One of the most common questions I get from clients starting to think about AI visibility is: ‘How does ChatGPT actually decide what to cite?’ The short answer is that it depends heavily on which mode ChatGPT is operating in — from memory or with web search — and the distinction matters a lot for how you build your strategy. I have watched clients waste months optimizing for the wrong mode because they never separated the two in the first place.

Two Modes, Two Very Different Mechanisms

ChatGPT operates in two fundamentally different citation modes:

  • Training-derived knowledge: The model answers from patterns it learned during pre-training. There are no live citations here — the model synthesizes from everything it absorbed. This mode dominates for general questions about well-established topics.
  • Browse mode (web search): When a user enables web search or asks about recent events, ChatGPT retrieves live web pages and cites them directly. This is closer to Perplexity’s default behavior.

Your strategy for each mode is different. Training-derived presence is built over time through consistent, widely-referenced content. Browse-mode citation is won through the same factors that earn featured snippets and top-of-SERP placement — authority, clarity, and query match.

In practice, most users do not consciously toggle between these modes — the product decides for them based on the query. Ask a question that sounds time-sensitive (‘best CRM in 2026’) and it is far more likely to trigger a web search than a question that sounds encyclopedic (‘what is a CRM’). That means the same brand can have strong training-data presence and weak browse-mode presence, or the reverse, and a business needs to check both.

What Training Data Favors

For training-derived answers, ChatGPT’s knowledge reflects the distribution of authoritative content on the web up to its training cutoff. That means:

  • Topics with more high-quality coverage in books, academic papers, and long-form journalism get more accurate, confident answers.
  • Brands and entities that appear frequently in independent, credible sources are named more reliably than those that only appear on their own websites.
  • Content that explains things clearly and completely — the kind humans share and reference — gets absorbed more thoroughly than thin, keyword-stuffed pages.

This is directly relevant to answer engine optimization. Building citation weight in training data means publishing genuinely useful content that independent sources link to and reference over time. There is no shortcut here — a page that goes up today does not retroactively appear in a model that was trained last year. This is one of the reasons AEO rewards patience in a way paid search never had to.

How Browse Mode Works

When ChatGPT uses its web search tool, it runs a query, retrieves a set of results, reads them, and synthesizes an answer with inline citations. The factors that influence which sources get selected and cited look like this:

  • Relevance to the specific query — the content must clearly address what was asked.
  • Source authority — established publishers, recognized industry sites, and high-domain-authority pages appear more often than newer or thinner sites.
  • Content clarity — pages that state their answer clearly and early, without burying it in navigation and promotional copy, are easier for the model to extract from.
  • Freshness — for time-sensitive queries, recency matters; an article from this year beats one from five years ago on the same topic.

I have run the same query through ChatGPT’s browse mode on different days and gotten different citations — not because the underlying facts changed, but because the retrieval layer pulled a slightly different set of results, or a competitor published something newer in the interim. That volatility is normal. It means one good citation is not a permanent win; it is a snapshot that needs maintaining.

What Does Not Work

Several things that used to work for SEO rankings actively hurt you in ChatGPT citation:

  • Thin pages optimized for a keyword but not actually answering the question.
  • Heavy keyword density without substantive content around it.
  • Content hidden behind modals, login walls, or heavy JavaScript rendering that a crawler cannot easily read.
  • Pages that describe a topic abstractly without providing a direct, citable answer.

I would add one more that clients are often surprised by: pages that try to answer too many questions at once. A page that covers ten loosely related sub-topics under one URL gives the retrieval layer a harder extraction job than ten focused pages, each answering one question completely. Comprehensive does not have to mean sprawling.

The Structural Signals That Help

ChatGPT’s retrieval layer benefits from the same structural clarity that helps any AI reader: a clear title that matches the query, a direct answer in the first paragraph, organized sections with descriptive headings, and FAQ-style Q&A where appropriate. For pages targeting specific answer opportunities, I structure the first 100 words as a direct answer to the target question — not a teaser, but the actual answer. More detail follows, but the citation-worthy statement comes first.

See conversational search optimization for more on structuring content around how people actually phrase questions to AI engines. One pattern worth calling out: people ask ChatGPT questions differently than they type Google searches. Google queries tend toward fragments (‘best schema plugin wordpress’); ChatGPT queries tend toward full sentences (‘what’s the best schema markup plugin for a WordPress site?’). Content written to answer the fragment does not always answer the sentence as directly, and that gap shows up in citation rates.

Entity Recognition in ChatGPT

ChatGPT is more likely to name your brand or recommend your product if it has encountered your entity in multiple independent contexts during training. A brand that appears in Wikipedia, major trade publications, credible review sites, and YouTube transcripts has a richer entity representation in the model than one that only appears on its own domain. This is why off-site presence is not optional for AEO — it is foundational.

At Salterra, I assess a client’s ChatGPT entity presence as a baseline before any AEO work starts — asking the model to describe the company or category, then auditing what it knows and where the gaps are. That gap analysis drives the content and outreach strategy. A common finding: the model knows the client’s name but attributes the wrong specialty to them, usually because a louder competitor with a similar name dominates the training data. Fixing that is a different job than building citation presence from zero — it is correcting a mistaken association, and it usually takes more third-party content, not less.

A Quick Way to Test Your Own Presence

Before building a strategy, I have clients run a simple audit themselves. Open ChatGPT and ask it to describe your company, your category, and your top three competitors, with web search turned off first, then on. Compare the two answers. If the training-mode answer is vague or wrong, that is an off-site authority problem — it takes months of independent coverage to fix. If the browse-mode answer misses you entirely, that is closer to an on-page and structural problem, and it can move faster because you are competing for that day’s retrieval, not for what got absorbed years ago.

Custom GPTs and Plugin-Style Retrieval

There is a third citation surface worth knowing about, even though it is smaller in volume: custom GPTs built by third parties, and any workflow where ChatGPT calls out to another tool mid-conversation. When a custom GPT is configured to search a specific set of sites or a specific API, the citation dynamics shift again — relevance to the narrow scope of that GPT matters more than general web authority. If your industry has a popular custom GPT that practitioners use as a first stop, understanding whether it favors certain domains is worth a few minutes of testing. It is a small surface today, but it behaves more like an old-fashioned directory listing than like open web search, and the rules for getting included are correspondingly more specific and static.

I would not build a strategy around this niche surface, but I do check it for clients in tightly defined B2B categories where a single well-known custom GPT has become a go-to reference tool. Being cited there once tends to be more durable than a browse-mode citation, because the underlying source list changes far less often than live search results do.

Putting It Together

To improve your ChatGPT citation: build an off-site entity footprint, write direct-answer content for the questions your target audience asks, make your pages structurally clean and crawlable, and earn mentions in sources the model weights as authoritative. None of this is a one-time project — training-derived presence shifts with each model update, and browse-mode presence shifts with every competitor’s publishing schedule. Treat it as an ongoing practice, the same way you would treat conventional SEO, and check back on your baseline audit every few months rather than only once. The AEO overview walks through each of these levers in detail.