Cited but not absorbed: GEO's missing metric

Citation absorption is the degree to which a cited page actually contributed language, evidence, structure, or facts to an AI-generated answer, as opposed to merely appearing in the source list underneath it. It is a different outcome from getting cited, and almost nobody is measuring it.

Most GEO dashboards, including the early versions of ours, treat the citation list as the scoreboard. Your URL appeared under the answer, so you won. That framing survives only until you read the answer text and notice that your page contributed nothing to it.

#Two stages, not one

An April 2026 paper from Zhang Kai, He Xinyue, and Yao Jingang proposes splitting the pipeline into two measurable stages: citation selection, where the platform triggers a search and picks sources, and citation absorption, where a cited page contributes to the final answer (arXiv 2604.25707). They analyze a public dataset of 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity, covering 21,143 valid search-layer citations and 18,151 successfully fetched pages (arXiv 2604.25707).

The useful part is the framing. Selection and absorption are separate filters, and passing the first tells you very little about the second. A page can clear retrieval, get listed, and still be decoration.

#Breadth and depth pull in opposite directions

The paper's central descriptive finding is that citation breadth and citation depth diverge (arXiv 2604.25707). Perplexity and Google cite more sources per prompt, while ChatGPT cites fewer and shows substantially higher average influence among the pages it fetches (arXiv 2604.25707).

That has an immediate consequence for reporting. If you count raw citations, Perplexity will look like your strongest channel simply because it lists more URLs per answer. If you weight by how much each source shaped the response, the ranking can invert. Two dashboards using the same crawl can hand you opposite platform strategies, and neither is wrong about its own metric.

This is the same denominator problem that makes rank-overlap studies contradict each other. Here it hides one layer deeper, inside a metric almost everyone reports as a single count.

#The commercial version of the gap

Semrush, working with Kevin Indig, ran the question from the brand side and got a name for the failure mode: ghost citations, where a platform sources a domain as a reference link but never mentions the brand in the answer text (Semrush). Across 3,981 domain appearances from 115 prompts in 14 countries, 61.7% were ghost citations, against 13.2% that were both cited and mentioned (Semrush).

The platform split is stark. In that dataset ChatGPT showed an 87% citation rate against a 20.7% mention rate, while Gemini showed the reverse shape: 83.7% mentions against a 21.4% citation rate (Semrush).

Ghost citations are not identical to low absorption. A page can be heavily absorbed and still not have its brand named, which is exactly what happens when an assistant paraphrases your definition into neutral prose. But the two measure the same underlying discomfort: the link list and the answer are different products, and only one of them reaches the reader.

#What absorbed pages have in common

The paper reports that high-influence pages tend to be longer, more structured, semantically aligned with the query, and richer in extractable evidence such as definitions, numerical facts, comparisons, and procedural steps (arXiv 2604.25707). Encyclopedia-style pages averaged higher influence than news media pages in their feature breakdown (arXiv 2604.25707).

A separate hand-coded study points the same direction from a much smaller sample. Advanced Web Ranking ran 40 SEO queries through Google AI Overviews and Bing Copilot Search on July 29, 2026, logged 265 citations, and manually coded 112 passages (Advanced Web Ranking). Comparing cited passages against uncited ones, two attributes separated cleanly: 6% of cited passages contained a hard number versus 0% of uncited ones, and 7% carried a novel claim versus 0% (Advanced Web Ranking).

Take those percentages for what they are. Six percent of ninety-five passages is a handful of pages, and the authors are explicit that they observed absorption patterns without proving them, because the retrieval pipeline is opaque. Read it as a direction, not a coefficient. The direction is consistent with the larger dataset: pages carrying something specific get used, pages restating consensus get listed.

That last point is the one worth sitting with. In the same coding pass, 82% of uncited passages were pure consensus restatement against 61% of cited ones (Advanced Web Ranking). Writing the same well-structured summary everyone else wrote is a strategy for being retrieved and ignored.

#Measuring absorption without a research budget

You do not need a feature pipeline to close most of this gap. Three changes to how you already track citations get you far enough.

Capture the answer text, not just the source list. Most citation trackers store the URLs and discard the response. Store the full answer. Everything below depends on having it.

Score each citation on a three-level scale. Not present in the answer, paraphrased without attribution, or named and quoted. That maps roughly onto absorption and takes a few seconds per answer to code by hand. Run it on a sample rather than the whole corpus.

Report the levels separately. A platform where you are cited constantly and named rarely is not the same win as one where you are named in the prose. Collapsing both into "citations: 47" hides the difference that actually moves pipeline.

Then treat the writing implication as the real work. If absorbed pages are the ones carrying definitions, numbers, comparisons, and steps, then the fastest way to move from listed to used is to publish something on the page that cannot be reconstructed from the other nine sources in the citation list. Original numbers you can stand behind. A comparison you actually ran. A procedure with real failure modes in it.

Being cited is a retrieval outcome. Being absorbed is an editorial one.