An AI citation source report groups the websites or pages that AI answers linked to as sources, and shows how often each one appeared in the answers the report collected. Reddit, YouTube, Wikipedia, LinkedIn and Google properties show up near the top of many published studies of the most cited domains AI engines use, but the order changes by engine, by question set and by date. So the top row of a report tells you what dominated that sample, and the percentage next to it only means something once you know what it was divided by.
If you are looking at one of these reports for the first time, it's normal for the numbers to feel a bit slippery, because two reports can show very different shares for the same site and both can be right. This guide covers what a report counts, what a few recent studies found, how to read the percentages, and which rows deserve your time.
What does an AI citation source report show?
A report starts with answers. Someone, usually one of the AI visibility tools or a research team, runs a set of questions through one or more AI engines and saves the answers. The report then pulls out every source link in those answers and adds them up by page or by domain.
That means the report describes the answers it collected. Unless the method says otherwise, it is not a count of everything people ask on an AI platform, and it isn't weighted by how many people ask each question. If the same question was run ten times, that's ten answers, not ten different people.
A few terms get mixed up a lot, so it helps to pin them down before you read any table.
| Term | What it means | What not to assume |
|---|---|---|
| Citation | A source reference attached to an AI answer, usually a link | That the cited site is recommended or praised |
| Citing answer | An answer that contains at least one citation to a source | That two links to the same site make two citing answers |
| Cited page | The exact page that was linked | That every page on the site gets cited equally |
| Brand mention | The answer names your brand | That the answer links to your website |
The brand mention and the citation are separate signals. An answer can name you without linking to you, and it can link to you without naming you. If you want to go deeper on that split, our guide on what counts as an AI citation covers it.
Domain grouping changes the leaderboard
Reports also differ in how they group rows. Some group by root domain, some by full hostname, and some by something else. Profound's documented domain leaderboard groups by root domain, while the Ahrefs tables show hostname-style rows. Those are not the same unit, so you can't quietly merge one into the other.
An English-language Wikipedia row is not every language edition added together, and Google and YouTube are separate rows in the Ahrefs tables even though one company owns both. A big site also holds many kinds of pages, such as articles, community threads, company profiles and comparison pages, and the domain name doesn't tell you which of those earned the citations.
Before you read any ranking, find three things: what sample was collected, what event was counted, and how rows were grouped.
Which sites do AI engines cite most?
There isn't one list of the most cited domains AI engines use that holds for every engine, topic and month. What you can do is look at a few published studies as bounded examples, each with its own sample and date. Please don't treat them as interchangeable estimates of some market share.
Google AI Overviews, September 2026
Ahrefs says it looked at domains cited across more than 3 million US queries covering all topics, and its table is labelled September 2026. The share it reports is a domain's citations divided by the combined citations of the top 50 most cited sources. That is a share of a limited leaderboard, not a share of everything on the web and not the percentage of answers that include the site.
| Rank | Site | Reported share of top-50 citations |
|---|---|---|
| 1 | YouTube | 22.9% |
| 2 | 18.5% | |
| 3 | 10.1% | |
| 4 | 8.8% | |
| 5 | 5.6% |
In Ahrefs' US Google AI Overviews snapshot labelled September 2026, YouTube led the top-source table with 22.9% of citations among the top 50 sources. That doesn't mean YouTube appeared in 22.9% of all AI answers. The page refreshes monthly, so today's table may differ from this one.
Google AI Mode, November 2026
Ahrefs publishes a similar Google AI Mode table, updated on November 3, 2026 and labelled November 2026, using broad US queries across all topics. In that table Google itself led with 23.4%, followed by Reddit at 18.1%, YouTube at 17.0%, Facebook at 8.3% and English Wikipedia at 5.9%.
The two tables carry different month labels, so this isn't a controlled same-period comparison of the two features.
What changed within weeks in 2025
Semrush published a three-month study on November 10, 2025. It looked at about 230,000 prompts across ChatGPT Search, Google AI Mode and Perplexity, with more than 100 million citation observations. The team took 13 weekly snapshots between July 14 and October 12, 2025 and looked at the 25 most frequent domains each week.
ChatGPT cited Reddit in close to 60% of prompt responses in early August 2025, and that fell to around 10% by mid-September. Wikipedia dropped from roughly 55% of ChatGPT prompt responses to under 20%.
Those are the share of responses that included the site, so they can't be set next to the Ahrefs top-source shares. Semrush says its data doesn't establish why the shifts happened, so it's best not to invent a reason. The useful lesson is that a leader can lose most of its presence in a matter of weeks, which is why a 2025 ranking shouldn't be quoted as a current one.
The long tail on Perplexity
Featured looked at 405 user-run audits from June 2 to August 21, 2026, with 1,983 buying questions and 22,881 citations across 11,499 different domains. The report talks about AI search in general, but its method says the figures come from Perplexity, so they shouldn't be labelled cross-engine.
YouTube made up 5.5% of citations and Reddit 3.7%, together 9.2%. The 100 most cited domains accounted for less than a quarter of citations, the top 1,000 accounted for 45.1%, and 76% of the distinct domains were cited exactly once.
That is a good counterweight to the big-name leaderboards, since large platforms can lead a table while most citations are spread across thousands of other sites.
What the studies do and don't tell you
Large community, video, reference, professional and platform-owned sites come up again and again. They don't give you a current, comparable top ten for every engine (our guide to the third-party sources ChatGPT cites most covers one engine in more depth), and mixing studies to build one would create something none of them measured. A benchmark is a dated observation with a method behind it, not a permanent list of sites every team must chase.
What are the top review sites cited by ChatGPT?
People ask for the top review sites cited by ChatGPT a lot, and the honest answer depends on the study. Two pieces of research reach different results because they measured different things.
AuthorityTech released a Machine Relations Index on November 3, 2026 that looked at ten review and ratings sites. Among the ten it picked, G2 was the most cited in ChatGPT runs.
| Review site | Runs citing the site | Share of 8,343 observed runs |
|---|---|---|
| G2 | 64 | 0.77% |
| GetApp | 17 | 0.20% |
| Capterra | 15 | 0.18% |
| Trustpilot | 8 | about 0.10% |
A few things sit next to those numbers. The ten sites were the author's own selection, the runs used OpenAI's API with web search on and not the signed-in ChatGPT app, and the prompts came from a library of category and tool questions, not real conversation logs. The Index was also created by the founder of AuthorityTech, so it isn't an independent audit. A safe way to put it is that G2 appeared in 64 of 8,343 observed API runs with web search enabled, the highest among the ten sites in that analysis.
Trustpilot commissioned research from Seer Interactive, with testing in March 2026. It analyzed 804,491 AI responses covering 1,926 US-based brands across eight industries, and it names Trustpilot as the most cited review platform among those tested, which included BBB, G2, Yelp and Capterra. The company that paid for the study has a commercial interest in the result, and the report labels its engines inconsistently, naming Google AI Overviews in its methods but a Google AI Mode column in the review chart.
So which one is right? Probably both, for what they measured. One used tool questions and ten chosen sites, and the other used nearly two thousand brands across eight industries. G2, Trustpilot, Capterra, GetApp, Software Advice, BBB and Yelp are all documented review-related sources, and which one leads depends on the questions and the method. Averaging the two or ranking the sites together wouldn't make sense.
Neither study shows that building a review profile causes more citations. For how review pages fit into a plan, see our piece on turning G2 and Capterra into citation channels.
What do citation share percentages mean?
This is the part of a report that trips people up most. There is no standard citation share formula, and different providers use the same label for different calculations.
The common ones are raw citation count, share of citation occurrences, answer incidence (answers citing the source out of all analyzed answers), share of source-bearing answers, a restricted top-source share, and a model-averaged share.
Providers really do define these differently. Topify's "Used" percentage is the share of all analyzed responses that cite a source. MentionFlow divides the answers citing your domain by the answers that have any citations, so three links to one website in one answer count as one citing answer. Profound averages each domain's share within each model across the models, so its raw count and share columns don't always sort the same way. The Ahrefs table uses a top-50 denominator even though it calls the figure mention share.
The habit to build is to look for a definition, tooltip or methods note before you read a percentage. A familiar column name is not enough.
One source, three correct percentages
Here is a made-up example. It is a teaching example only, and not DeepSmith data or a published finding. Say a report has 100 collected answers, 80 of them with at least one citation, and 250 citation occurrences in total. One source is cited in 20 answers, with 30 citation occurrences, across 6 different pages.
| Metric | Calculation | Result |
|---|---|---|
| Share of citation occurrences | 30 divided by 250 | 12% |
| Answer incidence | 20 divided by 100 | 20% |
| Share of source-bearing answers | 20 divided by 80 | 25% |
| Average citations when present | 30 divided by 20 | 1.5 |
| Distinct cited pages | Count of different pages | 6 |
These numbers don't conflict. They describe different things, and a report might show any of them next to a column called "share." If a provider counts repeated references differently, the occurrence total would change too, so state the counting rule instead of assuming this example works the same way everywhere.
Why percentages can add up to more than 100%
When a measure is based on answers, one answer can cite several domains and counts toward every one of them, so the percentages can add up past 100%. A table showing only a few rows will total less than 100% on a citation-occurrence share. Small rounding gaps are normal, and a big gap usually means different denominators.
When counts go up and share goes down
A source can gain citations and still lose share. Say it had 20 citations out of 200 in one period (10%) and 30 out of 600 in the next (5%). It got 50% more citations and lost 5 percentage points of share, because the total grew faster. The table alone can't tell you why, and it doesn't prove the engine trusts the source less.
Percentage points are not percent growth
A move from 8% to 10% is a rise of 2 percentage points and a relative increase of 25%. A leaderboard arrow showing a site moved up two places is a change in rank, which is neither.
Why do citation source reports disagree?
When two reports don't match, it's tempting to decide that one of them is wrong. Usually the reports are just measuring different things. These are the differences worth checking, in rough order of how often they explain the gap.
Prompt selection. A broad set of general queries, a library of software questions and user-chosen buying questions give different source mixes, so you can't compare their percentages as if only the provider changed.
Platform and interface. ChatGPT Search, an OpenAI API request with web search, Google AI Mode, Google AI Overviews and Perplexity are different places to collect answers. AuthorityTech says its ChatGPT runs used the API and not the signed-in app, and that detail belongs right next to its numbers. Our guide on API versus chat interface citations goes into why that gap shows up.
What counts as a citation. OpenAI's help documentation says the Sources view in ChatGPT can show cited sources and other relevant links. Its API documentation separates inline citations from the full list of URLs consulted, and that second list is often larger. A tool counting one isn't measuring the same thing as a tool counting the other.
Time window and geography. The Semrush study showed big changes inside a few weeks, and the Ahrefs and Seer examples are US samples, so a live leaderboard and a fixed study from last year shouldn't be expected to agree.
Weighting, duplicates and denominators. A provider might pool every observation, average across models, or count hostnames, root domains and repeated links in different ways. A share of the top 50 sources is not a share of all sources, so Ahrefs' 22.9% for YouTube in AI Overviews and Featured's 5.5% for Perplexity buying questions don't measure the same quantity.
Before you decide a report is inaccurate, line up the sample, the surface, the period and the formula. Most disagreements go away once you do.
Which source rows deserve attention?
Once you know what the percentages mean, the next question is what to do with the table. This is an editorial rule of thumb, not a tested scoring model, and it only applies to rows that are already in your report. If you want to go out and build a source list from scratch, that is a separate job, which our guide to running a citation source audit walks through.
The rule is to give attention to rows where the cited page is relevant to the questions you track, the pattern shows up more than once, and your team has a legitimate way to act. Raw share alone isn't enough. Ask three questions of each row.
Relevance. Does this source support an answer to a question that matters to your business, or is it just common across the web? A smaller domain that appears across several of your buying questions can be more useful than a giant one with nothing to do with them.
Recurrence. Does it appear across several relevant answers, prompts or dates, or did it pile up links inside one response? A site that appears once isn't a trend yet.
Agency. Is there a truthful, policy-compliant way to improve the information, publish a useful resource of your own or fix a mistake? A highly cited site is not automatically somewhere you can or should get placed.
No universal minimum share or count makes a row worth acting on, so I wouldn't invent one. These patterns can help you sort what you see.
- A smaller domain recurs across several buying questions. Often more useful than a bigger but unrelated row.
- A source keeps supporting a wrong or outdated description of you. Fix the facts before chasing more volume.
Domain first, page second, answer third
A useful way to move through a row is in three steps. The domain shows where citations pile up, the cited page shows what information got picked, and the answer shows what claim the citation backed up. A company profile, an independent comparison and a complaint thread are not the same opportunity, even on one domain. Our explainer on URL-level versus domain-level tracking shows what each level hides.
What can a citation source report not tell you?
A source report is observational evidence. It is not a window into an engine's reasoning, and it's worth knowing where it stops.
OpenAI's documentation confirms that ChatGPT can search the web and give links to sources, and that search results and citations can be incomplete, outdated or incorrect. Google's documentation says AI Overviews and AI Mode may use query fan-out, issuing related searches across subtopics and data sources, and may use different models, so the links they show can differ. A supporting page has to be indexed and eligible to show in Google Search with a snippet, and sites that appear in these features are included in Search Console's Web search reporting, per Google's AI features guidance. None of that publishes a universal domain ranking or a citation formula.
A citation also doesn't show that the answer feels positive about your brand, that you made a shortlist, that anyone clicked, that the page was understood correctly, or why the source was selected. It doesn't show that publishing similar content or appearing on the same domain will bring the same result.
Vendor words like authority and trust shouldn't be turned into confirmed engine scoring. Featured used Moz Domain Authority to describe domains, which isn't evidence that engines use the score.
A short checklist for your next report
Before you treat a source percentage as something to act on, see whether you can answer these. If the provider doesn't say, write "unknown" next to it and don't guess a formula from the column name.
- Which engines and interfaces were measured, and which questions were included?
- What dates were the answers collected, and how many observations are behind the table?
- Are answers without any citation counted in the denominator?
- Does the report count inline citations, source panel links or consulted URLs?
- Are rows pages, hostnames, root domains or owner groups?
- Are shares pooled or averaged across models, and does the table cover all sources or only the leading ones?
- Does a change column show rank, citation count, relative change or percentage points?
If you use one of the AI citation tracking tools, the same checklist applies to your numbers. DeepSmith's AI Visibility overview shows the sources AI cites most alongside your mention rate, citation rate and share of voice, and a free trial lets you check how your own report reads with these questions in mind.
A source report is most useful when it explains where relevant answers get their evidence, so what matters is what was counted, which pages supplied the information, and which observations deserve your attention.
Related guides
- Citation Tracking Tools With API Access: What to Ask For and What You Can Build: Know which API questions to ask a vendor (rate limits, fields, history) and what you could realistically build with the data.
- What to Look For in a Citation Tracking Tool as an Agency: List the citation-specific features agencies should test for, like per-client source reports, exportable evidence and alerting across accounts.
- How to Combine Citation Tracking With Analytics and Content Operations: Wire citation data into your analytics and content planning so cited pages, traffic and publishing work sit in one view.
- How to Find the Sources AI Engines Cite Most in Your Industry: Build a list of the domains and pages AI engines lean on for your category, using a seed prompt set and a simple tally.
- How to Track Backlink Sources That Appear in AI Citations: Connect cited third-party pages to your backlink profile so you can see where an existing link or a missing one lines up with AI citations.


