How ChatGPT and Perplexity Choose Which Sources to Cite

Ask ChatGPT and Perplexity the same question and you will often get similar answers with different sources. For an agency trying to get a client cited, that difference is the whole game. If you understand how each engine finds and picks its sources, you know which levers to pull.

This article explains the two ways AI engines get information, how they decide what to cite, and a simple test you can run to see where a client stands today.

The two ways an AI engine knows things

Every AI answer draws on some mix of two sources: what the model learned during training, and what it retrieves from the web at the moment it answers.

Model knowledge is what the model absorbed from the text it was trained on. It is broad but fixed at a point in time, and the model cannot point to where a given fact came from. When an answer relies on model knowledge, there is often no citation at all.

Live retrieval is a search the engine runs while answering. It fetches pages, reads them and uses passages to support the answer. Retrieved pages are what appear as citations.

The balance between the two is the biggest difference between engines.

How Perplexity chooses sources

Perplexity is built around retrieval. For nearly every question it runs a web search, reads the top candidates and writes an answer with numbered citations attached to specific sentences. That makes it the most transparent engine to study and the quickest to respond to changes on a page.

In practice, Perplexity tends to cite pages that:

• Can be fetched quickly and read without JavaScript. Its crawler and user-triggered fetcher need to see the content in the HTML.
• Contain a passage that directly supports a sentence. Clear, self-contained statements are easier to attach a citation to.
• Are recent where recency matters. For prices, product details and news, fresher pages have an advantage.
• Cover the specific angle of the question. A page about "CRM pricing for small teams" beats a general CRM overview for that question.

Because Perplexity cites inline on almost every answer, it is the easiest engine to track. Tools and manual checks can record exactly which domains were cited for which question.

How ChatGPT chooses sources

ChatGPT works differently. It can answer from its model's knowledge alone, or it can decide to search the web and cite what it finds. The decision depends on the question. Time-sensitive questions, explicit requests for sources and questions about specific products or places are more likely to trigger a search.

That has two consequences for agencies.

1. When ChatGPT searches, the same page-level rules apply. It needs to be able to fetch the page; OpenAI documents a dedicated search crawler, OAI-SearchBot, for surfacing sites in ChatGPT search, and a separate ChatGPT-User agent for fetches made on behalf of a user. The page needs a passage that clearly supports the answer.
2. When ChatGPT answers from model knowledge, the brand's reputation is what counts. If a brand has been widely and consistently described across the web, the model is more likely to mention it, even without a citation. No single page change will move that quickly.

This is why a brand can dominate Perplexity citations and be largely absent from ChatGPT's unprompted recommendations, or the reverse. Tracking both is necessary.

How Google's AI features choose sources

Google's AI Overviews and AI Mode draw on Google's own index and ranking systems, so there is more overlap with traditional rankings than on other engines. AI Mode also breaks questions into many background searches, which gives pages that answer specific sub-questions more chances to be used. We cover this in Google AI Mode SEO.

What all engines have in common

Despite their differences, every engine follows the same basic pipeline: find candidates, read them, pick passages that support the answer, and cite them. That gives you a short list of factors that matter everywhere.

For crawler access, see AI crawler user agents. For readability, see can AI crawlers read JavaScript?. For quotability, see answer-first content. For corroboration, see entity SEO for LLMs.

A citation test you can run today

You do not need a tool to get a first read on a client's position. Here is a simple manual test that takes about an hour.

Step 1: Write 20 prompts

Pick 20 questions a real customer would ask, spread across the buying journey. Mix broad questions ("best accounting software for freelancers"), specific ones ("does X integrate with Stripe") and local or comparison questions where relevant. Write them the way people actually talk.

Step 2: Run each prompt on each engine

Use a clean session on ChatGPT (with search available), Perplexity and Google. For each answer, record:

• Whether the client's brand is mentioned.
• Whether the client's site is cited, and which page.
• Which competitors are mentioned or cited.
• Which third-party sites are cited, such as review platforms or publishers.

Step 3: Look for the patterns

After 60 answers, patterns appear quickly. Typical findings:

• The client is cited on Perplexity but not mentioned by ChatGPT. Pages are strong; the wider brand presence is weak.
• Competitors are cited via third-party reviews. The client needs presence on the sites engines trust for that category.
• The client is never cited for a whole topic. There is a content gap, or the relevant pages are not readable by AI crawlers.

Step 4: Repeat monthly

AI answers vary from run to run, so a single test is a snapshot. Running the same prompts every month turns it into a trend. Our guide to AI share of voice explains how to score the results consistently.

What does not decide citations

A few things are often claimed to matter more than they do.

• Schema markup alone. A large 2026 study found that adding schema did not increase AI citations for already-cited pages. It helps clarity, but it is not a switch. See does schema markup get you cited by AI?.
• An llms.txt file. Analyses of server logs show AI crawlers rarely request it. See does llms.txt actually work?.
• Keyword density. Engines read for meaning. Repeating a phrase does not make a passage more useful.

A worked example

Imagine a client that sells project management software for architecture firms. You run "best project management software for architecture firms" on each engine.

On Perplexity, the answer cites a comparison article from a design publication, two vendor pages and a review platform. The client is mentioned in the comparison article but its own site is not cited. That tells you the client's product page does not contain a passage that directly answers the question, even though a third party has vouched for it. The fix is on-page: a clear section explaining who the product is for and how it compares, with specifics.

On ChatGPT without search, the answer names three well-known general tools and does not mention the client at all. That is a model-knowledge gap. The brand is not discussed widely enough for the model to associate it with the category. The fix is off-page: presence on the review platforms, comparison articles and communities where architects discuss tools.

On Google AI Mode, the answer breaks the question into sub-questions about drawing management, client approvals and pricing, and cites the client's help article on client approvals. That tells you specific, well-structured help content is already working, and similar pages for the other sub-questions are the next step.

One question, three engines and three different fixes. This is why tracking each engine separately matters.

How to influence each engine

How Rankwyn helps

Rankwyn's AEO Citation Radar runs a fixed prompt set across ChatGPT Search, Perplexity and Google's AI features for every client, records mentions and citations against competitors, and shows which pages earn the citations, so the manual test above becomes a monthly report.

The bottom line

AI engines cite pages they can fetch, read and quote, and they mention brands the wider web agrees about. Perplexity responds to page improvements quickly; ChatGPT's unprompted recommendations follow the brand's reputation more slowly. Make pages accessible and quotable, build corroboration across trusted sites, and track every engine separately. For the full strategy, start with our AEO and GEO guide.