What AI Actually Cites: 3,050 Answers Analysed

Most writing about AI citations leans on a handful of public studies. We wanted our own numbers, so we built a measurement engine and pointed it at the question directly: when you ask an AI a commercial question, what does it actually cite?

This is what 3,050 AI answers told us.

The dataset

We ran 230 distinct commercial questions — the kind a buyer types when they're close to choosing — through three answer engines (ChatGPT Search, Claude with web search, and Google AI Overviews), repeated across multiple runs, and recorded every URL each answer cited. After de-duplicating domains within each answer (one answer = one vote per domain), that's:

• 3,050 individual AI answers
• 11,703 domain-citations
• 2,662 distinct domains cited
• Two broad categories: local & service businesses and software / SEO / digital

Honest limits, up front. This is our own scan corpus, not a random sample of the entire web — the question set is weighted toward categories we work in, and it spans several months and model versions, so treat the cross-engine and cross-category contrasts as the signal, not the absolute percentages. Perplexity coverage in this run was too thin to report. We're publishing the aggregate table below so you can check our working.

Finding 1 — Two thirds of AI answers cite a source, and they cite the open web

67.1% of the answers carried at least one citation. The other third answered from the model's own parameters with nothing to click.

When they did cite, the overwhelming majority of citations — roughly 89% — went to individual web pages: company sites, blogs, editorial articles. The big "consensus platforms" people assume dominate AI answers (Reddit, YouTube, Wikipedia) were a small share of citation volume — under 4% combined.

That's the first myth punctured. AI mostly cites ordinary pages on the open web, not a short list of mega-platforms. If your page is a clear, liftable answer, it can be cited — you don't need a Wikipedia entry to get in.

Finding 2 — The engines do not agree on what to cite

Here's where it gets useful. We gave the three engines the same questions and they cited structurally different kinds of sources:

| Source type | ChatGPT | Claude | Google AI Overviews | |---|--:|--:|--:| | Individual web pages | 95.5% | 89.2% | 83.5% | | Directories | 0.1% | 6.0% | 1.2% | | Review platforms | 0.6% | 4.4% | 2.7% | | Reddit | 0.0% | 0.0% | 4.0% | | YouTube | 1.5% | 0.0% | 3.7% | | Social / community | 1.0% | 0.3% | 4.7% |

Read across the rows:

• ChatGPT is the purist — almost everything it cites (95.5%) is an ordinary web page. Win it with a strong, citable page.
• Claude leans hardest on directories and review platforms (over 10% combined) — claimed profiles and third-party listings carry real weight here.
• Google AI Overviews is the most diverse engine by far — it's the only one that meaningfully pulls Reddit, YouTube and social into answers.

The practical takeaway: "getting cited by AI" is not one game. The source that wins you a ChatGPT citation may be invisible to Claude, and vice versa. A serious AEO effort covers more than one source type on purpose.

Finding 3 — What's citable depends on your category

Split the same corpus by category and the claimable surfaces invert:

| Source type | Local & service | Software / SEO / digital | |---|--:|--:| | Individual web pages | 85.8% | 91.4% | | Directories | 6.1% | 0.0% | | Review platforms | 5.4% | 0.5% | | YouTube | 0.6% | 2.6% | | Social / community | 0.7% | 3.3% | | Wikipedia | 0.1% | 0.7% |

• Local & service businesses get cited through directories and review platforms — Yelp, HomeGuide, Expertise, Angi and the like make up over 11% of citations. These are claimable: a profile you can create, fill out and earn. That's the fastest lever in local.
• Software / SEO / digital categories barely touch directories. Instead the non-page citations skew to YouTube, community and Wikipedia. Different category, different doorway.

One nuance worth stating plainly, because it's easy to misread the volume numbers: in the digital category, YouTube shows up across a wide spread of distinct questions even though it's a small slice of total citation volume. Breadth and volume aren't the same thing — a single well-placed video can surface across many queries. For placement strategy, the question "how many different answers does this source appear in?" matters more than "what percentage of all citations is it?"

What this means for getting cited

The data points to one disciplined approach rather than a universal trick:

1. Match the source types your category actually uses. Local business? Claim and complete your directory and review-platform profiles first — the data says that's where the citations are. Digital/SaaS? Invest in YouTube, community presence and editorial inclusion.
2. Cover more than one engine. A page-only strategy wins ChatGPT and under-performs on Claude and Google AI Overviews. Add the directory/review layer (Claude) and the community/video layer (AI Overviews) deliberately.
3. Make your own page liftable anyway. Across every cut, ordinary web pages are ~85–95% of citations. A clear, self-contained answer on your site is the foundation everything else builds on.
4. Measure, don't assume. The contrasts above are large and they're specific to category and engine. Run the questions that matter to your business and read your own leaderboard before spending on placements.

That last point is the whole idea behind GetCited: point a measurement engine at your category, see exactly which sources the AI cites, and earn your way onto them. If you'd like that done for your brand, here's how it works.

The numbers

You can audit the aggregate behind this study: download the dataset (CSV). Every figure above is counted directly from stored AI answers — no estimates.

---

Methodology note: domains were de-duplicated within each answer so a source is counted once per answer regardless of how many of its pages were cited. "Individual web pages" covers company sites, blogs and editorial articles that aren't directories, review platforms or named consensus platforms; it therefore includes both genuine editorial and competitor business sites, which our bucketing does not separate. Figures span multiple model versions over several months and reflect our question set, not a random web sample.