Obsurfable

Retrieval Explains Citations — Not FAQ Blocks: What Chen and Koudas Found Across 468,000 AI Links

Obsurfable

The GEO industry sells page templates: FAQ blocks, schema.org markup, summary boxes, author bios, faster Core Web Vitals. Most vendor case studies compare your site to competitors — confounding page tweaks with domain authority.

Mahe Chen and Nick Koudas at the University of Toronto asked a harder question: within one website, which pages get cited and which do not? Their paper What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), posted to arXiv in October 2026 and circulating widely in industry coverage since, tracked 158,099 answers and 468,037 citations across 90 days of commercial buyer questions — then replayed every question through ChatGPT and Gemini APIs to observe the searches models run before they cite.

The headline is blunt: retrieval explains 32% of citation variation between pages on the same site. Page properties explain 2.8% for ChatGPT. The practices most often sold as “AI-ready formatting” — FAQ blocks, schema, summary boxes, author boxes — do not associate with more citations when you hold the website fixed. Matching titles and paragraphs to the questions people ask is the page signal that survives.

Why within-website comparisons matter

Observational GEO studies often show FAQ pages “winning” because strong brands implemented FAQs first. Chen and Koudas invert the design:

  1. Collect daily answers from ChatGPT (GPT-5.4 nano + web search), Gemini (3.1 Flash Lite + grounding), and Perplexity (Sonar) for 651 commercial queries in US fiber internet (June–September 2026).
  2. Download cited pages, extract 14 structural features, and model citation counts.
  3. Restrict analysis to pages on the same domain so brand strength cannot explain the gap.
  4. Replay each query through APIs that return rewritten sub-queries, ranked source lists, and selected citations.

The primary market is consumer ISPs — a concentrated category with recognizable incumbents. A second 55-query B2B software vendor panel replicates the direction of results for smaller brands models barely know.

This design speaks directly to marketing teams asking: “We added schema to 40 URLs — why did citations not move?” Chen and Koudas’s answer: you may be optimizing selection after retrieval already decided the candidate set.

Finding 1: FAQ, schema, and summary boxes fail the within-site test

Across 12,038 pages cited at least twice on 1,120 websites, the authors regress log citation counts on page features within each site.

SignalWithin-website effect
Title / paragraph match to buyer questionsPositive, robust (survives five falsification tests)
FAQ blocksNo lift within site; appears helpful only across sites
schema.org markupNo lift within site
Summary boxes, author boxesNo lift within site
Page speed (subset with data)No lift within site

At equal rank in ChatGPT’s retrieved source list — the step where the model chooses which of ~26 consulted pages to cite — the “recommended” practices trend slightly negative:

Practice at equal rankOdds vs baseline (ChatGPT conditional logit)
Author boxes× 0.77
FAQ blocks× 0.93
schema.org markup× 0.92

That does not prove FAQs harm citations. It means once retrieval puts you in the pool, those blocks are not what separates winners from losers — and naïve cross-site comparisons overstate their value.

This aligns with Indexably’s 18,129-page study, where domain factors dominated page factors — but Chen and Koudas go further by holding domain constant.

Finding 2: Every cited ChatGPT URL was retrieved first

API replay is the paper’s methodological contribution. For ChatGPT:

  • 100% of cited URLs (22,315 in replay; 2,107 in the B2B set) appeared in ChatGPT’s own search results first.
  • When replay runs without web search, collected answers had no citation in 83.1% of matching cases — versus 7.1% when search ran.
  • Each answer triggers about four rewritten sub-queries, consults ~26 pages, and cites ~3.
  • 43.8% of rank-1 retrieved pages are cited; 3.5% at ranks 11–20.

Sub-queries rarely match Google or Bing rankings for the same buyer intent. Optimizing for classic blue-link position is therefore an indirect bet on ChatGPT’s separate index and query rewrite behavior — consistent with why each engine reads a different search index.

McFadden pseudo R² for explaining which retrieved page gets cited:

ModelPseudo R²
Rank in ChatGPT list alone26.8%
Rank + query match + page features32.5%

Retrieval rank is most of the story. Content features nudge outcomes at the margin once you are on the list.

Finding 3: Content still matters — through semantic match, not formatting

Pages that match buyer questions in meaning (title, first paragraph, best paragraph cosine similarity to queries) cite more often. At equal rank, brand-owned pages cite more (× 1.71), more headings help (× 1.28), and shorter pages cite more (× 0.59 per SD length).

Paradoxically, listicles, comparison roundups, and Reddit threads are consulted often but cited rarely at equal rank (comparison × 0.34, Reddit × 0.30). Engines read them for context; they prefer to cite cleaner evidence pages when choosing footnotes.

Practical translation:

  • Write pages whose H1 and opening answer could be pasted into the model’s rewritten sub-query without semantic drift.
  • Split long guides into targeted URLs per question cluster — not one mega-FAQ.
  • Do not confuse being in the consideration set (lists/forums) with being the cited source (product, docs, primary research).

Finding 4: Model memory names brands without pages

Chen and Koudas separate citation (linked evidence) from mention (name in prose). ChatGPT names brands without any supporting page in a large share of mentions — 46% in their reporting for the consumer panel. It also writes familiar brands into its sub-queries and reorders otherwise identical product descriptions when brand labels change in controlled experiments.

For incumbents, that is free share of voice. For challengers in the B2B replication, “models barely know the brands” — visibility stabilizes only when search finds third-party evidence.

That rhymes with ghost citations and AI recognizes brands but rarely mentions them: entity memory and retrieval evidence are separate levers.

Engine differences in the same dataset

Table-level contrasts from the paper’s 90-day panel:

EngineCitations per answerShare to brand-owned sites
ChatGPT2.0 distinct URLs64.3%
Gemini10.1 distinct URLs16.8% (Reddit alone ~13%)
Perplexity0.3 (96% of answers cite nothing)6.8%

Strategies are not portable: ChatGPT concentrates citations on known brand domains; Gemini spreads citations across third-party pages. Perplexity in this sample barely surfaces links at all — mention-only answers dominate.

What this does not mean

It does not mean delete your FAQ pages. Cross-site correlations can still reflect that serious publishers ship FAQs and authority. Within-site, FAQs are not the differentiator.

It does not mean ignore crawlability. Retrieval cannot cite blocked pages. See robots.txt and AI visibility.

It does not replace the Aggarwal GEO experiments. Lab settings that fix the source set measure answer-stage rewrites. Chen and Koudas measure production retrieval, where most variation happens upstream.

It is not eternal truth. Models, indexes, and rewrite policies change weekly. The authors release data and code so others can re-run the replay — treat this as November 2026 evidence, not scripture.

Playbook: sequence retrieval before page polish

Step 1 — Confirm retrieval eligibility

  • Allow search bots (OAI-SearchBot, Googlebot, PerplexityBot) distinct from training opt-outs — see shared training opt-out plans.
  • Map which sub-queries your category triggers by replaying 20–50 real buyer prompts through ChatGPT/Gemini APIs or Obsurfable Tasks — not just ranking checks on google.com.

Step 2 — Match language to those sub-queries

  • One primary question per URL; title states the question; first paragraph answers it.
  • Prefer evidence pages over “best of” listicles for owned properties you want cited.

Step 3 — Only then tune on-page GEO

  • Headings, brevity, and technical HTML hygiene — not schema-for-schema’s sake.
  • For high-authority domains, compounding returns still appear in page optimization quartile work.

Step 4 — Measure both mentions and citations

Retrieval and memory can produce mentions without links. Track both, as in brand coverage vs citation volatility.

How Obsurfable fits

Chen and Koudas show why single-scrape rank trackers mislead: the same question rewrites into different sub-queries daily. Explorer stores repeated observations — prompts, answers, brands, citations — so you see whether a page rewrite moved retrieval, citation, or only mention. The App ties that corpus to your company for mention rate and share of voice.

FAQ

Should we stop investing in schema.org?

Keep schema for entity consistency (product, organization, FAQ where humans need it). Do not expect schema alone to separate two pages on your own site.

Does this contradict “first 30% of the page” guidance?

No. Where AI looks on the page applies after retrieval. Chen and Koudas explain who enters the candidate set.

How do we audit ChatGPT retrieval for our category?

Use API replay (the paper’s method), third-party monitors, or Obsurfable prompt panels. Compare retrieved hostname share and cited URL share — not just SERP position.

Is Gemini the same as ChatGPT here?

Directionally yes on query rewrite ≠ Google SERP, but Gemini cites far more third-party URLs. Platform-specific playbooks still matter.

Bottom line

Across 468,037 citations, Chen and Koudas show that generative engines cite pages their own search retrieved, ranked by semantic match to machine-written sub-queries — not by FAQ blocks or schema within a site. Retrieval explains 32% of within-site variation; page templates explain 2.8%. Brands are also named from model memory without links. Fix retrieval and question-match first; treat fashionable GEO markup as a distant second — and measure whether you moved citations, not just implemented a checklist.