Preprint finding — directional, not yet peer-reviewed. Zhang, Kai, He Xinyue, and Yao Jingang (2026), "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms," arXiv:2604.25707. Submitted April 28, 2026; revised April 29, 2026. This post describes its findings and what their preprint status means for how to use them.
You have been optimising for AI visibility. You have been optimising for the wrong thing. Or rather, for only half of it. A preprint submitted to arXiv in April 2026 establishes that citation selection (whether an AI platform includes your source at all) and citation absorption (how deeply that source shapes the generated answer) are empirically distinct phenomena with different drivers. The paper's most counterintuitive finding: Q&A formatting — one of the most widely recommended AI optimisation tactics — is associated with a 5.74% reduction in absorption depth, not an improvement. We think the selection/absorption distinction is the most important framework shift in GEO research since Aggarwal.
That is not a small qualification. It is a direct challenge to a specific category of common advice.
What existed before this paper
The original Aggarwal et al. 2024 study — peer-reviewed, published at KDD — established that specific content properties (statistics, source attribution, readability) were associated with higher AI citation probability. That study measured citation in terms of impression metrics: position-adjusted word-count presence in AI-generated responses. It answered the question of whether a source was retrieved and surfaced. It did not separately measure how much that source shaped the content of the answer.
The Zhang et al. 2026 preprint introduces a different measurement dimension alongside that existing work — absorption depth, as distinct from the citation frequency measures in earlier research. These are not competing measurements of the same thing. They are measurements of two different stages of the same process. A page that ranks well on Aggarwal's impression metric may still rank poorly on absorption depth, and vice versa. The framework does not supersede earlier work; it extends it.
The two-stage framework
The dataset behind this paper — geo-citation-lab — contains 602 controlled prompts, 21,143 valid search-layer citations, and 23,745 citation-level feature records across three platforms: ChatGPT, Google AI Overview, and Perplexity. The researchers extracted 72 features per citation. That is a credible measurement study at scale, even at preprint status.
The framework distinguishes two stages:
Citation selection is the first stage. The platform retrieves a source and includes it in the response. Being selected is a necessary condition for everything that follows. It is also, the paper argues, an insufficient one.
Citation absorption is the second stage. The cited source contributes language, evidence, or structure to the generated answer. A page can be selected — retrieved, cited, attributed — without its content materially shaping what the AI says. Selection without absorption is a citation in name only.
The paper measures absorption using an influence score: a quantified measure of how much a cited page's content appears in the generated response. This is not the same as Aggarwal et al.'s impression metric. Do not treat the absorption influence scores as updates to or improvements on Aggarwal's figures — they measure a different dimension of what happens to a cited page.
Platform profiles: breadth vs. depth
The three platforms in this dataset behave differently on both dimensions. These figures come from a single cross-sectional preprint dataset — they describe current best available evidence, not permanent platform profiles. Platform behaviour changes as models are updated.
| ChatGPT | Google AI Overview | Perplexity | |
|---|---|---|---|
| Selection breadth Mean citations per prompt |
6.88 Depth platform |
12.06 Breadth platform |
16.35 Breadth platform |
| Absorption depth Mean absorption influence score |
0.2713 Highest depth |
0.0584 Lower depth |
0.0646 Lower depth |
The gap is striking. Perplexity cites 2.4 times as many sources per prompt as ChatGPT. ChatGPT absorbs its cited sources at 4.2 times the depth of Perplexity. The same piece of content, cited once by ChatGPT, contributes substantially more to the generated answer than the same piece cited twice by Perplexity.
A second, independent line of evidence points the same direction. Schulte et al. (2026) — studying Swiss-German commercial verticals from January to March 2026 — found that ChatGPT cites fewer sources with higher stability across repeated queries, while Perplexity cites more sources with lower consistency. These are different measurements using different methodologies (Schulte uses Jaccard similarity values; Zhang uses absorption influence scores), and the figures cannot be combined. But the platform pattern they describe is consistent. Two independent studies, neither peer-reviewed, pointing the same way.
What the Q&A finding actually says
The paper measured absorption influence scores across content formats. The finding for Q&A formatting: –5.74%. Pages formatted as question-and-answer pairs showed lower absorption depth than pages without that structure.
This is the most counterintuitive result in the paper. Q&A formatting is actively recommended across practitioner GEO content on the basis that it mirrors how users phrase queries to AI systems. The logic is intuitive — if users ask questions, structure your content as answers.
The Zhang data does not support that logic at the absorption stage. Q&A formatting may still improve selection — the paper does not rule that out. But for absorption, the association runs negative. The mechanism is not established. One plausible account: Q&A structure fragments continuous argument into discrete answer units, each self-contained, which may reduce the depth of semantic contribution any single page makes to a generated response. The paper does not confirm this mechanism. The finding is what it is; the explanation is inference.
Treat Q&A as a selection tactic, not an absorption tactic. These are different targets.
What does improve absorption
The paper identifies several content dimensions associated with higher absorption depth. These are from a preprint dataset. All figures are directional.
Code and numerical content: +76.88% absorption association. The largest absorption-positive dimension in the dataset. Pages containing code blocks, equations, or structured numerical data showed substantially higher influence scores. This is also the most platform-specific finding — code content is not relevant to all content types, and the paper does not break out the effect by platform.
Statistics: +61.55%. This corroborates the Aggarwal et al. 2024 finding at a different measurement stage. Aggarwal found statistics as a citation predictor; Zhang finds statistics as an absorption dimension. Two different studies, different measurement stages, same direction. Statistics are the only content dimension in this paper with external corroboration from peer-reviewed research.
Definitions: approximately +57%. Explicit definitional language — sentences that state what something is — associated with higher absorption depth. This echoes the structural finding in Zhang et al.'s own earlier analysis of the same dataset.
Comparisons: approximately +55%. Comparative constructions ("compared to," "unlike," "whereas") associated with higher absorption. Comparative structure gives a cited page a discrete positional claim an AI system can extract and attribute.
A note on structural features in this context: the H1 Research Dispatch post (Junwei Yu et al., arXiv:2603.29979) established that structural changes alone produce a mean 17.3% improvement in citation rates. Structural dimensions are absorption-relevant. But in Zhang et al.'s data, content features — statistics, definitions, comparisons, code — contribute more to absorption uplift than structural properties alone. Structure is a reliable baseline contribution. Content type is where the larger absorption gains appear.
What this means for your content strategy
The selection/absorption framework has a direct implication for how you allocate optimisation effort.
If your goal is citation breadth — appearing in as many AI responses as possible across platforms — Perplexity and Google AI Overview are the high-volume environments. Selection-focused tactics (retrieval indicators, structural clarity, coverage breadth) address that target.
If your goal is citation depth — having your content materially shape what AI systems say — ChatGPT is the high-absorption environment. The content dimensions associated with absorption depth in this dataset are not structural; they are substantive. Statistics. Definitions. Comparisons. Numerical specificity.
A Q&A content strategy that performs well on selection-breadth metrics may be actively working against absorption-depth on the same platform. These are not equivalent outcomes for most content producers. Knowing which you are optimising for changes what you should do next.
The deeper implication: platform-specific optimisation is not a tactical refinement. It is a strategic question. The same content cannot be simultaneously optimised for Perplexity's breadth profile and ChatGPT's depth profile without understanding what each platform actually does with cited content.
What this paper does not tell us
The geo-citation-lab dataset is a controlled corpus of 602 prompts — it is not a random sample of the web. Generalisability to heterogeneous real-world content requires caution. The cross-sectional design means this study captures platform behaviour at a point in time: April 2026. Platform citation behaviour can shift substantially within weeks, as documented in Semrush's 2026 analysis of ChatGPT's Reddit citation collapse (from approximately 60% to 10% between August and September 2025). These platform profiles are current best available evidence, not permanent facts about how these systems behave.
The paper does not establish causal mechanisms. The –5.74% Q&A finding is an association, not a proof that Q&A formatting causes reduced absorption. The +76.88% code/numerical finding is similarly correlational. Acting on these as directional findings is warranted. Acting on them as guaranteed levers is not.
The study has not completed peer review. Findings may be qualified or revised. The dataset scale (21,143 citations, 72 features) is substantial for a preprint. The directional findings are worth understanding now, with the appropriate calibration applied.
About the tools referenced in this post
The Absorption Analyser measures a page's absorption dimensions against the Zhang et al. dataset findings — statistics presence, definitional language, comparative language, structural properties — in a single scored output. The Platform Variance tool surfaces how a piece of content is likely to perform differently across selection-breadth and absorption-depth environments based on its current content profile. Both tools are free, browser-based, and require no login.
The selection/absorption distinction is now the framework we use across all Psytable tool descriptions to explain what these tools measure — and why that differs from what SEO tools measure.
References
Zhang, Kai, He Xinyue, and Yao Jingang (2026), "From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms," preprint. arXiv: 2604.25707
Aggarwal, Manas et al. (2024), "GEO: Generative Engine Optimization," ACM SIGKDD 2024. arXiv: 2311.09735
Yu, Junwei, Yang MuFeng, Yepeng Ding, and Hiroyuki Sato (2026), "Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior," preprint. arXiv: 2603.29979
Semrush (2026), "The Most-Cited Domains in AI: A 3-Month Study." Industry study, January–March 2026. semrush.com/blog/most-cited-domains-ai
Measure your absorption depth.
The Absorption Analyser scores the content dimensions associated with higher absorption in the Zhang et al. 2026 dataset — statistics, definitions, comparisons, structural properties — with evidence tiers clearly labelled.