empirical finding; the mathematical result concerns bounded amplification of a single feature under a concave transformation.
Standard web-spam baselines perform better at detecting degradation, and the score misses the tested attack families outside its calibration distribution. The authors propose layering the score over those baselines for quality filtering.
Related terms
FAQ
Can I interpret a higher score as a citation forecast?
The paper's conclusion positions question-blind scores as quality filters. A score computed from page text cannot resolve which source an engine selects for a particular question; the proposed uses include editorial quality control and filtering document collections.
Can I reproduce the results on my computer?
The repository's quick start installs pinned dependencies and runs the following command from the package directory:
./reproduce.sh
It reruns analyses from released captures and executes a consistency audit of 163 headline numbers. Computing fresh captures requires a separate scorer, which the repository excludes from its release.
Sources
- arXiv, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", 7 September 2026
- arXiv, "GEO: Generative Engine Optimization", submitted 16 November 2023, revised 28 June 2024
- TW3 Partners, "Scoring Without the Engine: reproducibility package", accessed 17 October 2026
Check how a page reads to answer engines. Open the GEO score analyzer