What sources do AI answer engines cite?

When someone asks ChatGPT, Gemini, Perplexity or Google AI Overviews which tool to use, the answer arrives with sources. This is a count of what those sources are, read by host, over the answers our product has stored for customer questions about B2B software since July 2026. It is one sample, from one category of question, and every figure below says which engine and which window it came from.

Source use differs by engine

Measured as the share of cited answers that include at least one host from each family, on 18,625 answers with citations (product corpus, 19 July to 2 September 2026):

A second sample of 152 companies across categories (15,023 answers) was within three percentage points on every cell. The largest difference was Perplexity’s Reddit share: 46.3% here and 43.5% there. The broad source pattern was similar across those samples, although that does not establish how other categories would behave.

Publications, the row a press strategy cares about, lead only on the ChatGPT API path, where TechRadar and the PR wires carry it. Elsewhere the row is 16.1% on Perplexity, 9.2% on Gemini and 5.9% on Google AI Overviews.

Tap to zoom

Reddit results depend on the measurement surface

On 3,183 customer questions asked to all four engines in the same window (22 to 23 August 2026), the same questions, the same companies, using non-error records (including Google checks that returned no AI Overview):

On 1,735 of those questions at least one engine cited Reddit; the API path cited it on none of them, while citing YouTube, LinkedIn, Medium, GitHub and G2 on the same questions. The zero was observed on that API retrieval path in this sample. Our ChatGPT column is the OpenAI API with web search, not the ChatGPT app; this analysis does not establish the app's Reddit citation rate.

Tap to zoom

Directory counts depend on the question being counted

We took SEOmade's list of 975 directories "worth submitting to", plus our own checklist and Dodo's Index, and counted how often each host appears among 225,143 citation instances.

The 40-host directory classification used for this count is 1.68% of citations (3,787 of 225,143). The table above uses a separate 35-host bucket and counts answers containing a citation, rather than citation instances. These percentages measure different things; they should not be compared as a trend. By engine: Perplexity 2.45%, ChatGPT API 1.78%, Gemini 0.83%, Google AI Overviews 0.54%.

The ten most cited directories carry 82% of all directory citations (3,118 of 3,787). This concentration does not tell us what a new listing would achieve. A directory being cited is not evidence that it carried the company named in the answer.

Of 453 classifiable platform citations, none was classified as first-party

We could check who wrote a cited platform post only where the URL names the author: Medium, dev.to and Substack. On the 2,619 answers that resolve to a brand domain, 2,429 citations pointed at a platform, and 453 of those could be classified. The retained classification recorded zero first-party posts among those 453. YouTube and Reddit URLs carry no author, so 81.4% of platform citations could not be classified this way, and the question for those two platforms is open.

Most GitHub citations in this sample point to repository roots

GitHub is 0.75% of all citations (1,683 of 223,815, in 946 answers). Of those 1,683, 1,092 (64.9%) are repository roots; 213 are github.io pages; 115 are markdown files in a repository; issues, discussions and pull requests are 54.

What to do with it

These are observed shares of what was cited, by engine. They tell you where the answers draw from for these kinds of questions; they do not show that adding your own post or listing to one of those surfaces earns a citation, and none of the 453 author-classifiable platform citations was a first-party post. Authorship of the remaining 1,976 citations was unresolved. Read them as a prior for where to look, engine by engine: Reddit and YouTube on Google AI Overviews, Reddit, LinkedIn and the startup-data directories on Perplexity, publications and the wire on the ChatGPT API path, with the app possibly different. Our working policy is to treat a relevant review-site listing as a one-time name-discovery task. This analysis did not measure its cost or effect. It also did not test domain rating as a predictor, so it does not support a domain-rating-based submission priority.

Scope and caveats

One product, one category of question, one summer. Customer questions about B2B software, weighted to developer tooling, and a YC cohort for the reproduction. Shares of cited answers say the surface appears in the answer, not that it carried the company. The ChatGPT column is the OpenAI API with web search, not the app. Gemini hosts are read from citation titles. The directory host lists are hand-selected, overlap with the review-site bucket, and are retained with the analysis. The table columns are not exclusive categories and do not add up to 100%. The platform-author count is partial by construction and filed as partial.