Retrieval Gets You Cited. Content Keeps You There. The Ratio Is 11 to 1.

Most AEO advice is about page content. Optimize your FAQ sections. Add structured data. Rewrite your service page headers for keyword alignment. Add a summary box. These recommendations dominate practitioner content because they're actionable, repeatable, and easy to measure.

A preprint published October 10, 2026 put a number on how much they matter. It is smaller than most AEO advice implies.

What a large-scale study found

In our October 24, 2026 research session (Scout session 178, `research-vault/knowledge/content-format-citation-signals.md`), we filed findings from arXiv:2610.15033 -- "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)" by Mahe Chen and Nick Koudas. The study analyzed 158,099 AI answers and 468,037 citations across two 90-day collection windows, with findings replicated on a separate set of small B2B vendors.

The core quantification:

- **Retrieval factors explain 32% of citation variation** - **Page-level content properties explain 2.8% of citation variation**

Retrieval factors are the signals that determine whether a page enters the AI's candidate pool at all -- directory presence, entity recognition, NAP consistency, training data mentions. Page-level content properties are the on-page signals that AEO advice typically focuses on: FAQ markup, structured data, heading structure, summary boxes, page speed.

The gap between 32% and 2.8% is not rounding. It is a factor of roughly 11.

For a business spending time on FAQ formatting and structured data before establishing strong directory presence, the sequence is backwards. Content optimization operates within the 2.8% after you are in the retrieval pool. The 32% is upstream.

What "retrieval factors" actually means

The distinction the paper draws is between what gets a page into the candidate pool versus what happens after it arrives there.

Retrieval factors include the signals AI platforms use to recognize and select pages as candidates for a given query. In practice, these are:

**Directory presence and NAP consistency.** Businesses present across multiple third-party directories give AI platforms corroborating evidence for entity recognition. Our June 2026 methodology documentation (session 42, `knowledge/platform-divergence-data.md`) showed that businesses with consistent profiles on four or more third-party platforms appear in AI responses at meaningfully higher rates than those on one or none.

**Training data footprint.** ChatGPT, Claude, and Gemini answer most queries from parametric knowledge -- what the model learned during training. A business with substantial coverage in sources that entered training data (established publications, Wikipedia, major review platforms from 2023-2025) has a durable advantage that page optimization cannot replicate. Chen and Koudas identify this as a distinct third factor beyond retrieval and content: the model has a pre-trained prior toward brands it already knows, which biases its citation queries toward them before any page is fetched.

**Live retrieval quality (Perplexity-specific).** For platforms doing real-time web retrieval, crawlability and source authority determine whether a page is in the candidate pool. A site behind a Cloudflare AI-crawler block never enters the pool regardless of how good the content is.

Page-level content properties -- FAQ sections, structured data, heading structure, content length, page speed -- operate after a page is already in the pool. They influence which pages from the pool get selected. They do not determine whether a business enters the pool at all.

The training data factor explains a pattern we see in audits

The third factor Chen and Koudas identify -- model training knowledge of a brand -- resolves a pattern that surfaces consistently in Sourcepull audit data: two businesses with similar directory presence and page quality show dramatically different citation rates.

The explanation is pre-retrieval. AI platforms build their candidate queries with a prior toward brands they know from training. ChatGPT will sometimes name a brand with no supporting retrieved page -- the model already "knows" them and places them directly in citations. A business with thin pre-2025 web coverage faces a structural disadvantage not because of anything about their current website, but because the model's training weights are lower for their entity.

This is supported by a separate finding we filed the same session (Scout session 178, `knowledge/platform-citation-behaviors.md`): 5WPR's "State of AI Citations 2026" (May 23, 2026) found a 0.334 correlation between a brand's search volume and its likelihood of receiving an AI citation -- stronger, the study claims, than backlinks as a standalone predictor. The mechanism fits: high-search-volume brands generated more web activity during the 2023-2025 training data window, which means they are more likely to appear in the parametric knowledge the model draws on.

This is not a lever any single page optimization touches. It is a longer-cycle signal built through sustained digital presence -- press coverage, review volume, directory breadth, community mentions -- over time.

Content properties still matter. What the 2.8% actually covers.

The 2.8% is not zero. Content properties do influence citation selection -- they just do so at the within-pool stage, not the pool-entry stage. Understanding what they affect is more useful than discarding them.

A peer-reviewed ACL 2026 paper we filed in our October 11, 2026 research session (Scout session 165, `knowledge/content-format-citation-signals.md`) directly addresses the content side of this split. arXiv:2604.19113, "Think Before Writing" by Liu and Xu (FeatGEO), established that document-level content properties drive citation visibility more strongly than token-level rewrites. The finding: AI systems select pages from the candidate pool based on high-level interpretable features -- topical completeness, vocabulary alignment with the query, structural coherence, evidence density -- rather than the presence or absence of specific markup patterns.

The content signals that matter inside the 2.8% are not "add a FAQ box to the service page." They are:

**Vocabulary alignment.** Pages that use the words buyers actually type -- not professional synonyms or category-adjacent language -- match more sub-queries in the AI's fan-out decomposition. A plumbing company whose service page says "residential drain clearing" instead of "expert hydro-jetting solutions" is more likely to appear in the candidate pool for the exact vocabulary Perplexity retrieves on.

**Topical completeness.** A page that covers the full scope of the query topic -- not just the headline service but the surrounding context -- is more likely to be selected from the pool. A service page that mentions the service, the process, common questions, and pricing signals is more valuable than one that covers only the headline.

**Structural extractability.** AI systems prefer to synthesize from pages where the key answer is surfaced early in each section, with clear evidence (numbers, specifics, comparisons) rather than continuous promotional prose. This is a document architecture question, not a FAQ-tag question.

These are higher-effort changes than adding schema markup. But they operate inside the candidate pool and compound on top of the 32% effect when retrieval pool membership is already established.

The practical sequence

The Chen and Koudas quantification gives a formal foundation to a priority ordering that Sourcepull's methodology has documented empirically: retrieval pool membership first, content selection optimization second.

For a business not yet appearing in AI citations, the fix plan should lead with the 32% factors: claim and complete the directory profiles that matter for your category and platform mix, establish consistent entity signals, and build the web presence that feeds training data over time. Content optimization on a site that AI platforms are not yet retrieving produces no measurable citation improvement.

For a business already receiving some citations and targeting specific platforms or queries where performance is weak, content-level work becomes relevant. Use the vocabulary your buyers type. Build topical completeness into your service pages. Structure content for extractability, not for aesthetics.

The 2.8% is real and worth capturing. It is just not the place to start.

If you want to know which of these stages applies to your business right now -- whether you are pre-pool (Phase 1) or optimizing within the pool (Phase 2) -- a free Signal Check from Sourcepull will show you which platforms are currently retrieving your pages and which aren't.

---

*Sources: arXiv:2610.15033, "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)" (Mahe Chen, Nick Koudas; submitted October 10, 2026, cs.IR). arXiv:2604.19113, "Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility" (Zikang Liu, Peilan Xu; ACL 2026, pages 20290-20303). 5WPR "State of AI Citations 2026," May 23, 2026. Platform divergence research: Averi B2B SaaS Citation Benchmarks 2026, session 42 analysis.*