What gets cited by ChatGPT? What the evidence actually shows

Some people say schema is the answer. Others say you need to rank first on Google, build more backlinks, publish longer articles, add an FAQ section, or repeat your brand name across the web.

The problem is that a single citation does not prove a theory.

A page appearing in one ChatGPT response could be the result of the wording of the prompt, the date the question was asked, the search system used for that particular response, the user's location, or simple variation between repeated runs.

So instead of treating isolated screenshots as proof, I reviewed the available research on generative search citations and designed a structured 100-query experiment to examine the issue properly.

The central question

The question is not simply: does a page rank on Google?

The more useful question is: does the page contain information an AI search system can retrieve, understand, attribute, and confidently incorporate into its answer?

Those are related questions, but they are not identical.

Traditional search engines generally return a ranked collection of pages. Generative search systems retrieve information and then construct a new answer from the material they select. That additional synthesis step changes what it means to win visibility.

A page does not only need to be discovered. Its information must survive retrieval, interpretation, selection, and answer generation.

How ChatGPT citations work

When ChatGPT uses web search, its response may include inline citations linking readers to external sources. However, not every query activates web search, and not every response contains citations.

That distinction matters. A business could be absent from a response because its website was not retrieved, because the retrieved page was not selected as evidence, because search was never activated, or because the answer was generated from another source.

This is why AI visibility should not be reduced to one screenshot or one prompt. A reliable test requires repeated queries, controlled conditions, and a record of exactly which pages were cited.

What the existing evidence suggests

Research into generative search is still developing, but several patterns are becoming clearer.

1. Extractable evidence matters

Generative systems need passages they can incorporate into an answer. Pages containing clear definitions, factual statements, numerical evidence, comparisons, procedures, and directly relevant explanations are easier to use than pages filled with vague marketing language.

Consider these two passages:

The first sentence may sound polished, but it provides little usable information. The second gives the system a clear definition it can quote, summarize, or use to support an explanation.

This does not mean every page should be written for robots. It means important information should be stated plainly enough that both people and machines can understand it.

2. Semantic alignment appears more useful than general relevance

A broadly authoritative page is not automatically the best source for every question. A page discussing digital marketing in general may be less useful for a question about Google Business Profile suspensions than a focused page explaining suspension causes, evidence requirements, and reinstatement procedures.

This suggests that page-level relevance matters. Topical authority is not just about publishing hundreds of loosely connected articles. It is about demonstrating that a specific page fully addresses the specific information need behind the query.

3. Structure helps a system interpret the page

Clear document structure makes content easier to process. Useful structural elements can include:

• A descriptive title and a clear primary heading
• An answer near the beginning
• Question-based subheadings
• Numbered procedures and comparison tables
• Definitions
• Visible authorship
• Publication and update information
• Supporting internal links

No single element guarantees a citation. The advantage comes from reducing ambiguity. The system should not have to guess what the page is about, who published it, or where the relevant answer is located.

4. Source identity and authority still matter

A source is more useful when the system can determine who is responsible for it. That can involve identifiable authors, organization information, consistent business details, third-party references, professional credentials, and a broader web presence that confirms the entity exists.

Structured data can help describe these relationships, but markup alone cannot create authority. Adding Organization or Person schema to a weak, anonymous website does not automatically make it trustworthy. The information in the markup must correspond with visible and externally supported information.

"Entity optimization is not a schema project. It is a consistency and corroboration project."

5. Technical accessibility acts as an eligibility layer

A page cannot contribute much if its primary content is unavailable to the retrieval system. Potential obstacles include:

• Blocked crawling
• Broken canonical tags
• Accidental noindex directives
• Server errors
• Content hidden behind interaction requirements
• Important text rendered unreliably
• Redirect chains and duplicate pages
• Conflicting page signals
• Slow or unstable delivery

Technical health should not be described as a guaranteed citation factor. It is better understood as part of citation eligibility. The page must first be accessible before its quality can be evaluated.

6. Google ranking is relevant but not decisive

It would be a mistake to conclude that traditional rankings no longer matter. Pages ranking prominently in search often possess the same characteristics an AI search system may value: relevance, authority, strong links, understandable content, and technical accessibility.

But overlap is not equivalence.

Generative search systems can select a source that is not the highest-ranking Google result. They may also exclude a highly ranked page when another source provides clearer evidence for the answer being constructed. That means ranking position should be measured as one variable, not treated as the explanation for every citation.

7. Third-party sources may have an advantage in recommendation queries

Commercial queries create a difficult trust problem. When someone asks, "What is the best company for this service?" a business's own website is naturally self-promotional. A third-party publication, professional association, review platform, or independent comparison may appear more suitable as support.

This does not make brand-owned content useless. It means businesses should build visibility across several source types:

• Their own website
• Reputable industry publications and local media
• Professional associations and relevant directories
• Expert interviews and independent reviews
• Original data that other websites can reference

Why single-query citation tests are unreliable

One of the most important findings from recent visibility research is that generative search results can vary. The same prompt can produce different citations across multiple runs. Small wording changes can also change which sources are retrieved and selected.

For example:

• "Best SEO agency in Fort Myers"
• "Who should I hire for SEO in Fort Myers?"
• "Which Fort Myers SEO company is best for a small business?"
• "Recommend a local SEO expert in Southwest Florida"

These questions express similar intent, but they are not identical retrieval tasks. A meaningful visibility study therefore needs repeated runs and multiple prompt variations.

Reporting that a company "ranks number one in ChatGPT" after one test is not a reliable measurement.

The 100-query experiment design

To investigate citation selection properly, the experiment is structured around 100 commercial-intent queries across five industries. Each query is tested three times in a new conversation, resulting in 300 response observations.

For every run, the following information is recorded:

• Exact query, date, and time
• Whether web search was activated
• The full response text
• Cited URLs and citation order
• Source type
• The claim each source supports
• Whether the citation accurately supports the associated statement

Every cited page is then evaluated for page-level, domain-level, and technical characteristics.

The study also records Google's top organic results for the same query. This makes it possible to compare cited pages with highly ranked pages ChatGPT did not select.

Variables being evaluated

The experiment examines whether citation likelihood is associated with: a direct answer near the beginning of the page, close alignment between the page and the query, topic-specific depth, clear headings, definitions and factual statements, tables and comparisons, original data, identifiable authorship, organization information, external references, structured data, crawlability, indexability, content freshness, Google ranking position, source type, domain authority indicators, and local relevance.

The study distinguishes correlation from causation. For example, if cited pages frequently contain a concise answer near the top, that does not automatically prove that moving an answer higher will cause ChatGPT to cite the page. The feature may coexist with other qualities such as stronger writing, better relevance, or more authoritative sourcing. A later controlled test would be needed to isolate the effect.

The difference between citation and influence

Citation visibility is only one measurement. A page may appear in the source list without contributing materially to the answer. Another page may supply the main definition, recommendation, or numerical evidence.

For that reason, the experiment separates three outcomes:

Retrieval

Was the page apparently considered by the system? This is difficult to observe fully because the complete retrieved candidate set is not always visible.

Citation

Did the page appear as a visible source? This is the most straightforward outcome to record.

Answer influence

Did the generated response clearly use information from the page? Influence can be evaluated by comparing the response with the cited content and identifying whether the page contributed a definition, number, recommendation, comparison, or procedural step.

This distinction matters because businesses do not merely want a link hidden among several sources. They want their information to shape the answer.

What businesses should do now

The evidence does not support chasing a single GEO trick. The more defensible strategy is to improve the entire path between discovery and answer generation.

Establish the entity

Make the business and its experts easy to identify. Use consistent names, biographies, service descriptions, contact details, and professional information. Connect relevant profiles and make sure important facts agree across the web.

State the answer clearly

Do not force readers or retrieval systems to dig through a long introduction before reaching the useful information. Open important pages with a concise explanation of the primary question, followed by the details, evidence, and qualifications.

Build page-level depth

Create pages that completely address real customer questions. Include explanations, costs, limitations, comparisons, processes, suitability factors, and related decisions. Avoid publishing multiple thin pages that repeat the same general information.

Provide evidence

Use original observations, named methodologies, screenshots, calculations, case data, and credible external sources. Claims become more useful when they can be inspected and verified.

Improve technical eligibility

Confirm that important pages are indexable, canonical, available in HTML, and internally linked. Remove retrieval barriers and conflicting signals.

Earn third-party corroboration

Build legitimate mentions outside the company's website. Expert contributions, interviews, association profiles, local coverage, and independently referenced research can help establish that the business is recognized beyond its own claims.

Measure repeatedly

Track multiple prompts, multiple runs, and multiple platforms. AI visibility should be treated as a changing distribution of answers, not a permanent ranking position.

What this means for SEO

AI search does not eliminate SEO fundamentals. It changes the final stage.

Traditional search optimization focuses heavily on whether a page can be discovered and ranked. AI search adds further questions:

• Can the system understand the page?
• Can it identify the source?
• Can it isolate a useful passage?
• Can it support a claim with that passage?
• Is the source suitable for the specific type of question?
• Will the information be incorporated into the answer?

This is why I describe the goal as Agent-Ready SEO. The objective is not simply to create content that ranks. It is to create a digital presence that an AI system can retrieve, interpret, verify, and confidently use.

The bottom line

ChatGPT citations are not controlled by one optimization technique. The available evidence points toward an interaction between relevance, source identity, extractable information, technical accessibility, external authority, and the specific wording of the query.

Google rankings remain important, but they are not a guaranteed ticket into an AI-generated answer.

That means publishing information worth using, making the information easy to interpret, clearly establishing who is responsible for it, and earning enough external corroboration that the source is credible beyond its own website.

The next phase of this research will apply the 100-query methodology across multiple commercial industries and compare ChatGPT citations with traditional organic results. The goal is not to manufacture another list of GEO "secrets." It is to create a repeatable measurement system businesses can use to understand where their AI visibility actually comes from.