What Types of Content Do AI Search Engines Cite?

First-party pages lead, but every engine has a different source mix

Days 1 through 5 of this research series measured the cross-engine citation graph at the domain level and the vertical level. The Q2 2026 dataset has another lens that has not been pulled yet: what kind of content each engine reaches for. The 1,430 citations in the classified subset were tagged with one of fifteen structural categories. The patterns by engine turn out to be striking.

ChatGPT cites the vendor's own first-party site 68% of the time. The other four engines run between 46% and 52%. ChatGPT also cites zero video, zero community discussion, zero personal blogs, and zero review aggregators. Claude cites zero video, zero community discussion, and zero tech press. Perplexity is the only engine in the five that cites YouTube at scale, and Gemini is the only engine that cites Reddit at any meaningful rate. The engines are running on visibly different slices of the web even when they're answering the same buyer-intent question.

The per-engine summary table

Each engine returned 75 responses, but the count of citations per response varies by a factor of nearly three. Google AI Overview leads at 9.3 citations per response. Gemini ships 9.3. Perplexity 7.3. Claude 5.3. ChatGPT 3.3. The classified subset within each engine is the right denominator for any structural comparison.

The full per-engine, per-category matrix

Each cell is the share of the engine's classified citations going to that category. The categories are ordered by total classified citation volume across all five engines. Cells where the engine cites zero of a category are shown as 0%, which is what they are. There are eleven such zeros in the matrix, all five clustered around ChatGPT and Claude.

Coverage: how broad each engine's citation diet is

The fourteen classified categories are not equally represented across engines. Gemini and Google AI Overview cite from all fourteen. Perplexity covers thirteen (the only gap is community discussion). ChatGPT and Claude each cover nine of the fourteen. The five categories ChatGPT skips entirely are video, personal blog, review aggregator, community discussion, and search-engine cache URLs. The five Claude skips are video, tech press, community discussion, lifestyle media, and developer platform.

Five engine signatures

Run the per-engine columns through the eye and a personality emerges for each one. Below is the simplest version of that read.

The vertical overlay: SaaS is vendor-first, CPG is publisher-first

Collapse the 25 verticals into three super-categories: tech SaaS, consumer services, and CPG / retail. The engine-level patterns above are real, but they are partly an artifact of which verticals each prompt set covered. Looking at the vertical lens directly, the gap between SaaS and CPG / retail is much larger than the gap between any two engines.

Tech SaaS responses cite the vendor's own first-party 76% of the time and a niche publisher hub 0.6% of the time. CPG / retail responses cite the vendor first-party 9.4% of the time and a niche publisher hub 48.5% of the time. The buyer asks "best CRM software" and gets salesforce.com. The buyer asks "best mattress" and gets sleepfoundation.org. Two different AI-search products, depending on the vertical.

The single engine-exclusive category

The taxonomy includes one category that is effectively a one-engine signal in Q2 2026:

• Community discussion (Reddit, Quora, HN): Gemini accounts for 69.2% of all 13 citations to this category in the dataset.

Community discussion is concentrated in Gemini in this dataset. Google AI Overview accounts for 4 of 13 community citations, while ChatGPT, Claude, and Perplexity account for none. The total citation volume to community UGC across the entire dataset was only 13 URLs (0.9% of classified citations). This panel does not show that community participation caused those citations or that another prompt set would produce the same split.

What this means for a publisher

This is an observational source-share panel. It records which source types appeared in one frozen set of answers. It does not identify why an engine selected a source or estimate the citation lift caused by publishing, editing, or earning a mention. The actions below are channel hypotheses to test with stable prompts and engine-level before-and-after measurement.

One. Test vendor first-party surfaces first. Every engine puts the brand's own .com in the top three content types (ChatGPT 68%, Claude 52%, Perplexity 48%, Google AI Overview 49%, Gemini 46%). That makes product pages, resource centers, glossary entries, and FAQs the clearest owned-surface hypothesis in this sample. An intervention test is still required. Improve the page that answers the tracked prompt, then rerun the same prompt by engine. The related AI Readiness study found that 23.2% of scanned sites ship no structured data at all, which is a separate technical observation rather than evidence of citation lift.

Two. Test niche-publisher-hub placement for consumer-facing categories. In CPG, retail, and health prompts, the observed mix includes category publishers such as Sleep Foundation, Healthline, OutdoorGearLab, and The Good Trade. Across all engines combined, niche publisher hubs are the second-largest classified content type after vendor first-party. That makes an independent placement test reasonable; it does not prove that a placement will produce a citation.

Three. Test video and community discussion by engine, not as universal channels. In this sample, video was concentrated in Perplexity, with smaller shares on Gemini and Google AI Overview and none on ChatGPT or Claude. Community citations appeared on Gemini and Google AI Overview only. Use those distributions to choose a test, then measure the target engines independently instead of assuming the sample predicts future behavior.

How to use this source-layer map

The dataset is useful because it separates the owned-page work from the outside-source work. When an engine's observed mix is mostly first-party vendor pages, the next test can start on product, docs, FAQ, comparison, and research pages. When the observed mix includes publisher hubs, community threads, reviews, or video, test the matching outside-source channel separately.

Use the table above to pick the channel by engine, then use the playbooks below to turn that channel into a concrete action queue.

Other reports in this series

This is Day 7 of an ongoing research cadence. Days 1 through 5 draw from the Q2 2026 AI Search Citation Benchmark and measure the engine side of citation. Day 6 measures the publisher side: how technically prepared the typical scanned domain is to be cited. This Day 7 report slices the engine side by content type, the lens Day 2 collapsed away into binary aggregator-vs-vendor.

• AI engines agree on brands more than sources. A post-fix production-data study of answer agreement vs. source overlap.
• AI Readiness across 194 websites. Median AI Readiness 60 vs median SEO 90. The 30-point publisher-side gap.
• Top 100 most-cited domains in AI search (Q2 2026). Only 12 of 1,119 domains are cited by all five engines.
• ChatGPT vs. Google AI Overview: the same prompt, two different webs. 4.1% Jaccard, 64% zero-overlap.
• Buyer intent reshapes AI citations. Discovery, shortlist, and variation cite different domain sets.
• When AI engines cite the reviewer vs. the brand. The 70-point vendor-vs-aggregator gap by vertical.
• AI Search Citation Benchmark, Q2 2026. The underlying dataset.

To turn the content-type mix into a monitoring plan, compare tool coverage in the AI search visibility comparison hub and track engine-specific prompt behavior with Foglift AI search monitoring.