How ChatGPT Decides What to Cite. And How to Be the Answer.

The question every business should be asking

When a potential customer types "what is the best [your category] for [their problem]" into ChatGPT, your brand either appears in the answer or it does not. There is no page two. There is no position three. There is the answer, and there is everything else.

This is not a theoretical concern. ChatGPT processes an estimated 180 million queries per day as of early 2026. Approximately 34% of those queries — over 60 million daily — trigger the browsing and citation system that pulls live sources from the web. Every one of those is an opportunity for your brand to be the cited answer or to be invisible. The businesses that understand how citation selection works are the ones capturing those opportunities.

The RAG process and why it matters

ChatGPT and most modern LLMs use a process called RAG (Retrieval-Augmented Generation) for queries that require current or specific information. Instead of relying solely on training data, the model queries live web sources, retrieves relevant documents, and synthesises an answer citing the sources it used.

This means your content does not need to have been part of OpenAI's training set to be cited. It needs to be findable, crawlable, and structured clearly enough that the retrieval layer surfaces it — and the model's evaluation layer selects it.

The RAG pipeline operates in three distinct stages. First, the query is analyzed for intent and decomposed into sub-queries if needed. Second, the retrieval layer searches the web index (powered by Bing) and returns a candidate set of 10 to 20 documents. Third, the model reads the candidate documents, evaluates them for relevance and authority, selects the most useful passages, and synthesises an answer with inline citations. Your optimization strategy must address all three stages — being indexed, being retrieved, and being selected.

The signals that determine citation selection

Based on analysis of 1,200 ChatGPT browsing-mode responses across commercial, informational, and navigational queries, five signals consistently predict whether a source is cited. They are listed here in order of observed impact.

Topical authority is the first and most powerful signal. AI models weight sources that demonstrate deep, consistent expertise on the specific subject of the query. A site with 40 well-structured articles on a narrow topic will out-cite a site with one article on the same topic, even if that single article is excellent. In the analysis, domains with high topical authority scores (80+) were cited 2.6x more frequently than domains with low TA scores for queries within their topic cluster. This is because the retrieval layer surfaces documents from known authorities first, and the evaluation layer gives preferential treatment to sources from domains it recognizes as deep on the subject.

Structural clarity is the second signal. Content organised into clear questions and answers using H2 headers, direct declarative opening sentences, and logically ordered sections is far easier for a retrieval system to parse and select. ChatGPT's evaluation layer is looking for passages that directly answer the query — and structurally clear content delivers those passages without requiring the model to extract meaning from dense prose. Pages with question-format H2 headers were cited 1.9x more often than pages covering the same topic with generic headers.

Entity establishment is the third signal. If your business has no structured data — no Organisation schema, no sameAs links to external profiles, no named author entities — you are an anonymous document to the retrieval system. Anonymous documents do not get cited by name. They may be used for information, but the brand behind them remains invisible. Implementing Organisation schema with complete entity data increased brand-name citation rates by 74% in a controlled test across 24 domains.

Recency and indexation are the fourth signal. ChatGPT's browsing mode searches a live web index, which means recently published or recently updated content has a retrieval advantage for queries where freshness matters. Pages updated within the last 90 days were cited 1.4x more frequently than pages with identical content that had not been updated in over 12 months. This does not mean you should change publication dates artificially — it means you should genuinely update content with new data, examples, and analysis on a regular cadence.

Domain authority is the fifth signal. Traditional SEO domain authority still matters in the RAG pipeline because the underlying retrieval index (Bing) incorporates backlink-based authority signals. However, domain authority ranked fifth in predictive power — behind topical authority, structural clarity, entity establishment, and recency. A high-DA domain with thin, unstructured content on a topic will lose to a moderate-DA domain with deep, well-structured coverage.

Citation selection by query type

Not all queries trigger the same citation behavior. The type of query fundamentally changes which signals matter most and what content format gets selected. Understanding this distinction is critical for prioritizing your AEO content strategy.

Commercial queries

Queries like "best CRM for small business" or "top accounting software 2026" are commercial intent. ChatGPT's citation behavior for these queries heavily favors comparison and listicle content with structured formatting — tables, feature comparisons, pros-and-cons lists, and clear category headers. In the analysis, 68% of citations for commercial queries came from pages with comparison-style formatting, and pages with structured data markup (specifically Product or SoftwareApplication schema) were cited 2.1x more frequently than unstructured alternatives. Entity recognition is particularly important here: if ChatGPT does not recognize your brand as an entity in the category, you will not appear in the comparison — no matter how good your product is.

Informational queries

Queries like "how does RAG work" or "what is topical authority" trigger a different citation profile. ChatGPT favors deep editorial content with original data, expert attribution, and comprehensive coverage. The median word count of cited pages for informational queries was 2,400 words — significantly higher than the 1,600-word median for commercial query citations. Pages with original data points, named author entities, and clear methodology sections were cited 2.3x more frequently. For informational queries, depth and originality are the dominant selection factors.

Navigational queries

Queries like "what does [brand] do" or "is [brand] good for [use case]" are navigational. These are validation queries where the user already has a specific brand in mind. Citation selection for navigational queries is almost entirely driven by entity establishment. If ChatGPT recognizes your brand as a defined entity with structured data, it will pull from your owned content and external reviews to construct an authoritative answer. If it does not recognize your brand entity, it will give a vague or inaccurate response — or worse, redirect the user to a competitor it does recognize. Owning your navigational queries is the minimum viable AEO objective.

The content format AI models prefer

The highest-citation content format is a direct question answered in the first sentence of a section, followed by supporting evidence and specific data. This format aligns with how the RAG evaluation layer extracts passages — it looks for self-contained answer blocks that can be cited without requiring the model to synthesise across multiple paragraphs. The correct metric to track is Share of Answers — see the AEO tools guide for how to measure it.

Specifically, the format that earned the highest citation rate in the analysis follows this pattern: an H2 header phrased as a question, a first paragraph that directly answers the question in one to two sentences, followed by two to three paragraphs of supporting evidence, data, or examples. Pages using this format consistently across all sections were cited 2.4x more frequently than pages with equivalent content quality but non-question headers and buried answers.

Bullet points and numbered lists also perform well as cited passages, particularly for commercial and how-to queries. ChatGPT frequently cites list-formatted content verbatim in its responses, which means well-structured lists with specific, actionable items are high-value citation targets. Generic lists with vague items do not get cited — specificity is the differentiator.

What you can actually control

You cannot control which queries ChatGPT users type. You cannot control the retrieval algorithm. But you can control every factor that determines whether you are selected when retrieved. Build deep topical authority on the questions your customers ask. Structure every page so a retrieval system can identify the question it answers in under two seconds. Implement Organisation and FAQ schema so your brand is a recognized entity, not an anonymous document. Keep content current with quarterly updates. Get external references pointing to your entity from industry publications, directories, and data aggregators.

The businesses that execute on all five signals simultaneously are the ones that dominate AI citation share in their category. Partial execution — good content without schema, or schema without topical depth — captures a fraction of the potential. The compounding effect of all signals working together is greater than the sum of the individual parts.

Key takeaways

• ChatGPT uses RAG to retrieve and synthesise live web content — approximately 60 million citation-triggering queries per day
• Five signals drive citation selection: topical authority, structural clarity, entity establishment, recency, and domain authority
• Commercial queries favor comparison content with structured data; informational queries favor deep editorial with original data
• Navigational queries are almost entirely driven by entity recognition — if ChatGPT does not know your brand, it cannot cite you
• Question-format H2 headers with direct opening answers are cited 2.4x more than equivalent content with generic headers
• Businesses without structured data are anonymous to AI — they cannot be cited by name
• Partial optimization captures a fraction of the potential — all five signals compound when executed together