An AI strategy is not finished because an engine selected our page. Selection is the engine pointing at a URL. Absorption is whether that page contributes language, evidence, or structure the answer actually reuses. A strategy that stops at the link can collect citations that change nothing a buyer reads.
The dataset is Citation Absorption vs Citation Selection. The order I run the layers is the stack on Machine Relations. When a pitch renames a layer, the stable names are in the Machine Relations glossary.
What the measurement separated
A GEO measurement paper separates selection from absorption across 602 controlled prompts, 21,143 valid search-layer citations, and 18,151 fetched pages on ChatGPT, Google AI Overview and Gemini, and Perplexity (arXiv, From Citation Selection to Citation Absorption). The same study extracted 72 features from the pages it fetched, looking for what predicts reuse and not only what predicts being chosen. Those features are not a template to stuff into a CMS. The split is the operating fact. A platform can choose a page and absorb nothing from it. A selection rate measures half the system.
The original GEO paper is the earlier half of the same lesson. It introduced GEO as a way to improve visibility in generative engine responses and found methods that could boost visibility by up to 40% on its benchmark (arXiv, GEO). The paper studied how content becomes visible inside an answer. It did not say "publish more." The citable claim, the source behind it, and the structure that lets a model lift it are the assets. Keep the 40% as evidence that composition matters: quotations, structure, and cited numbers. Do not hire a team to chase 40% on a page that has no primary fact.
Citation behavior is observable in other work too. A 2026 PMLR paper models retrieval and citation as processes over query-document pairs and studies citation on "Your Money or Your Life" queries from the MS MARCO Web Search dataset (PMLR). Where the answer can move money or health, the engine's choice of source is an object to audit. Being findable is a different event from being used.
Adoption is not a source system
Stanford's 2025 AI Index reported organizational AI use at 78% in 2024, and generative AI use in at least one business function more than doubling, from 33% to 71%. The same report put corporate AI investment at $252.3 billion in 2024 and private generative AI investment at $33.9 billion (Stanford HAI). That is a lot of buying. It is not a source system. Those figures explain the inbox and why a tool-first strategy feels responsible. They do not rank our sources, and they are not evidence that a purchase absorbs anything. The first question is which sources an answer can cite when a buyer asks who matters in the category, and whether those sources contribute anything once they are cited.
I have made the other mistake. I used to think the remedy for a fuzzy market was more speed: more tools, more people, more output. The dependency I was outrunning was that the market could not say what was being built. Answer engines make that expensive sooner, because they assemble an explanation from whatever they can retrieve. If what they can retrieve is thin, the strategy is thin, regardless of the software line on the budget.
Google's Search Central documentation says AI Overviews and AI Mode surface relevant links so people can find and explore content, and that there are no additional requirements to appear in those features beyond existing Search guidance (Google). That is the floor: a page that can be crawled, a purpose a parser can see, claims with sources, ordinary technical hygiene. It is not "do last year's SEO and wait." A page that cannot be fetched cannot be reused, and a page that is fetched can still contribute nothing. If the page cannot be crawled, the fetch date never arrives, and absorption is undefined.
Four layers, then one query
| Layer | The question | What gets built | How it fails |
|---|---|---|---|
| Owned clarity | Can a machine explain who we are? | Definition, category, and comparison pages in direct language | We are visible and misunderstood |
| Earned authority | Can a machine verify us off our own site? | Third-party reporting, a named interview, a sourceable proof | The graph is only our own voice |
| Structured proof | Can a machine lift the claim? | Tables, definitions, answer-first blocks, cited numbers | The page exists and is not reusable |
| Measurement | Can we see use, not only publication? | A fixed prompt set, citation tracking, whether the answer reused the page | Shipping gets reported as progress |
I take those four layers over six disconnected AI subscriptions. GEO and AEO sit inside that system as work on answer surfaces. They are not a substitute for a third-party source, and a third-party source is not a substitute for a page that can be parsed.
Then delete the work that never produces a fetch date: the prompt library nobody opens, the calendar that exists to "have AI content," the chatbot on a process that is broken before it is automated, and the dashboard that counts output and never checks whether an answer reused the page. The smaller thing, kept: name the category query, write the direct answer, support it with a primary source or a third-party proof, structure it so a model can lift it, publish it where it can be crawled, earn a corroboration we do not own, and measure retrieval, citation, and reuse. Publishing is not the last step. Use is.
I do not try to reproduce 72 page features. I pick the category query that shows up in sales calls and log three dates. The day a page of ours was fetched. The day an answer selected it. The day a sentence, a number, or a table from that page showed up in the answer in a form I could match. The absorption paper's scale is what convinces me those dates can come apart. If I only celebrate the second date, I fund pages that are linked and unused. A strategy memo that says "we were cited" and cannot say whether anything we wrote survived into the answer is a memo about the pointer.
Owned clarity makes the fetched page say the right thing. Earned authority makes a second domain agree, so the model is not trusting only us. Structured proof makes absorption possible: a table or a number that can move into the answer without the model's paraphrase washing the claim out. Measurement is the log of the three dates, kept on a frozen query so a wording change cannot impersonate progress. I review that log monthly. I do not add a tool whose screen cannot show the third date.
Pick one category query and track the three events for a quarter. The company this fits already has something true to say, and a buyer question where being named without being used would create a false sense that the strategy is working. This is not a tool-procurement project.
If a measurement on our own query, over a disclosed window, showed selection rate and absorption moving together so tightly that tracking reuse added nothing, the two metrics collapse and the saved attention goes to earning the next source. The paper exists because that collapse is not what they found across those 21,143 citations. Until our own query shows it, a selected URL does not count as a win by itself.
FAQ
What is the difference between selection and absorption?
Citation selection is the engine choosing a page. Citation absorption is whether that page's language, evidence, or structure shows up in the answer. The paper that separates them covers 602 controlled prompts, 21,143 search-layer citations, and 18,151 fetched pages, and it extracted 72 features aimed at reuse. A selected page that contributes nothing is a citation I do not celebrate.
Where do the Stanford numbers fit?
They describe adoption and spend: 78% organizational AI use in 2024, generative AI in at least one function at 71% after 33% the year before, $252.3 billion of corporate AI investment, $33.9 billion of private generative AI investment. They explain why a tool-first strategy feels responsible. They do not choose a tool. The source a buyer's question can cite is the prior decision.
Does Google's guidance mean ordinary SEO is enough?
Google says AI Overviews and AI Mode add no extra requirements beyond existing Search guidance, and that those features surface links. That is the floor: crawlable, understandable, useful pages. The absorption study is why the floor is not the whole strategy. A page can clear the floor, be selected, and still contribute nothing the answer reuses.
What gets deleted before another tool is added?
Anything that reports activity without a retrieval, citation, or reuse event attached. The GEO benchmark's "up to 40%" is a finding about visibility methods on a benchmark, not a reason to ship more undifferentiated pages. Keep the page that a fixed query can fetch and reuse. Drop the calendar whose only job is volume.
About the author
Jaxon Parrott founded AuthorityTech in 2018 and coined Machine Relations in 2024. He writes about earned authority, AI discovery and founder judgment.
Continue reading
The Category Ownership Trap
Owning a category name leaves the buyer's question open. I coined Machine Relations in 2024, and the commercial question is still a separate job.
How to Buy Performance PR on Citations
Tie a PR fee to citations now that earned media can be counted. Muck Rack put earned media at 84 percent of AI citations and paid content at 0.3 percent.
Who Owns the Next Move After an AI Mention
An AI mention becomes a decision once someone verifies the observation and checks the claim. Then you decide whether a commercial action changes.