Elevating generative search: a comprehensive framework for optimization in AI platforms

Most GEO advice stops at one question: were you cited? A 2026 research paper, "From Citation Selection to Citation Absorption," argues that is only half the story. Being listed as a source is not the same as actually shaping the answer. The paper proposes a two-stage framework that separates the two, and it reframes what optimizing for AI search really means.

The two stages of GEO

The framework splits generative search performance into two distinct stages. Citation selection is whether a platform triggers a search and picks your page as a source at all. Citation absorption is whether that cited page actually contributes to the final answer, its language, evidence, structure, or facts.

The distinction matters because the two do not move together. You can be selected and still ignored, listed in the sources but absent from the words the user reads. Optimizing only for selection, the thing most tools count, misses whether you influenced the answer at all.

Selection vs absorption

The research found that citation breadth and citation depth diverge across engines. Some engines cite many sources but lean on each one lightly; others cite fewer sources but draw far more heavily from the ones they do pick. In the study, Perplexity and Google cited more sources on average, while ChatGPT cited fewer but showed substantially higher influence per fetched page.

The practical takeaway: a raw citation count tells you little. Appearing in a long source list on one engine is not the same as being the page another engine actually builds its answer from. Absorption, not just selection, is the outcome worth optimizing for.

What makes content get absorbed

The paper found that high-influence pages, the ones actually absorbed into answers, share clear traits. They tend to be longer, more structured, and semantically aligned with the question, and they are rich in extractable evidence: definitions, numerical facts, comparisons, and procedural steps.

That gives a concrete content brief. Vague, narrative pages may get cited but rarely absorbed, because there is nothing clean to lift. Pages that state definitions plainly, include real numbers, lay out comparisons, and give step-by-step procedures give the model something it can pull directly into the answer.

Best practices and checklist

Optimize for absorption, not just selection. Write so the model can lift a clear statement, then supply the evidence that makes your page the one it leans on. And because engines differ in how they cite, treat each one separately rather than assuming one strategy transfers.

A quick checklist:

• Measure absorption, whether you shape the answer, not just whether you are cited

• Lead with clear definitions and direct answers

• Include real numbers, comparisons, and procedural steps

• Structure content cleanly so evidence is easy to extract

• Keep pages semantically aligned with the questions buyers ask

• Track selection and absorption per engine, since they behave differently

Measuring absorption is hard to do by hand, since it means comparing your content against what actually appears in answers across engines. This is where a platform helps. As the marketing stack for the agentic web, Limy tracks how AI engines discover, cite, and draw on your content, and connects it to traffic, pipeline, and revenue. Start now to turn AI search into a measurable growth channel.