Where Does ChatGPT Get Its Information? How To Find The 15-25 Sources That Control Your Category's AI Answers

ChatGPT gets its information from training data, live web search and licensed content, and only the live web is something you can influence. For a category question, the answers lean on a handful of third-party pages that recur across prompts, so the useful work is finding the 15 to 25 that recur most in your category.

This holds for category, comparison and recommendation questions, where ChatGPT searches the web and shows sources rather than answering from memory.

No public source gives ChatGPT's exact ranking weights, so the audit below measures what gets cited in your category today and does not predict what will be cited next quarter.

Where ChatGPT Gets Its Information: Training, Live Search And Licensed Content

Most explainers stop here. They tell you ChatGPT "learns from the internet" and leave you with nothing to do. That is the consumer answer, and it is thin. Here is the short version, then the part that matters for a marketing team.

ChatGPT's answers draw on three streams. Each one means something different for a brand.

• Training data: ChatGPT learns from a pre-trained corpus that stops at a knowledge cutoff. Anything written about your brand after that date is invisible to this stream, and nothing you publish today changes it.
• Live search: ChatGPT searches the web for newer material and builds the answer from the pages it finds. Live search is the only stream a brand can influence, because the pages it pulls are pages that exist right now.
• Licensed content: ChatGPT also draws on publisher content that has been licensed to it. A brand cannot edit that stream, though it can earn coverage in the publications involved.

Notice the pattern. One stream is frozen, one is closed, and one is open. Every sensible AI visibility plan puts its effort into the open one, which is why how to get cited in AI answers is mostly a question about live retrieval.

So the real question for you is not where ChatGPT gets its information in general. It is where it gets its information about your category, which pages it keeps reaching for when a buyer asks it to compare, recommend or shortlist. That list is shorter than you would expect, and it can be found.

Why Your Category's Answers Come From A Short List Of Third-Party Pages

Buyers do not ask ChatGPT about your brand first. They ask about the problem, or they ask for the best tools in the category. The answer is then built from whichever pages the search step returns, and those pages repeat.

The clearest evidence comes from a DerivateX study of ChatGPT with web search. Across 40 B2B SaaS categories, with one prompt per category, DerivateX counted 233 recommendations, and 188 of them carried a third-party citation. Vendor-owned sites supplied 51.1 percent of those cited recommendations (96 of 188). Niche or anonymous sites supplied 22.9 percent (43 of 188). Analysts, review platforms and major press together supplied about 16 percent, and G2 and Capterra supplied none.

Read the sample limits before you quote any of this. It is 40 categories and one prompt each, so it is a snapshot and not a census. The shares show how concentrated sources can be. They are not your category's numbers, and your own audit is the only way to get those.

The same study found a second pattern that matters more for planning. A single page can supply several brand recommendations within one answer, and the vendor-owned pages that were cited were mostly comparison posts that rank the vendor itself. One strong listicle can therefore carry a whole answer. Win a place on that page and you appear in the answer. Miss it and you are absent from every prompt that page serves.

That is why the list is short. A few pages do a lot of the work, and the same pages come back across many prompts. The 15 to 25 figure is the working size of the list TRM's audits target. It is not a measured constant, and it is not a law of how ChatGPT behaves. It is the size at which the list is still short enough to act on and long enough to cover most of what gets cited. Your data decides the real cut-off, and step 4 shows how.

If you want the wider concept this belongs to, AI search visibility covers what it covers and why citations are one half of it.

Does ChatGPT Cite Its Sources, And Which Ones Does It Cite?

Yes, when it searches the web, and not everything it retrieves makes it into the answer. In an Ahrefs analysis of 1.4 million ChatGPT prompts, about half of the URLs retrieved through search were ultimately cited, at 49.98 percent. The other half were read and dropped.

That gap is the first thing to understand. Being retrieved is not the same as being cited, and only the citation puts your name in front of a buyer.

The second finding is bigger. In the same analysis, results from the general search index were cited 88.46 percent of the time, against 1.93 percent for results from the Reddit channel. ChatGPT treats its retrieval channels very differently, and a thread that surfaces through the Reddit channel is rarely the thing that ends up in the answer.

Two more findings help you decide what to do with this. Ahrefs found that cited pages' titles match the prompt and its sub-queries more closely than uncited titles do. Cited pages in search results also have a median age of about 500 days, while the uncited ones are mostly very young.

Put those together and the practical reading is plain. A page that names the question in its title has an edge, and a page that has been live for well over a year has another. A post you publish this week to chase a prompt starts behind. A page that has aged into a reference starts ahead.

One limit applies to everything above. A citation shows where an answer leaned. It does not show that the answer was correct, and a cited source can be wrong, thin or out of date. Count citations as a map of influence, never as a verdict on truth.

If you want to watch this over time instead of by hand, ChatGPT answer monitoring tools are the next step.

How ChatGPT And Perplexity Cite Different Sources

The engines do not share a source list, and blending them hides that.

Ahrefs Brand Radar data across 3.1 million US queries shows that Reddit makes up 24.6 percent of Perplexity citations, YouTube 14.2 percent and Wikipedia 7.2 percent. The consequence for your audit is direct. Run each engine separately, and never average them into one list.

Set that beside the ChatGPT figure from the last section. In the Ahrefs ChatGPT analysis, results from the Reddit channel were cited just 1.93 percent of the time. On Perplexity, Reddit is the largest single source in the data. Same site, very different treatment. A tactic that earns citations on one engine can do close to nothing on the other.

This matters most for user-generated sources. Community threads and video are answer sources on Perplexity at a scale that would be easy to miss if you only tracked ChatGPT. If your buyers use more than one engine, a Reddit presence is a real lever for one and a weak one for the other, and only separate runs will show you which.

So the rule for the rest of this guide is simple. Every step below is run once per engine. You end up with one ranked list per engine, and you compare them afterwards. If Perplexity is part of your mix, Perplexity visibility tracking is where to look for ways to record it.

Step 1: Build A Query Set That Mirrors How Buyers Ask

A query set is the fixed list of full prompts you run through every engine, written the way a buyer in your category would type them. Everything downstream depends on it. A weak query set produces a confident, wrong source list.

TRM's working practice is a first baseline of about 25 prompts. That is enough to see recurrence and small enough to run by hand in an afternoon.

1. Define the category in the buyer's words. Write the category as a customer would describe the job, not as your product page does. If your own product vocabulary is the only vocabulary in the set, the audit will measure your marketing and not your market.
2. Pull the questions buyers type. Collect them from sales calls, support tickets, search suggestions and community threads. Sort them into four prompt types, shown below.
3. Write each as a full prompt. Use complete sentences with the context a real buyer would add, such as team size or budget range. Short keyword strings do not behave like real prompts.
4. Balance the four types. Spread the 25 prompts across the types so no single one dominates the counts.

These are the four prompt types, each with an example for a placeholder category, "project management software for agencies".

• Problem prompt: "How do agencies stop projects running over budget without adding more meetings?"
• Comparison prompt: "How does tool A compare with tool B for a 30-person agency?"
• Best-of prompt: "What is the best project management software for a creative agency?"
• Alternatives prompt: "What are good alternatives to tool A for agencies that need time tracking?"

Semrush is a practical place to start the list. Its keyword research tools show the questions people already type about a category, which gives you the first draft of the query set before you add the comparison and alternatives prompts.

Wording matters more than most teams expect. Two prompts with the same intent can pull different pages, so build a few variants for your most important questions. Prompt variation testing covers how to do that without inflating the set. If you want the whole baseline to run automatically, see how daily prompt runs handle a 25-prompt setup.

Step 2: Run Every Query Through Each Engine And Record The Citations

This step is a method, not a tool review. You need a sheet, a few hours and some discipline.

1. Open each engine with web search on. Without search, ChatGPT answers from its training data and there are no live sources to record.
2. Run every prompt in a fresh session. Carry-over from earlier prompts in the same chat can bias the sources, so start clean each time.
3. Run each prompt twice. Citations shift between runs. A source that shows up once may be noise, while a source that shows up on both runs is a candidate for the list.
4. Record every cited URL. Log the prompt, the engine, the run number, the date, the position in the answer and the full URL, not just the domain.
5. Tag the source type as you go. A quick label now saves a second pass in step 3.

Use one row per cited URL, per prompt, per run. The layout below is enough.

Keep the full URL for a reason. Recall the Ahrefs finding that cited pages carry titles close to the prompt and its sub-queries. When you later open the top pages, the page-level detail tells you why that page won, which a bare domain never will.

Doing this by hand for 25 prompts, two runs and three engines is 150 runs. That is fine for a first baseline and painful as a habit. When you are ready to automate the recording, citation tracking tools are built for it.

Step 3: Classify And Count The Sources

Now you turn a pile of URLs into a picture. Two jobs here, classification and counting.

Classify each cited domain into one of five types. This is the same kind of hand classification the DerivateX study used.

• Vendor-owned: A company's own site, often a comparison or listicle post that ranks the company itself.
• Niche or independent site: A small publisher, blog or specialist site with no large brand behind it.
• Media: A news or trade publication with editorial staff.
• Analyst or review platform: A research firm, review site or directory.
• Community: A forum, discussion board, video platform or Q&A site.

Then count recurrence, per domain and per page. Recurrence share is the number of answers that cite a domain, divided by the total number of answers. Count a domain once per answer, even if it appears twice in the same answer.

Here is a worked example. The numbers are invented for the arithmetic, so do not read them as a finding. Say you ran 25 prompts across three engines, which gives 75 answers. Domain A is cited in 30 of those answers.

• Recurrence share for Domain A: 30 divided by 75, which is 40 percent.
• Recurrence share for Domain B: 24 divided by 75, which is 32 percent.
• Recurrence share for Domain C: 21 divided by 75, which is 28 percent.

Notice that the shares add up to more than 100 percent. That is correct, because one answer can cite several domains. Do not normalise them to 100.

Do the same count per page as well as per domain. A domain can look dominant because of a single URL, and the page is the thing you will actually try to appear on. For a category-wide view of the same counting, see category-level AI visibility tracking.

Step 4: Rank The 15 To 25 Sources That Control The Answers

Sort the domains by recurrence share, highest first. Then add up how much of all citations the top of the list covers as you go down. The point where that curve flattens is where your list ends.

Use this cut-off rule. Stop where adding the next source adds less than a stated share of all citations. Pick the stated share before you look, such as 2 percent, so the line is not drawn to flatter the result.

Carry on with the invented example. Suppose the 75 answers produced 200 domain citations in total, and Domains A to H account for 120 of them, with the rest spread across about 60 smaller domains. Each source's share of all citations is its count divided by 200. Cumulative coverage is the running total.

With a 2 percent rule, the list stops at Domain G, because Domain H adds only 1.5 percent. Seven sources cover 58.5 percent of everything cited. That is a small, invented set, so the cut lands early.

Real categories have a longer head. In TRM's audits the line tends to fall at 15 to 25 sources, which is why that range is in the title. Treat it as a typical landing zone, not a target. If your curve flattens at 11, stop at 11. If it flattens at 28, keep going. Cutting at a fixed number throws away the one thing the audit measures, which is where your category's concentration actually ends.

Once you have the ranked list, the work becomes concrete. Every source on it is a place a buyer's answer is being decided. That is the point where Generative Engine Optimization turns from a concept into a task list.

Step 5: Judge Which Sources You Can Realistically Earn A Mention On

A ranked list tells you where the answers are decided. It does not tell you where you can get in. Some sources take a phone call. Some cannot be bought at any price.

Use this decision rule. Score each source on reachability, cost and fit, then start with the sources that have the highest recurrence and are the most reachable. Leave the rest for later, and drop any source that is not earnable at all.

Sort each source into one of five routes.

The own-content row deserves a note. In the DerivateX study, the vendor-owned pages that were cited were mostly comparison posts that rank the vendor itself. Vendor-owned sites also supplied the largest share of cited recommendations. So brand-owned comparison content is a lane of its own, and it is one you fully control. Do not treat it as a fallback.

You do not need a mention on every source. Cumulative coverage tells you why. If the first five sources cover half of all citations, those five matter far more than the last ten put together. Pick by recurrence and reachability, and ignore the tail until the head is done.

For scoring and choosing between outreach and outranking, tools that suggest where to publish for citations help once your list is long. For the pitching itself, Earned Mentions is the work of getting onto those pages.

Common Mistakes In Citation-Source Audits

Most bad audits fail in the same few ways. Each one produces a list that looks authoritative and points the wrong way.

• One run per prompt. Citations shift between runs, so a single run mistakes noise for signal. Run every prompt at least twice and count a source only when it recurs.
• Blending engines. Perplexity and ChatGPT cite very different sources, so an averaged list describes neither. Keep one list per engine and compare them afterwards.
• Treating one study's shares as your category's. The DerivateX and Ahrefs figures are samples with limits, and the DerivateX one is 40 categories with one prompt each. Use them to understand the shape, then measure your own.
• Counting mentions instead of cited URLs. A brand named in an answer is not the same as a source cited in it. Record the URLs, because the URLs are the pages you can act on.
• Stopping at the domain. A domain can hide one dominant page. Open the top pages and note what they are and why they were chosen.
• Building the query set from your own product words. Your vocabulary is not your buyer's. Take prompts from sales calls and support threads, and test your wording against theirs.

Every third-party study in this article has sample limits, and each one is stated where it appears. That is a feature of the evidence that exists, not a flaw in the method. It is also the best argument for running your own.

What To Do After The Audit

A ranked list sitting in a sheet changes nothing. Turn it into a plan with an owner and a date.

1. Pick the first five reachable sources. Take the highest-recurrence sources you scored as reachable in step 5. The output is five named sources.
2. Assign an owner and a route for each. Decide who pitches, who writes and who maintains the profile. The output is a name next to every source.
3. Set a re-run date. The audit is a snapshot and the answers move. For a fast-moving category, monthly is TRM's suggestion, offered as a judgement and not a rule. The output is a date in the calendar.
4. Watch for citation loss. A source that cited you last month and does not now is the earliest warning you get. The output is a flagged list of lost citations.

If you want the wider picture first, run an AI visibility audit shows where this citation work fits. If you want the ongoing execution handled, the GEO + SEO Growth Programme covers the re-runs and the follow-through.

And if you are not ready to talk to anyone yet, the free AI Visibility Report is a lower-pressure way to see where you stand before you commit to a full audit.