Two preprints — directional findings, not yet peer-reviewed.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. Schulte et al. (2026), "Don't Measure Once: Longitudinal Citation Stability in AI Search Platforms," arXiv:2604.07585. The Schulte findings come from Swiss-German commercial verticals (January–March 2026) and may not generalise across all geographies and industries. Both papers are preprints. This post describes their convergent findings and what that convergence is — and is not — sufficient to conclude.
You have been treating ChatGPT and Perplexity as two doors into the same room. They are not. Two independent research teams, using different methods, different time windows, and different measurement approaches, reached the same conclusion about how these platforms cite: one is built for depth, one is built for breadth. The distinction is not cosmetic. It changes what optimisation is worth doing and for whom.
Here is what the evidence shows — and what we think the convergence actually means.
Two measurements. Same platform picture.
Zhang et al. studied 602 controlled prompts across three platforms — ChatGPT, Google AI Overview, and Perplexity — and extracted 72 features per citation from 21,143 valid citations. Their focus was on citation selection (whether a source is included) and citation absorption (how deeply it shapes the generated answer). The absorption influence score is their unit of measurement: a quantified estimate of how much a cited page's content appears in the response.
Schulte et al. took a different approach. Over 45 to 46 days, they returned to the same prompts across four platforms — including ChatGPT and Perplexity — and measured whether the same sources came back each time. Their primary stability metrics are Jaccard similarity (the proportion of sources that overlap between two days) and RBO (rank-biased overlap, which weights consistency at the top of the source list more heavily). Their dataset covers four Swiss-German commercial verticals from January to March 2026.
Two preprints. Two measurement frameworks. Neither team was responding to the other's work.
Both arrived at the same platform characterisation. That convergence is what we find significant here.
The Depth Platform and the Breadth Platform
Zhang et al. find that ChatGPT cites an average of 6.88 sources per prompt. Perplexity cites 16.35. Perplexity selects more than twice as many sources as ChatGPT for the same query. On selection breadth alone, Perplexity operates at a higher volume.
The absorption picture inverts that relationship entirely.
ChatGPT's mean absorption influence score: 0.2713. Perplexity's: 0.0646. The same piece of content, cited once by ChatGPT, contributes roughly four times as much to the generated answer as the same piece cited once by Perplexity. ChatGPT selects fewer sources — and absorbs each one more deeply.
That is the first line of evidence.
The second comes from Schulte et al. On source-set stability across repeated daily queries, ChatGPT shows a Jaccard overlap of approximately 0.42 — meaning that roughly 42% of sources cited on a given day also appear the following day. Perplexity's Jaccard: approximately 0.34. ChatGPT's source pool is more consistent day-to-day. Perplexity's rotates more frequently.
These are not the same measurement. Absorption influence scores and Jaccard similarity values cannot be combined — they measure different properties of citation behaviour. The former captures how deeply a cited source shapes a response; the latter captures how consistently a source is included across repeated queries. Do not treat them as two columns of the same table.
What they are is two independent lines of evidence pointing the same direction. The Depth Platform and the Breadth Platform are not an artefact of one study's methodology. Two teams found the same pattern. Our read: when two methodologically independent preprints converge on the same platform characterisation, the direction is worth acting on even before peer review closes.
What the pattern means in practice
On the Breadth Platform — Perplexity — your content is competing in a larger pool. More sources are selected per prompt. Each cited source carries less absorption weight on average. The Gini coefficient Schulte et al. report for citation concentration is 0.715 across platforms — a high concentration score, meaning most citations cluster around a small number of frequently-returned sources. Getting into that pool at all is the hard problem. Once inside, deep influence per citation is lower.
On the Depth Platform — ChatGPT — selection is harder. Fewer sources make the cut per prompt. But when your content is selected, its influence on the generated answer is substantially higher. Source sets are also more stable: once ChatGPT is citing a page, it tends to continue citing it day-to-day. The barrier to entry is higher. The yield per citation is greater.
This asymmetry has a direct strategic implication — one that changes how we recommend approaching optimisation. Selection-focused tactics — improving retrieval indicators, structural clarity, broad topic coverage — address the Perplexity problem. They target inclusion in a larger, more volatile pool. Absorption-focused content properties — statistics, definitional language, comparative constructions, numerical specificity — address the ChatGPT problem. They target deeper influence once a source is selected.
A content strategy designed to perform on Perplexity will not automatically perform on ChatGPT.
These are different targets requiring different choices.
Why cross-platform consistency is rare
We built the Platform Variance tool to measure something practitioners have been noticing without a framework to explain it: the same content appears on one platform and not on another, without obvious reason. The breadth/depth divergence is the structural explanation for why that happens.
ChatGPT cites fewer sources and holds them more stably. Perplexity cites more sources with a rotating pool. A piece of content that succeeds on ChatGPT — deep, authoritative, statistically grounded — may not reach the volume threshold that gets into Perplexity's wider selection. A piece that reaches Perplexity's selection threshold — broad, high-retrieval, structurally clear — may not carry the absorption properties that make ChatGPT cite it deeply. The citation architecture difference is not random variance. It is built into how each platform relates to its sources.
Cross-platform consistency is the exception, not the default. Understanding why makes it possible to target it deliberately instead of hoping for it. That is what we built the Platform Variance tool to surface.
What this doesn't tell us
The Zhang et al. dataset is a controlled corpus of 602 prompts. It is not a random sample of the web, and generalisability to heterogeneous real-world content requires caution. The cross-sectional design means it captures platform behaviour at a point in time. Platform citation patterns can shift substantially — and fast. The Schulte et al. findings come specifically from Swiss-German commercial verticals in January through March 2026. Whether the same platform ordering holds in other geographies, other industries, or other query types is not established by this data.
Neither study is peer-reviewed. The preprint status is stated once here because it matters once: these are directional findings, not confirmed facts about permanent platform architecture. Act on them as directional evidence. Do not treat the specific figures — 6.88 vs. 16.35, 0.2713 vs. 0.0646, Jaccard 0.42 vs. 0.34 — as fixed properties of these platforms. Model updates change citation behaviour. The ChatGPT Reddit citation pattern collapsed from approximately 60% to approximately 10% between August and September 2025 (Semrush, 2025). Any specific figure in this post may already be out of date by the time you read it.
What is more likely to be durable is the directional pattern itself: ChatGPT cites fewer sources more deeply; Perplexity cites more sources more shallowly. That structural asymmetry is consistent across two different measurement approaches. The specific figures will shift. The architecture question is worth watching.
About the tools referenced in this post
The Platform Variance tool measures how a piece of content is likely to perform differently across ChatGPT's depth environment and Perplexity's breadth environment based on its current content profile. The breadth/depth framework described in this post is the structural explanation for the variance our tool surfaces.
The Absorption Analyser is the adjacent tool. If you are optimising for ChatGPT specifically — where absorption influence per citation is high and source sets are stable — the absorption dimensions our Analyser measures are the properties most relevant to that platform's citation architecture. Statistics. Definitions. Comparisons. Numerical specificity. These are the content properties we identified as associated with higher absorption depth in the Zhang et al. dataset.
Both tools are free, browser-based, and require no login.
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
Schulte et al. (2026), "Don't Measure Once: Longitudinal Citation Stability in AI Search Platforms," preprint. arXiv: 2604.07585
Semrush (2025), "The Most-Cited Domains in AI: A 3-Month Study." Industry study. Published November 10, 2025. semrush.com/blog/most-cited-domains-ai
Find out where your content lands.
The Platform Variance tool surfaces how your content profile maps to ChatGPT's depth architecture and Perplexity's breadth architecture — based on the citation properties identified in the Zhang et al. and Schulte et al. datasets.