In short. Published intervention evidence supports no reliable citation gain from adding statistics across current engines.
Definition: Statistics addition is an intervention that inserts numerical claims into source content. Published experiments report early visibility gains and later null or negative results, so the evidence supports no reliable increase in citations across current engines.
Key facts
- Citead's GEO score paper reports a pooled statistics effect of -0.28 percentage points on its citation visibility measure, with a confidence interval spanning zero.
- Aggarwal and colleagues' original GEO paper tests statistics addition using a GPT-3.5-based answer system and a separate Perplexity experiment using uploaded sources.
- Puerto and colleagues' C-SEO Bench paper reports that its Statistics method decreases citation rankings in 19 of 24 evaluated settings.
What does Citead's later intervention test measure?
Citead's September 2026 paper, Scoring Without the Engine, remeasures interventions across ten engine families. Its outcome is a position-weighted share of answer words attributed to a source through sentence citations, described as a measure resembling position-adjusted word count (PAWC).
The experiment compares edited source content with neutral filler of matched length. This volume control separates the tested edit's effect from the effect of adding text. The pooled statistics result covers 1,087 paired observations: -0.28 percentage points, with a 95% confidence interval from -0.97 to +0.43. The interval includes zero, and the point estimate is negative.
The paper also reports a July 2026 replication on gpt-5.5, gpt-5.4 and gpt-5.4-mini. Across 450 paired queries, the statistics effect is -0.0066 in PAWC share units, with p = 0.25. These observations overlap the broader pooled analysis; they are separate views of the evidence.
The test uses five curated sources and a fixed citation prompt. Its deterministic edits approximate the original model-generated edits. The result therefore concerns those implementations and that controlled answer-generation setting; transfer to live search with retrieval and deduplication remains untested by this protocol.
The paper names the report containing the pooled modern intervention results:
modern_anchors_report.json
What does the original GEO experiment establish?
In GEO: Generative Engine Optimization, statistics addition replaces qualitative discussion with quantitative statistics where possible. The main answer system uses gpt3.5-turbo with five Google search sources. One source is selected for modification and kept constant across the compared methods.
The June 2024 revision's main table reports an overall position-adjusted word count score of 25.2 for statistics addition against 19.3 without optimization. This metric weights words attributed to a source by their position in the answer. It measures the source's share of answer visibility.
The separate Perplexity experiment covers 200 test samples. Its corresponding scores are 26.2 with statistics and 24.1 without optimization. Source text is supplied through file uploads, and answers are restricted to those files. These results establish gains within the paper's tested systems and metrics.
What does the citation-ranking evidence add?
C-SEO Bench: Does Conversational SEO Work? evaluates citation ranking: how early the answer cites the modified document. Its experiments cover question answering and product recommendations, using GPT-4o-mini, Claude Haiku 3.5, o3 and o4-mini as answer models.
The paper reports worse rankings for statistics addition in 19 of 24 settings. Its outcome differs from PAWC: an earlier citation and a larger share of attributed words are distinct measurements. Read together, the studies support treating statistics addition as a tested content intervention whose result depends on the model, task, edit and outcome measure.
Related terms
- The GEO score paper (arXiv 2609.07559)
- The 2023 GEO paper (Aggarwal et al.)
- Can early GEO experiments establish what works on current engines?
- Does a correlation with citations prove that a page feature causes them?
FAQ
Does a higher GEO score prove that adding statistics improves citations?
Citead's paper positions its query-independent content score as a quality filter. You need an intervention comparison with an answer-level outcome to assess whether an edit changes citations.
What should I check when someone quotes an improvement percentage?
Check the model, source delivery method and metric. The original GEO paper reports both position-adjusted word count and subjective impression, and its Perplexity test uses uploaded files.
Sources
- Citead, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", 7 September 2026
- Aggarwal et al., "GEO: Generative Engine Optimization", 16 November 2023, revised 28 June 2024
- Puerto et al., "C-SEO Bench: Does Conversational SEO Work?", 6 June 2025, revised 20 October 2025
Check how a page reads to answer engines. Open the GEO score analyzer