In short. Added quotations show gains in the original GEO experiments but no reliable positive citation effect in Citead’s later controlled tests.
Definition: Added quotations are attributed passages inserted into a source page as a content intervention. They show positive visibility results in the original GEO experiments, but Citead's later controlled tests find no reliable positive citation effect across the tested engine families.
Key facts
- Citead's September 2026 GEO score paper reports a pooled quotation effect of -0.33 percentage points in citation-weighted answer share, with a 95% confidence interval from -0.84 to +0.17.
- Aggarwal and colleagues' original GEO paper reports a 22% relative improvement in position-adjusted word count, a visibility measure that gives earlier citation-linked words more weight, for quotation addition in its Perplexity experiment.
- The original paper’s arXiv record dates its first submission to 16 November 2023 and its latest revision to 28 June 2024.
What does the original experiment measure?
The original GEO paper defines quotation addition as inserting relevant quotations from credible sources. Its main experimental engine uses gpt3.5-turbo to answer questions from five sources retrieved through Google search.
Position-adjusted word count is a visibility measure that credits a source for words in sentences citing it, with greater weight for sentences earlier in the answer. The reported 22% Perplexity improvement is relative to the unmodified-source baseline on this measure. It describes the amount and placement of cited material within answers.
The Perplexity experiment tests 200 samples and constrains answers to uploaded source texts. Its result therefore concerns quotation edits within supplied material. Applying the result to a public website also involves retrieval and source selection, stages outside that file-based comparison.
What changes in the repeated tests?
Citead's GEO score paper repeats quotation interventions across ten engine families, including open-weights models and proprietary models. Its protocol gives an engine five sources and compares the target source under unmodified, quotation-edited and length-matched neutral-filler conditions. Matching added length helps separate the quotation intervention from the effect of providing more text.
The outcome is a share of position-weighted words in sentences citing a source. The quotation effect subtracts the neutral-filler change from the quotation-edit change. Pooling every complete quotation arm gives -0.33 percentage points across 1,531 paired engine-query observations; the confidence interval spans zero.
A separate July 2026 replication on gpt-5.5, gpt-5.4 and gpt-5.4-mini also finds no positive quotation effect. Its nominally negative pooled result fails the paper's correction for testing multiple interventions. Together, these results support the paper's conclusion that the earlier positive intervention effect fails to transfer to its tested settings.
How should a website interpret the comparison?
The Citead study describes its fixed citation prompt, curated sources and answer-share measure as a controlled proxy. It uses a deterministic quotation edit to operationalize the intervention. The comparison tests transfer across experimental settings; it leaves the effect of every possible quotation edit in live search unresolved.
The units also matter: a relative percentage improvement in an older visibility metric and a later percentage-point change in answer share cannot be subtracted to calculate a decline. For a website, the evidence supports treating quotation addition as an editorial choice without a reliable cross-engine citation uplift. Citead positions its content score as a quality filter rather than a citation predictor.
Related terms
- The 2023 GEO paper (Aggarwal et al.)
- The GEO score paper (arXiv 2609.07559)
- Can early GEO experiments establish what works on current engines?
- Do added statistics reliably increase citations across current engines?
FAQ
Should I add quotations specifically to increase citations?
The repeated quotation intervention in Citead's study gives no reliable positive effect across its tested engine families. Its results provide no basis for expecting a consistent citation increase from that edit.
Does the original Perplexity result describe ordinary website retrieval?
The original paper evaluates uploaded source texts and restricts answers to those files. Its 22% result concerns visibility within that supplied-source experiment.
Can I use a higher GEO score to predict citations after adding a quote?
Citead's paper positions the score as a content quality filter. It separates the score's response to edits from the measured citation outcome.
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", 28 June 2024 revision; first submitted 16 November 2023
Check how a page reads to answer engines. Open the GEO score analyzer