Does ChatGPT tell you its sources?

Only when it searched the web to answer. In that case it shows links or a sources panel. When it answers from training data alone — common for established categories — there are no citations to see, and the influence came from pages it read during training rather than at answer time.

How does an AI decide which sources to cite?

For live answers it runs a search, retrieves a handful of pages it judges relevant and trustworthy, and composes the answer from those. Pages that state facts plainly, cite their own sources and are widely referenced elsewhere are more likely to be picked up. Research on generative engine optimization points to primary sources, quotations and statistics as the strongest factors.

What is the owned-versus-third-party split?

Of the pages behind an answer about your category, how many you control (your site, your docs, your blog) versus how many you don't (review sites, forums, press, competitors). A low owned share means publishing on your own site moves the needle slowly; the work is getting cited on the third-party pages instead.

Can I see sources for Google AI Overviews?

Yes. The Overview places links beside the claims they support, and expanding the block shows the fuller list. These are the pages Google's system used for that specific answer.

See which sources an AI answer is built from

Perplexity lists its sources inline. ChatGPT and Gemini show them when the answer used live search — as links or a citation panel. Google's AI Overviews link the pages beside each claim. Collect those URLs for your buyer questions, then sort them into pages you control and pages you don't. The split tells you whether your visibility problem is fixable on your own site.

Key takeaways

Perplexity is the clearest: every claim carries a numbered citation. Use it to get the fastest picture of who the engines are reading.
ChatGPT and Gemini cite sources only when they ran a live search for that answer. When they answer from training alone, there is nothing to inspect.
Google AI Overviews place links next to the claims they support; hover or expand to see the full list.
Sort every cited URL into 'we own this' and 'we don't'. A category dominated by review sites, forums and competitor pages is won off your own site, not on it.
In one GeoShark run on nike.com, the brand's own site was cited twice out of 341 sources the engines read — 0.6%. Owned pages are rarely the deciding factor.

Whether your brand is named is the score. Why is in the sources — the web pages the engine read to write the answer. If you only track the score, you know you have a problem but not what kind, and the fixes for the different kinds do not overlap.

Where each engine shows its sources

Perplexity — every sentence carries a numbered citation, and the full list sits under the answer. This is the easiest place to see the citation picture, so start here even if Perplexity is not your priority engine.
ChatGPT — shows links or a sources panel when it searched the web for that answer. If it answered from training data alone, there is nothing to click, and the influence came from what it read during training rather than at answer time.
Gemini — same rule as ChatGPT: citations appear for search-backed answers, and not for answers it produces from memory.
Google AI Overviews — links sit next to the claims they support. Expand the block for the full set. These are exactly the pages Google's system used for that answer.

A practical consequence: to see sources at all, phrase questions so the engine wants to search. Current, specific, comparative questions ("best X in 2026", "how does A compare to B") trigger a search more reliably than broad ones.

Collect the URLs

For each of your buyer questions, run it on the engines that cite, and copy out every source URL — domain and page. Put them in one list with the question beside each.

You are building a picture of the domains that decide answers in your category. After twenty or thirty questions the same names keep coming up. Tally them by frequency and a pattern appears fast: a category usually has ten to twenty pages doing most of the work, and they are rarely the pages a brand would guess.

Sort them into two piles

Owned — your website, your documentation, your blog, anything you can edit today.
Third-party — review sites, directories, forums, news, roundups, competitors' pages.

The ratio is the finding. If the answers about your category are built mostly from third-party pages, then writing more on your own site changes little in the short term — the engines are barely reading it. The work is to get accurate, favourable mentions onto the third-party pages they are reading.

In one GeoShark run on nike.com, across ten buyer questions and forty answers, nike.com was cited twice out of 341 sources. The most-read sources were YouTube, Google's own properties, a shoe-review site and Reddit; the highest-placed brand site in the whole run belonged to a competitor. A brand that size, and its own pages were 0.6% of what the engines read. Most categories look like this — the owned share is small, and knowing the number changes where you spend effort.

What the split changes

A high owned share is the good case: your pages are in the retrieval set, so improving them — clearer answers, cited facts, structured data — feeds through to the answers.

A low owned share means two tracks, and only one of them is on your website:

On-site, so that when an engine does reach your page it can lift a clean answer. Necessary, but slow to show up on its own.
Off-site, the larger lever — being mentioned, reviewed and listed on the pages that actually make it into the answers. Directory and review profiles, inclusion in the "best X" roundups, answers in the threads where buyers ask.

You cannot tell which case you are in, or which of the two tracks to weight, without reading the sources. That is the whole reason to look past the score.

Why ChatGPT doesn't mention your company — the reasons, in order
How to track your brand in AI search answers — the full picture
GEO vs SEO: what changes for a small agency — why the sources are the part SEO work does not cover

FAQ

Does ChatGPT tell you its sources?: Only when it searched the web to answer. In that case it shows links or a sources panel. When it answers from training data alone — common for established categories — there are no citations to see, and the influence came from pages it read during training rather than at answer time.
How does an AI decide which sources to cite?: For live answers it runs a search, retrieves a handful of pages it judges relevant and trustworthy, and composes the answer from those. Pages that state facts plainly, cite their own sources and are widely referenced elsewhere are more likely to be picked up. Research on generative engine optimization points to primary sources, quotations and statistics as the strongest factors.
What is the owned-versus-third-party split?: Of the pages behind an answer about your category, how many you control (your site, your docs, your blog) versus how many you don't (review sites, forums, press, competitors). A low owned share means publishing on your own site moves the needle slowly; the work is getting cited on the third-party pages instead.
Can I see sources for Google AI Overviews?: Yes. The Overview places links beside the claims they support, and expanding the block shows the fuller list. These are the pages Google's system used for that specific answer.

Sources

GEO: Generative Engine Optimization (Aggarwal et al.) — arXiv, 2024
AI features and your website — Google Search Central, 2026

GEO
citations
AI search

How to see which sources an AI answer is built from