Can early GEO experiments establish what works on current engines?

Definition: Early generative engine optimization (GEO) experiments establish findings for the engines, queries and measurement procedures they test. Transferring those findings to current answer engines requires renewed experiments, as illustrated by Citead's remeasurement of the original GEO paper's content edits.

Key facts

• Aggarwal and colleagues' GEO paper is first submitted in November 2023; its latest listed revision is June 2024.
• The original paper's experimental setup supplies the top five Google sources to a gpt3.5-turbo model and samples five answers per query at temperature 0.7.
• Citead's GEO score paper, submitted in September 2026, reports no citation movement in the published positive direction for the tested quotation, statistics and source-citation edits across ten engine families.
• The later paper's abstract positions scores computed without the query as quality filters rather than citation predictors.

What does the original experiment measure?

Generative engine optimization means modifying website content to change its visibility in answers. In the original GEO study, the researchers select a source for each query, apply content edits separately and compare the resulting answers with an unmodified baseline. The paper reports increased visibility for edits that add quotations, statistics or source citations within the source text.

Visibility has a specific operational meaning. Position-Adjusted Word Count measures the words in answer sentences attributed to a source, with greater weight for earlier sentences. Subjective Impression uses model-based judgments about aspects of the citation's presentation. Reported improvements therefore describe these answer-level measurements.

The Perplexity experiment also has a defined boundary. Appendix C.1 supplies source text through file uploads and requires answers to use those files. That setup tests answer generation from supplied material. The paper's limitations explicitly anticipate adaptation as engines and query distributions evolve.

What changes in the later remeasurement?

Citead's outcome-validation experiment compares an unmodified target source, a deterministic quotation edit and a neutral addition matched for length. This control helps separate the effect of the added quotation from the effect of adding text. Its visibility measure is a position-weighted share of answer words attributed to each source.

The paper extends the comparison to statistics and source-citation edits and reports results across ten engine families. The tested interventions show no positive transfer of the earlier findings. This is an experimental result about those edits under the later protocol, with deterministic edits approximating the original model-generated modifications.

The later protocol still supplies five curated sources and fixes the citation-format prompt; the paper identifies open-web retrieval and competition between pages as a separate validation setting. Its content score consequently serves editorial quality control and corpus filtering, while citation selection remains dependent on the question and competing sources.

How should a website owner assess a transferred claim?

The two papers suggest a concrete reading procedure: identify the tested content change, the supplied sources and the answer metric before applying a result to a website. The original study's uploaded-source Perplexity test and Citead's fixed-source protocol make source selection an explicit boundary of the evidence.

For a claimed replication, the comparison also needs the engine identifiers, query set and control condition. Citead's paper documents pinned model identifiers for its initial outcome experiment and publishes raw records. It names this report for the pooled modern intervention measurements:

modern_anchors_report.json

The report concerns measured intervention effects. A content-score increase answers a different question about the page's measured properties. Keeping these outcomes distinct prevents a historical visibility result from becoming an unsupported current-engine recommendation.

Related terms

• The 2023 GEO paper (Aggarwal et al.)
• The GEO score paper (arXiv 2609.07559)
• Generative engine optimization (GEO)
• AI answer reproducibility

FAQ

Should I add quotations because the early study reports gains?

Treat that result as evidence from its original setup. Citead's later tests find no positive transfer for the tested quotation intervention, so that study provides no basis for promising a current citation gain from adding quotations.

Does testing Perplexity make the original result a live-web result?

The original paper's Appendix C.1 says the experiment uploads source text as files and confines answers to those sources. When assessing its relevance to your website, retain that supplied-source condition in the description of the finding.

Sources

1. arXiv, "GEO: Generative Engine Optimization", revised 28 June 2024
2. Aggarwal et al., "GEO: Generative Engine Optimization", 2024
3. Bajemon and Rochet, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", 7 September 2026
4. arXiv, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", 7 September 2026

---

Check how a page reads to answer engines. Open the GEO score analyzer