What AI Engines Hear When a Founder Speaks (And Why It Matters for Citation)

Last updated: 2026-10-13

TL;DR

• The default chunk is small. Google’s Vertex AI RAG documentation shows a default of 1,024 tokens per chunk with 256 tokens of overlap, so an engine does not retrieve your essay. It retrieves a passage [1][2].
• Specifics win the passage. In the KDD 2024 paper that introduced GEO, adding quotations and statistics produced some of the largest visibility gains on the benchmark, up to about 40 percent on the “position-adjusted word count” metric for citation sources [3].
• Context around a chunk is a known lever. Anthropic’s September 2024 Contextual Retrieval work reports a 49 percent reduction in failed retrievals, and a 67 percent reduction when combined with reranking, by prepending chunk-specific context before embedding [4].
• Tree-structured retrieval exists. Stanford’s RAPTOR builds a hierarchy of recursive summaries over long documents and improves QA accuracy on NarrativeQA, QASPER, and QuALITY, with up to a 20 percent absolute gain on QuALITY paired with GPT-4 [5].
• Scope. No vendor publishes the exact ranker that decides whether your passage gets cited, and I did not run a head-to-head founder-voice benchmark. The mechanisms below are documented in papers and vendor docs. The sentence shapes I recommend are my inference from how those mechanisms work.

In Google’s Vertex AI RAG defaults, retrieval operates on 1,024-token chunks with 256 tokens of overlap, not whole pages [1][2]. That is roughly 750 words with about 190 words of overlap between neighbors. It matters because the thing an engine quotes about a founder is almost never the whole post. It is one of those passages.

I run a podcast with 100+ AI founder interviews on it [6]. Some of our guests get cited by ChatGPT, Perplexity, and Google AI Mode. Others get indexed but not cited. The parent of this post, The Founder-Voice Moat, argues that specificity and voice provenance are what separate them [7]. This post goes one layer down. What does retrieval actually do with a founder’s transcript, and which sentence shapes make a passage easier to lift?

What is retrieval doing with a founder’s words?

A modern AI answer is produced in two steps, retrieval and generation. The retrieval step takes a query, splits every candidate page into chunks, scores them, and hands the top passages to the model. The generation step writes the answer, often with attribution back to the chunk it used [8].

The chunk is the unit. Stockebrand’s O’Reilly book on answer engine optimization puts it directly: “Most modern systems also use semantic chunking with overlap windows, not fixed word counts. Google’s Vertex AI RAG documentation shows defaults of 1,024 tokens per chunk with 256 tokens of overlap. Coherent paragraphs that stand alone semantically perform better than long, meandering sections that only make sense in full context” [9]. Google’s own documentation shows the same defaults [1][2].

The newer systems add structure on top of flat chunks. Stanford’s RAPTOR, published at ICLR 2024, builds a tree of recursive summaries over a document, so a question about a long book can be answered from a high-level summary chunk at the top of the tree or a leaf chunk near the bottom, depending on scope [5]. Perplexity’s own Search API documentation describes an in-house index built for “sub-document, fine-grained relevance,” so results are often “passages or snippets” rather than full pages [10]. Different vendors, same pattern. The retrieval target is a slice, not a document.

Three practical consequences follow.

• A sentence that only makes sense if you read the three before it loses in retrieval, because its chunk no longer stands alone.
• A paragraph that packs a name, a company, a specific number, and a claim into itself is a self-contained unit that an engine can quote whole.
• A post split across pagination loses: “A chunk from Page 2 might reference ‘the process described above’ without including the context from Page 1. Those incomplete, context-poor chunks perform badly in retrieval because they can’t stand on their own as meaningful answers” [11].
• Structured data on the page can guide the chunker. Dell Technologies describes “schema-aware chunking” that keeps related attributes together as metadata on each chunk, and Microsoft’s October 2025 guidance says AI assistants “break content into smaller, usable pieces, a process called parsing” and that structural signals shape how those pieces are built. I unpack the mechanism and the controlled schema-only experiments across five AI engines in How Schema Markup Might Actually Work in AI Search [18].

Why does “Hector Hernandez and Speakeasy” chunk better than “the CEO at a company”?

Modern retrieval combines two signals: dense vector similarity between the chunk and the query, and sparse term matching such as BM25. Anthropic’s Contextual Retrieval work, published September 19, 2024, reports that combining the two reduces failed retrievals by 49 percent, and by 67 percent with a reranker, over single-signal baselines [4]. Dense retrievers do most of the semantic work. Sparse retrievers catch exact strings, which matters for names.

Named entities are exact strings. In Hector Hernandez’s episode on TwoSetAI, he said, “the first thing that breaks when people scale their use of AI is often not cost. Most of the time it’s complete loss of visibility of what’s going on” [12]. That one sentence contains the entity “Hector Hernandez” via byline, “Speakeasy” via page title, and the specific claim “visibility breaks first.” A query like “what breaks first when a company scales AI” matches on semantics. A query like “Hector Hernandez Speakeasy” matches on sparse exact strings. A query like “Speakeasy agent identity” matches on both.

Compare it to “the CEO of a leading AI startup said scaling can be challenging.” The semantic match degrades because the sentence is generic. The sparse match collapses because there is no exact name. Both retrieval signals weaken on the same text. This is my inference from the papers above, not a vendor-published scoring rule.

Co-occurrence is the second half of the pattern. Research on open information extraction and entity co-occurrence shows that named entities appearing together in the same window are treated as related in downstream knowledge graphs, with the window usually being a sentence or a short chunk [13]. If the founder’s name, the company name, and the key claim all sit inside the same 1,024-token chunk, an engine building a knowledge graph has an easier time storing them as a triple. If the name sits in paragraph one and the claim sits in paragraph nine, the chunk boundary can split them.

Attribution scaffolding helps the generation step, not retrieval. Stockebrand notes that AI systems “place significantly more weight on claims about your brand when they’re supported by multiple independent sources” and that bare claims without corroboration get softened or dropped [14]. Attribution patterns such as “According to Hector Hernandez, co-founder of Speakeasy” pre-format the sentence into something an engine can cite with attribution intact. That is a style guess backed by Stockebrand’s language, not a benchmark.

Why do specific numbers beat vague adjectives?

The KDD 2024 paper that introduced GEO ran the largest documented head-to-head on this. The authors defined nine content optimization methods and measured how each changed visibility in generative-engine answers across 10,000 queries [3]. Their results table, reported on arXiv, shows some of the largest gains for “Cite Sources,” “Quotation Addition,” and “Statistics Addition,” at roughly 30 to 40 percent lift on their position-adjusted word count metric for citation sources. Keyword stuffing gave little [3].

That is a benchmark, not a promise. The paper evaluated on a specific generator and specific benchmark queries. Our own observations are directional. In our August 2026 audit of past TwoSetAI guests, 86 percent were cited by at least one AI engine; a recent cohort of Founder Visibility Program participants were cited more often than an earlier cohort on every surface, which we attribute to packaging, not only voice [15]. We did not isolate specificity as a single variable.

The mechanism I would guess is this. A specific number is a verifiable string. A dense retriever picks up its semantics. A sparse retriever picks up its exact digits. A generator can quote it without rephrasing, which keeps the citation attachable to a real page. A vague adjective supplies none of that: “many,” “most,” and “significant” look the same across every page an engine has indexed, so there is no retrieval reason to pick yours.

Sentence feature	What retrieval sees	Why it matters
Named founder + company in the same chunk	Entities linkable via co-occurrence [13]	Knowledge-graph triple is formable from one passage
Specific number with a unit	Exact string match plus semantic match	Sparse signal fires; generator can quote verbatim
Attribution scaffolding (“According to [Founder]…”)	Pre-formatted citation shape	Generator keeps attribution attached in the output
A self-contained thesis sentence	Chunk stands alone semantically [9]	Passes the “stands alone” test the Vertex docs encode
Vague adjective, no entities	Weak on both dense and sparse signals	Interchangeable with ten other pages, no reason to cite you

Does attribution scaffolding help?

Attribution scaffolding is the sentence shape “According to [person], [role] at [company], [claim].” Stockebrand’s framing on independent-source corroboration, cited above, treats attribution as something an engine’s generation step weights [14]. I did not find a published study that isolates the scaffold as a single variable, so I treat it as an inference from the mechanisms, not a benchmark finding.

The reasoning is narrow. A generator that is grounded in a retrieved chunk tends to carry the chunk’s attribution forward into the answer, because the alternative is to produce an unattributed claim that may be flagged as a hallucination [8]. If your original sentence already pre-formats the attribution, the generator has less work to do and less reason to rewrite you into generic voice. If your original sentence is “we help enterprises win,” the generator has to invent attribution, and often simply drops you.

The practical form I use in our transcripts is: “According to [Founder Name], [Role] at [Company], [one specific claim with a number or named entity].” It is a template, but a template of the right shape costs an engine nothing to lift.

Four sentence-shape rules a founder can apply this week

These are rules I would give a founder preparing for a podcast or writing a LinkedIn post. They come from the mechanisms above. I believe they are correct in direction. I do not claim they guarantee a citation.

1. Put the name, the company, and the claim in the same paragraph. Not “the problem we see” in paragraph three with “Speakeasy addresses this” in paragraph seven. One passage, three entities, one claim. If an engine’s chunk boundary falls in the middle, you still have most of the triple on each side.
2. Replace one adjective with one number in every spoken answer. “A lot of agents” becomes “a hundred agents inside a Fortune 500 Salesforce instance.” “Significant time saved” becomes “about 40 hours of audit review per week.” The number is a sparse-retrieval anchor. The unit tells a generator you are not guessing.
3. Pre-format at least one quotable sentence with its own attribution. “I’d say the first thing that breaks when people scale their use of AI is not cost, it’s visibility” is a sentence that keeps you attached when it gets lifted. Say it in interviews. Write it on your site. The Quote That Carried the Episode covers how we mine these out of transcripts.
4. Keep your most important paragraph a self-contained unit. A reader who starts in the middle of your page should still be able to tell what the paragraph is about, who said it, and why it matters. That is the same test the Vertex RAG defaults encode. If the paragraph fails the stand-alone test, chunk retrieval fails on it too [9][1].

If you want the broader framework behind these rules, How to Find Your Founder Voice is the companion piece. If you want the counterexample, The Ghostwriter Trap covers why a brief-driven template produces the opposite of all four rules. For how engines read the next layer up, How AI Engines Parse Video handles transcripts and chapters, and How AI Engines Decide What to Cite covers the ranking step after retrieval.

So, what do engines actually hear?

Not what you wrote. A passage of what you wrote. The passage that gets quoted is the one that stands alone, carries your name and your number, and keeps the attribution attached long enough for a generator to repeat it. Everything else is background. That is the technical reason the Founder-Voice Moat compounds: a founder with a story and a number produces passages that pass every retrieval test by construction, and a template produces passages that fail them.

I cannot prove any specific engine will cite you after you follow these rules. These systems update their rankers quietly, and no vendor publishes the exact scoring formula. What you control is the sentence shape. If your sentences lose on the mechanisms documented here, your odds are near zero. If they win on the mechanisms, you have given the system something it can lift. That is our work at TwoSetAI: we turn one podcast conversation into lasting AI-search presence for AI founders whose buyers check AI before they check the website. When your buyers ask AI, be the answer. See the Hector Hernandez watch page for the concrete example this post is built around.

Who is TwoSetAI?

TwoSetAI is one of the top AI podcast and video studios in Silicon Valley, and the team behind this article, hosted by Angelina Yang. The show has featured 100+ AI founders and leaders on how real AI products get built, sold, and trusted. Alongside the podcast, TwoSetAI runs the Founder Visibility Program that turns each conversation into the articles, clips, and pages that make its guests easy to find. Watch full episodes on YouTube at youtube.com/@TwoSetAI, and find every guest and episode at twosetai.com.

Glossary

• Chunk (also called: passage, retrieval unit). A fixed-size slice of a document, typically around 1,024 tokens, that an AI retrieval system scores against a query [1][9].
• Named-entity co-occurrence (also called: entity co-mention, in-window co-occurrence). The pattern in which two or more named entities appear inside the same chunk and are treated as linked by downstream knowledge extraction [13].
• Attribution scaffolding (also called: citation scaffolding, pre-formatted attribution). A sentence shape such as “According to [person], [role] at [company], [claim]” that keeps attribution attached when the sentence is lifted into a generated answer.
• Tree-structured retrieval (also called: hierarchical retrieval, recursive summary retrieval). A retrieval design, exemplified by RAPTOR, that stores recursive summaries of a document in a tree so queries can be answered at the right level of granularity [5].
• Voice provenance (also called: source trail, voice traceability). The trail from a published sentence back to something a founder actually said, did, or decided; coined in the parent pillar [7].

About Angelina Yang

Angelina Yang is VP of AI, a fast.ai fellow, and the host of the TwoSetAI podcast, where she has interviewed 100+ AI founders and leaders about how real AI products get built, sold, and trusted. Find her on LinkedIn, talk to her AI clone on Delphi, watch the podcast on YouTube, or browse twosetai.com.

Frequently Asked Questions

How big is the chunk that AI engines actually retrieve? In Google’s Vertex AI RAG defaults, a chunk is 1,024 tokens with 256 tokens of overlap, which is about 750 words with 190 words of overlap between neighbors [1][2]. Other vendors vary, but the pattern of “a passage, not a document” is consistent across RAPTOR, Perplexity’s Search API, and the systems Stockebrand surveys [5][9][10].

Do AI engines use named entity recognition on my content? Modern retrieval combines dense vector similarity with sparse exact-string matching, and named entities are exact strings that fire on the sparse signal [4]. Downstream knowledge extraction pipelines also use co-occurrence of entities inside a chunk to build triples [13]. I did not find a vendor that confirms “we run NER on your blog post before retrieval,” so I treat the mechanism as evidence-supported, not vendor-confirmed.

Will “according to [my name]” phrasing improve my chance of citation? I believe it helps, because an attribution-formatted sentence survives into the generated answer with your attribution intact, and because Stockebrand’s framing on independent-source corroboration treats attribution as weighted by the generator [14]. I did not find a benchmark that isolates the scaffold as a single variable. The practice costs nothing.

Is there research on which content changes increase AI citations? Yes. The KDD 2024 GEO paper tested nine optimization methods across roughly 10,000 queries and found that adding citations, quotations, and statistics produced the largest visibility gains, around 30 to 40 percent on their primary metric. Keyword stuffing gave little [3].

Does Perplexity index pages differently than Google AI Mode or ChatGPT? Perplexity’s documentation describes an in-house index built for “sub-document, fine-grained relevance” and returns passages or snippets rather than full pages [10]. ChatGPT Search is powered by Bing’s index [16]. Google AI Mode draws on Google’s own index. The chunk-level retrieval target is the common pattern. The ranking signals and training vary.

What is Anthropic’s Contextual Retrieval? A method, published September 19, 2024, that prepends chunk-specific context to each chunk before embedding and before BM25 indexing [4]. On the benchmarks Anthropic reports, it cuts failed retrievals by 49 percent, or 67 percent with a reranker on top. It argues that chunks without surrounding context lose retrieval accuracy, which is why “write passages that stand alone” matters.

Does any of this guarantee my post will be cited? No. These systems update their rankers quietly, and no vendor publishes the exact scoring formula [8]. The best I can offer is: your sentence shapes either win on documented retrieval mechanisms or lose on them. Winning is necessary and not sufficient.

How does this connect to the Founder-Voice Moat? A founder’s specific, named, numbered sentence is the one input a template cannot produce, and it is also the input retrieval scores highest. The content moat and the retrieval mechanism point to the same page. The Founder-Voice Moat is the full argument.

References

[1] Google Cloud (2026). “Chunk documents.” Vertex AI RAG Engine documentation, accessed October 13, 2026. https://cloud.google.com/vertex-ai/generative-ai/docs/rag-engine/chunk-documents

[2] Google Cloud (2026). “Transform API.” Vertex AI RAG Engine documentation, default chunk_size 1024 and chunk_overlap 256, accessed October 13, 2026. https://cloud.google.com/vertex-ai/generative-ai/docs/rag-engine/transform-api

[3] Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). “GEO: Generative Engine Optimization.” KDD 2024; arXiv:2311.09735. https://arxiv.org/abs/2311.09735

[4] Anthropic (2024). “Introducing Contextual Retrieval.” Anthropic blog, September 19, 2024 (reports 49% reduction in failed retrievals and 67% with reranking). https://www.anthropic.com/news/contextual-retrieval

[5] Sarthi, P., Abdullah, S., Tuli, A., Khanna, S., Goldie, A., and Manning, C. D. (2024). “RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval.” ICLR 2024; arXiv:2401.18059 (reports up to 20% absolute gain on QuALITY paired with GPT-4). https://arxiv.org/abs/2401.18059

[6] TwoSetAI (2026). “TwoSetAI podcast guest archive.” Accessed October 13, 2026. https://www.twosetai.com

[7] TwoSetAI (2026). “Why Organic Founder Voice Beats Ghostwriters in AI Search: Patterns From 100 AI Founders.” TwoSetAI blog, October 12, 2026. https://www.twosetai.com/blogs/the-founder-voice-moat/

[8] Lewis, P., Perez, E., Piktus, A., et al. (2020). “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020; arXiv:2005.11401. https://arxiv.org/abs/2005.11401

[9] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026 (Chapter 2 on semantic chunking defaults; Chapter 4 on self-contained passages). https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[10] Perplexity (2026). “Search API.” Perplexity documentation, accessed October 13, 2026 (“sub-document, fine-grained relevance”; results as passages or snippets). https://docs.perplexity.ai/getting-started/overview

[11] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026, Chapter 3 on single-URL long-form over paginated splits. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[12] TwoSetAI (2026). Hector Hernandez (Co-founder and COO, Speakeasy) episode, published October 5, 2026; quotes at 22:32 and 23:11. https://www.youtube.com/watch?v=yhOLsoqSSEY

[13] Angeli, G., Premkumar, M. J., and Manning, C. D. (2015). “Leveraging Linguistic Structure For Open Domain Information Extraction.” ACL 2015 (foundation for entity co-occurrence in RE/KG pipelines). https://nlp.stanford.edu/pubs/angeli2015openie.pdf

[14] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026, Chapter 6 on independent-source corroboration weighting. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[15] TwoSetAI (2026). Guest search audit, checked August 10 to 25, 2026, incognito, screenshots on file. Internal data; protocol in the Methodology note.

[16] Microsoft Bing (2024). “ChatGPT Search is powered by Bing.” Bing Blogs, October 31, 2024. https://blogs.bing.com/search/October-2024/ChatGPT-search-is-now-powered-by-Bing

[17] Loktionova, M. (2026). “We Analyzed 89K LinkedIn URLs Cited in AI Search: Here’s What Drives Visibility.” Semrush Blog, March 10, 2026 (325,000 prompts, January to February 2026; ChatGPT Search, Google AI Mode, Perplexity). https://www.semrush.com/blog/linkedin-ai-visibility-study/

[18] Yang, A. (2025). “How Schema Markup Might Actually Work in AI Search.” Medium, December 17, 2025 (republished at twosetai.com/blogs/how-schema-markup-might-actually-work-in-ai-search). https://angelina-yang.medium.com/how-schema-markup-might-actually-work-in-ai-search-d017b1d2ef83

Methodology note. Last verified 2026-10-13. The mechanisms in this post are documented in the cited vendor docs and papers; the specific sentence-shape rules at the end are my inference from those mechanisms, not benchmark findings, and I label them as such. The KDD 2024 figures [3] are reported from the paper’s own evaluation on their benchmark set and specific generator; “up to 40 percent” refers to the largest reported gains on the position-adjusted word count metric and does not transfer to arbitrary content. The Anthropic Contextual Retrieval numbers [4] are the authors’ reported benchmark results. Google’s Vertex AI RAG defaults [1][2] are the published documentation defaults and may be configured differently per project. Guest quotes are verbatim from the published episode cited, with punctuation and capitalization normalized and timestamps taken from the episode transcript. Several references cite Stockebrand’s O’Reilly book [9][11][14]; his cited third-party figures (AirOps, Vertex defaults) are directional. The TwoSetAI audit [15] covered a cohort of past guests, not the full catalog. “Attribution scaffolding,” “voice provenance,” and “Founder-Voice Moat” are TwoSetAI’s own terms, not industry standards. Nothing here guarantees a citation, a ranking, or a business outcome.