Getting Retrieved Is Only Half the Problem. The Other Half Is Absorption.

Most AEO advice is about getting retrieved. Show up in the right directories. Use FAQ schema. Structure your headings as questions. All of that is real, and it matters. But it targets only the first stage of a two-stage problem.

The second stage -- what researchers call citation absorption -- is where most of that work fails quietly.

There Are Two Gates, Not One

In October 2026, we filed our analysis of arXiv:2604.25707 (Zhang, He, Yao, session 164, October 10, 2026), a study covering 602 prompts and 21,143 citations across major AI platforms. The paper introduced a formal distinction that most AEO practitioners have been collapsing into one:

**Citation selection** is whether your page gets chosen as a source at all. This is what most AEO tactics address. Schema, heading structure, directory presence, crawlability -- these improve the odds that a retrieval system picks up your page.

**Citation absorption** is how much your page actually shapes the generated answer. A page can be cited in the sources list while contributing almost nothing to the answer text. Absorption measures the degree to which the model's output reflects your specific claims, data, and framing.

The distinction matters because the tactics that improve selection do not automatically improve absorption. You can optimize aggressively for selection and still contribute almost nothing to what the AI actually says.

The 15% Problem

The AirOps 2026 State of AI Search report, which we incorporated in session 164 (October 10, 2026), quantifies where the pipeline actually leaks.

Of all pages retrieved by an AI system during query processing, only 15% become actual citations in the final answer. Roughly 85% of fetched pages are discarded between retrieval and response generation.

Most AEO strategy focuses on getting into the retrieved set. That is a real threshold, and clearing it matters. But clearing it only gets you into the pool that the system then filters again. The 15% stat is a reminder that retrieval and citation are not the same event.

The same AirOps analysis found that pages with 50% or greater overlap between title and query had a 20.1% citation rate versus the 15% baseline -- a meaningful signal, but a narrow one. Title-to-query vocabulary alignment improves your odds of moving from retrieved to cited. It does not, by itself, determine absorption.

Why Q&A Format Helps Less Than You Think

The most widely cited structural intervention in AEO is the FAQ format. Add questions as H2 headings, answer them directly, use FAQ schema. This advice is not wrong, but it describes a selection tactic, not an absorption tactic.

Our analysis of arXiv:2604.25707 (session 164, October 10, 2026) found that Q&A formatting alone does not improve citation absorption. The structural signal tells a retrieval system where answers might live. Whether those answers are actually absorbed into the generated response depends on what is inside them.

Thin FAQ answers -- "Yes, we offer 24/7 emergency plumbing across the Greater Toronto Area" -- pass the structural check and fail the absorption check. The format signals that an answer exists. The retrieval system includes the page. The model reads it, finds nothing it can build a response around, and cites it nominally or not at all.

The distinction between selection and absorption reframes the FAQ question: the heading structure is a selection mechanism. The evidence density inside the answer is the absorption mechanism. You need both, and they require different fixes.

What Actually Drives Absorption

In session 165 (October 11, 2026), we filed our analysis of arXiv:2604.19113 (FeatGEO, ACL 2026), a peer-reviewed study that tested which document properties most reliably predict citation absorption across platforms.

The result contradicted a lot of surface-level AEO advice. Sentence-level lexical edits -- adding keywords, reformatting individual sentences, inserting citation language -- showed weak and inconsistent effects on absorption. Document-level properties drove citation significantly more than sentence-level interventions.

The four document-level properties that predicted absorption in the FeatGEO analysis:

**Topical completeness.** Does the page cover its claimed subject thoroughly? A page that answers one question well but omits adjacent questions leaves model-detectable gaps. Completeness signals that a page is authoritative on the topic, not just present for the query.

**Vocabulary alignment.** Does the page use the terminology that appears in the query and in related high-authority sources? Our session 58 analysis (June 25, 2026) found this was the only page-level signal to survive domain fixed effects in Discovered Labs' regression analysis (beta = +0.37). FeatGEO confirms it operates at the document level -- vocabulary coherence across a page, not individual keyword insertions.

**Structural coherence.** Is the document architecture internally consistent? Pages that mix abstraction levels, jump between topics without transitional logic, or use headings that don't match section content produce fragmented retrieval chunks. The model's synthesis process uses document structure as a navigation tool. Incoherent structure produces incoherent chunks, which produce low absorption.

**Evidence presence.** Are claims substantiated within the page? A specific measurement, timeframe, named credential, or cost range is not just a selection signal. It is what the model quotes. Generic claims do not get absorbed -- they have nothing to anchor a quote to.

The Order of Operations

There is a caution embedded in this research that matters operationally. The GEO Survey (arXiv:2607.14035, session 164, October 10, 2026) reviewed over 120 papers on generative engine optimization and found that citation-oriented surface rewrites -- adding bold text, inserting FAQ schema, reformatting for visual emphasis -- can impair retrieval when applied before infrastructure is solid. The mechanism is coherence damage: surface rewrites that don't account for the semantic flow of a document can fragment it in ways that hurt the retrieval system's ability to extract useful passages.

The sequence matters: get the infrastructure right first (crawlability, schema, directory presence, homepage structure), then address document-level content properties (topical completeness, vocabulary alignment, evidence density), then optimize structural elements. Inverting that order can produce a page that looks AEO-compliant and performs worse than it did before.

The Practical Implication

Most AEO audits identify selection failures. The page isn't being retrieved; fix the schema, check robots.txt, verify the directory listing. That is the right first step.

The second step is less common in AEO practice: test absorption. A page that appears in citations but never shapes the answer text has a structural or evidence density problem. The fix is not adding more schema. The fix is answering the question more completely, with specific data, in a document architecture that a retrieval system can navigate cleanly.

Getting retrieved is a threshold. Absorption is a spectrum. The businesses that show up in AI answers with their actual claims, prices, and service details intact have cleared both gates -- not just the one that AEO checklists measure.

If you want to see which gate you're failing at -- whether your pages are getting retrieved, cited, or actually shaping the generated answer -- Sourcepull's Signal Check runs live queries across ChatGPT, Perplexity, and Gemini and returns the actual citation text, not just a presence score.