More AI citations did not mean more of your page in the answer

What separates citation selection from citation absorption

The paper is From Citation Selection to Citation Absorption, by Zhang Kai, He Xinyue and Yao Jingang, all listed as independent researchers in China, dated 29 April 2026 and carrying the identifier 2604.25707v2 in cs.IR. Its argument is a vocabulary change before it is a finding. Selection is defined as the stage where a platform triggers a search and chooses sources. Absorption is defined as the stage where a cited page contributes language, evidence, structure or factual support to the final answer.

Once those are named separately it becomes obvious that most reporting collapses them. A dashboard that tells you your domain was cited 40 times last week has counted selection events. It has said nothing about whether any of those 40 answers borrowed a sentence, a number or a definition from the page, or whether the citation was a footnote the model attached to a claim it made without reading you. Both are called a citation and they are not the same event, which is roughly the distinction answer engine optimization has been trying to make in looser language for two years.

This is also the second recent paper to argue that the field measures the convenient thing rather than the causal thing. The survey of 45 GEO studies published in July reached its own version of the same complaint, reporting that no reviewed technique showed a stable cross-platform causal effect on discoverability. Splitting one outcome into two does not by itself fix that. What it does is stop a rise in citation count being read as a rise in influence, when the paper's own figures show those can move in opposite directions at the same time.

What the study counted, and what one AI citation record contains

The paper does not run its own crawl. It analyses a public dataset called geo-citation-lab, whose repository is at github.com/yaojingang/geo-citation-lab, given here as plain text because this site does not link that host. The third author maintains that repository, which the paper states in an author and data note rather than leaving to be discovered, and figures throughout the paper are captioned as coming from the geo-citation-lab public report.

The public snapshot it describes holds 602 prompts across three platforms, 21,181 cleaned search-layer rows, 21,143 valid search-layer citations, 23,745 citation-level feature rows, 72 feature dimensions and 18,151 successfully fetched citation pages. The prompts are split into four layers of 432, 60, 60 and 50, covering a main experiment, a style contrast, a language contrast and a set of realistic or extreme scenarios. A citation record is therefore not a URL in a list. It is a URL joined to up to 72 measured attributes of the page behind it, which is what makes the second stage computable at all.

One number is conspicuously absent, and it matters for a fast moving subject: the paper states no data collection period. There is no window given for when the 602 prompts were issued, so the most a reader can say is that the work was published on 29 April 2026 and the collection preceded it. Anyone citing these figures a year from now is citing a snapshot of unknown age. That is a real limit on the research value of the numbers and it is not a criticism the paper avoids elsewhere, since it is candid about most of its other constraints.

It is worth saying what 602 prompts buys, because sample size is the recurring weak point in this literature. An earlier study on how many prompts an AI visibility measurement needs put a confidence interval five points wide on citation share at roughly 40 queries on one platform and 150 or more on another. At 602 prompts across three platforms this dataset is comfortably past that bar for platform-level means, which is the level almost every figure below sits at.

Why the broadest citing platform had the lowest mean influence

Search triggering turned out not to be the interesting variable. The paper reports ChatGPT triggering a search on 579 of 587 observed prompts, a rate of 98.64 percent, Google on 600 of 602 at 99.67 percent, and Perplexity on 602 of 602 at 100.00 percent. When all three platforms search almost every time, the question of whether generative engines look at the live web is settled, and what remains is how many sources they take and what they do with them.

On breadth the three separate cleanly. Perplexity averaged 16.35 citations per prompt with a median of 17 and a maximum of 27. Google averaged 12.06 with a median of 12 and a maximum of 37, the widest single answer in the dataset. ChatGPT averaged 6.88 with a median of 6 and a maximum of 21. Read alone, that table says a page has more chances to be named by Perplexity than by ChatGPT, which is true and is the finding most likely to be quoted from this paper.

The absorption table inverts it. Mean influence among fetched pages was 0.2713 for ChatGPT, 0.0646 for Perplexity and 0.0584 for Google AI Overviews. ChatGPT's figure is more than four times Perplexity's on the mean and nearly eight times on the median, 0.2611 against 0.0333. The paper's reading is that ChatGPT cites fewer sources and uses the ones it cites more deeply, and that a single visibility metric collapsing both objectives loses information. Its practical version is blunter: for referral exposure a long citation list may be worth more, and for answer-shaping power a small number of high-absorption citations may matter more.

There is a precedent for the ranking not matching the intuition. A separate study found that nearly a third of AI Overview citations are not on page one of the same query's results, which is another case of an established metric failing to predict where citations actually come from. The pattern in both is the same, which is that the number everybody already collects turns out to be measuring a different thing from the one everybody assumes.

Roughly a quarter of the cited records never reached content analysis

The figure in this paper that belongs most directly to what a scanner measures is the one the paper spends least time on. It reports 18,151 successfully fetched citation pages and an overall fetch success rate of 76.44 percent, and then moves on.

The denominator is worth pinning down, because the paper does not state it. The three per-platform fetch-ok counts in the absorption table are 3,323 for ChatGPT, 6,385 for Google and 8,443 for Perplexity, which add to exactly 18,151. And 18,151 of the 23,745 citation-level feature rows is 76.44 percent, while 18,151 of the 21,143 valid citations would be 85.85 percent. So the rate is against the feature rows. That is arithmetic on the paper's own printed figures rather than a quotation from it, and it is the sort of thing worth checking before repeating a percentage, because the two candidate denominators are nine points apart.

Either way, between one in seven and one in four of the cited records did not yield a page the pipeline could read. The paper does not say why. Its section on robustness lists a fetch-failure model, described as predicting fetch_ok from source type, platform and domain in order to diagnose selection into the absorption sample, as part of a validation plan rather than as work already done. So the missingness is acknowledged and unmodelled, and every absorption figure quoted above is computed on the pages that were readable.

That gap is the whole subject of this site. A page an AI crawler cannot fetch, or can fetch but cannot turn into text, is absent from the second stage no matter how well it would have performed there. The failure modes are ordinary and individually boring: a robots.txt rule, a challenge page, a status code, or markup that arrives empty and fills in later in the browser. Some are not even failures in the way they look, since a 404 on robots.txt allows every crawler while a 503 blocks them all. The check that matters is whether the text a crawler receives resembles the text a person receives, which is what prose parity means, and what GPTBot sees on a given URL is answerable without any of this study's machinery.

What separated the top influence quartile from the bottom

Sorting the fetched pages by influence score and comparing the top quarter with the bottom quarter produces the largest gaps in the paper. Top-quartile pages averaged 1,943.30 words against 169.82, a ratio of 11.44x. They carried 10.59 headings against 0.85, a ratio of 12.50x, and 47.49 paragraphs against 8.34. List density was 0.428 against 0.048. Answer-citation semantic similarity was 0.570 against 0.247, and an LLM-rated relevance score was 3.535 against 1.856.

The obvious reading is that longer pages win, and the paper explicitly declines it. Among features not used inside the influence formula, the strongest reported correlation is that LLM relevance score at r = 0.4322, followed by answer-citation embedding similarity at r = 0.3561, LLM content quality at r = 0.2917 and question-citation embedding similarity at r = 0.2548. The paper's summary is that word count and structural markers matter but semantic fit is stronger than a simple length signal. A 2,000 word page about the wrong thing is still about the wrong thing.

The genre table is the part most likely to change what somebody writes. Pages containing code averaged 0.1747 influence against 0.0988 without, a relative difference of 76.88 percent. Numbers or statistics gave 0.1171 against 0.0725, definition markers 0.1252 against 0.0795, comparison content 0.1389 against 0.0894 and how-to content 0.1296 against 0.0918. The exception is the one worth remembering: Q and A formatting scored 0.0947 against 0.1005 for non-Q and A pages, a relative difference of minus 5.74 percent. Packaging a page as questions and answers, on this dataset, associated with slightly less absorption rather than more. By domain type, encyclopedia pages averaged 0.2144 influence and news media pages 0.0726.

None of that is causal and the paper says so, in a section on threats to validity and again in a claim-level self audit appendix. The honest summary is that pages which already function as evidence containers, with definitions, figures and comparisons a model can lift, are the ones that end up quoted. That is close to what structured data is for, though the mechanism here is the prose rather than the markup, and this site has already published the case that a perfect schema score and a zero can be measured on the same page. Google's own AI features documentation makes a similar point in weaker terms, which is that there is no special markup for AI answers and the page itself is the input.

Which of these stages an external scan can actually check

The influence score is built from five weighted components: reference count capped at three, position of first reference, paragraph coverage, TF-IDF cosine similarity and an average of bigram and trigram overlap, weighted 0.20, 0.15, 0.20, 0.25 and 0.20. The paper is careful about what that makes it, calling it a constructed observational proxy rather than a direct measure of hidden model attention, retrieval ranking or causal dependence, and warning that the variables inside the formula must not then be presented as independent drivers of the score they define. Reproducing any of it requires issuing prompts, capturing answers and diffing them against source pages.

Lantad does none of that, and the boundary is worth stating plainly rather than blurring. A scan fetches a URL as a named crawler, compares what that crawler receives against what a browser receives, evaluates the robots.txt rules per token, and reads the structured data. It produces no citation count and no influence figure, and our methodology is the place where what each part of the score does and does not cover is written down. The reason we withhold a grade when the measurement failed is the same instinct that makes this post keep saying whose numbers these are.

The connection between the two is one direction only, and it is a precondition rather than a promise. Every figure in this paper is conditional on a fetch having succeeded, both the platform's and the researchers'. A page that cannot be fetched contributes to neither stage and appears in no citation table, which means it will not show up as a poor result. It will show up as an absence, and an absence looks identical to not having been chosen. That asymmetry is why the readable check is worth running even though it predicts nothing about absorption: it is the difference between competing badly and not being in the set at all.

For the platforms in this study the fetching agents are documented. OpenAI's crawler documentation names OAI-SearchBot for search surfacing and ChatGPT-User for user-initiated fetches, and states that robots.txt rules may not apply to the latter because a person triggered it. Reading those pages and then checking your own file against them is a short task with a definite answer, which is more than can be said for anything downstream of it. What this paper adds is a reason not to stop at the citation count once you are past that gate, because on its own data the count and the contribution disagreed.