Citation selection and absorption study

Definition: From Citation Selection to Citation Absorption proposes a distinction between an answer engine selecting a source and that source contributing to the answer. It measures the second outcome with an observational influence score combining citation placement, answer coverage and textual similarity.

Key facts

• The arXiv abstract describes 602 controlled prompts across ChatGPT, Google AI Overview/Gemini and Perplexity, with 21,143 valid search-layer citations.
• The paper's platform comparison reports fewer citations per prompt for ChatGPT alongside higher mean influence among successfully fetched pages.
• The authors' public repository characterizes the cross-platform experiment as a static research snapshot with incomplete record-level collection timestamps.

How does the study distinguish selection from absorption?

Zhang Kai, He Xinyue and Yao Jingang's paper, revised on 29 April 2026, defines selection through search triggering and source choice. Absorption concerns a cited page's contribution to the answer's language, evidence, structure or factual support. A source can therefore appear as a background reference or supply material used throughout an answer.

The paper's influence score combines repeated references, the position of the first reference, coverage across answer paragraphs and textual similarity between the answer and the source. Its equation can be written as:

influence_score =
    0.20 * min(ref_count / 3, 1)
  + 0.15 * (1 - first_position_ratio)
  + 0.20 * paragraph_coverage_ratio
  + 0.25 * tfidf_cosine
  + 0.20 * (bigram_overlap + trigram_overlap) / 2

Here, tfidf_cosine, bigram_overlap and trigram_overlap are the textual-similarity components. The score represents observable traces of answer participation; it remains a proxy for absorption. The authors exclude its own components from the proposed independent predictors of that same score, preventing circular explanations.

What observational evidence supports the distinction?

The April 2026 manuscript reports the following descriptive means for its research snapshot:

The ordering changes between citation breadth and the influence proxy. This is the central observational reason for measuring them separately.

The denominators also differ. Search-layer summaries cover 587 observed ChatGPT prompts and 602 prompts each for Google and Perplexity. Absorption analysis uses 18,151 successfully fetched citation pages, drawn from 23,745 citation-level feature records. Fetch failures therefore affect which sources enter the absorption comparison.

Pages in the highest influence quartile have more headings, paragraphs and words than those in the lowest quartile. Pages containing definitions, numerical evidence or comparisons also have higher mean influence in the reported feature comparisons. Q&A-format pages average 0.0947, compared with 0.1005 for other pages. These are descriptive associations; causal effects of rewriting a page remain unestablished.

How can a website use this framework?

The paper's proposed measurement dashboard separates citation occurrence, answer influence and support quality, meaning whether the cited material substantiates the answer's claims. For a website, this suggests examining the answer text alongside the linked URL when reviewing visibility.

The authors also propose sentence-level annotation that connects answer claims to supporting source pages. This supplies a concrete review method: identify the claim, locate its supporting passage and record the page's role. Their proposed roles include definitions, comparisons, factual evidence and background material.

Related terms

• AI citation
• Citation tracking
• Citation faithfulness
• Citation completeness

FAQ

Does a citation count measure absorption?

The study defines citation count as breadth. Its absorption proxy additionally uses paragraph coverage, citation position and textual overlap, so you need answer-level evidence to apply the distinction.

Should you convert every page into Q&A format?

The study's Q&A group has slightly lower mean influence than its comparison group. The authors propose testing page structure and evidence quality with controls before drawing intervention conclusions.

Can you interpret the influence score as model attention?

The authors define it as an observational proxy. You can use it to describe citation and textual patterns in the recorded answers; direct model attention requires a different measurement.

Sources

1. arXiv, "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms", revised 29 April 2026
2. Zhang Kai, He Xinyue and Yao Jingang, full manuscript, 29 April 2026
3. GEO Citation Lab, public repository, accessed 17 October 2026

---

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