Which sources AI engines cite, and what to do about each
Review sites, forums, YouTube, Wikipedia, news, your own docs and listicles each signal something different to an engine. What each does and your move for it.
- Answers pull from seven kinds of sources, and each tells the engine something different about you.
- Engines do not cite the same mix, and every published share is a snapshot of a moving source pool.
- Third-party sources are the ones you cannot edit and therefore the ones worth working on first.
- Your own source list, ranked by how often domains appear, is the target list for outreach.
The source categories answers pull from
Answers are built from seven kinds of pages: review and rating sites, forums like Reddit, YouTube, Wikipedia, news and editorial, your own docs and blog, and comparison or listicle posts. Each tells the engine something different, so each needs a different move from you.
| Source category | What it signals to the engine | Your move |
|---|---|---|
| Review and rating sites (G2, Yelp and their peers) | Independent verdict with a score and a date attached | Claim the profile, correct stale facts, and get reviews from customers who actually use you |
| Reddit and other forums | Unprompted experience, including complaints and workarounds | Reply where the rules allow it, and fix the problem the threads keep repeating |
| YouTube | A demonstration, quotable through transcripts and descriptions | Publish demos with the claim stated in the spoken line and the description |
| Wikipedia | The entity definition, used for “what is X” prompts | Keep public facts consistent so the article stays accurate; never edit it for promotion |
| News and editorial | Independent coverage: launches, funding, rankings | Pitch corrections when they are wrong, and keep a dated press page on your own site |
| Your own docs and blog | How the product actually works, from the party that knows | Write answer-first sections; this is the one category you fully control |
| Comparison and listicle posts | A shortlist, which is what recommendation prompts ask for | Get into them: pitch the author, supply current specs, correct the ones that list you wrongly |
Engines do not cite the same sources
The mix changes by engine, and both published studies you can check today disagree with each other in instructive ways: ChatGPT leaned on Wikipedia and Reddit plus editorial sites, while Perplexity leaned on Reddit, LinkedIn and G2 for business queries.
The Peec AI analysis of 30 million sources, reported by Search Engine Land, found Reddit the most cited domain overall, with YouTube, LinkedIn, Wikipedia and Forbes behind it, and different favourites per surface: Wikipedia, Reddit and editorial sites on ChatGPT; Facebook and Yelp on Google's surfaces; Reddit, LinkedIn and G2 on Perplexity for business queries.
Semrush's own three-month study (230,000 prompts, weekly snapshots, over 100 million citations) in The Most-Cited Domains in AI tells a slightly different story, and the disagreement is instructive: ChatGPT's leaders were Wikipedia and Reddit with Medium, Forbes and LinkedIn behind; Google AI Mode favoured LinkedIn, YouTube, Reddit and Google properties, with Wikipedia in around 2% of its responses; Perplexity's leaders were Reddit, LinkedIn, NIH, Microsoft and Google.
Why ignoring third-party sources does not work
The usual excuse is “we cannot edit them”, and that is precisely why the engines use them: a source with nothing to sell is what a recommendation prompt is asking for.
You do not edit third-party sources, you get represented in them. That means accurate profile data on the review sites, coverage from publications that already write about your category, correct facts in the directory listings, and a reply in the forum thread where the complaint lives. It also means feeding writers the numbers they copy, because a spec or price they publish with your name on it becomes citable material on a domain you do not control.
The waiting version of this strategy does nothing. A review from 2022 that ranks you fourth will keep feeding answers until newer reviews exist on the same domain, and no amount of work on your own blog changes that. The only situation where the advice flips is a category with almost no third-party coverage at all: then your own documentation carries the load, and the writing rules in getting cited by ChatGPT and Perplexity are the whole job.
Turn the sources behind your answers into a target list
advisible's Sources view lists the domains that appear in the answers about you, with counts per domain, and exports the list as CSV. Ranked, that is your outreach list.
The reading order we would use: separate your own domain from the rest, then sort the third-party domains by how often they appear. The top of that list is where your category's answers actually come from, which is rarely the same as the outlets your PR plan already targets. Next, look at which of those domains appear in answers that skip you entirely, because a review site cited for a competitor and not for you is the fastest win on the page.
- Domains you can edit (your profiles, your listings): fix the facts this week.
- Domains that write about the category: pitch the author the current numbers.
- Domains nobody has touched for years: find the stale claim and request a correction.
Export it rather than reading it in the app once: the list changes as answers accumulate, and a CSV gives you something to diff next month. How the counts themselves are read without over-reading them is in reading citation and source data.
Common questions
Which source should I chase first?
The one your own answers already cite. Rank the domains appearing in your category's answers and start at the top, rather than chasing household names that never show up for your prompts.
Does a Wikipedia page about us help?
It helps on definitional prompts. Do not edit the article for promotion; keep your public facts consistent so the article stays accurate on its own.
Are YouTube citations real?
Engines can quote transcripts and descriptions, so a demo with a specific claim spoken and written in it can be cited. Vague titles and padded descriptions give nothing to extract.
Do the study percentages transfer to my niche?
Treat them as category-level hints. What matters is the source mix inside your own tracked prompts, which you can read directly from your answers.
How to get cited by ChatGPT and Perplexity
How engines choose sources at answer time, how to write a page worth quoting, and a 30-day procedure for earning citations in ChatGPT and Perplexity answers.
How to read AI citation and source data
Mention, citation and retrieved source defined in one sentence each, plus how to read citation share and a sources list without fooling yourself.
Using competitor gaps in AI answers to pick pages to write
A competitor gap is an answer that names a rival and leaves you out. How to read the gap list, choose which gap to work, and pick the page to write.
GEO vs SEO: what changes and what carries over
GEO optimises how AI engines describe and cite you. SEO optimises ranking. The buyer is the same and much of the work carries over. A side-by-side comparison.
Track this for your own brand
advisible runs your prompts on ChatGPT, Perplexity, Gemini, Grok and Claude every day, and every number shows how many answers it is based on.