AEO · By Aidan Shaw · 13 min read

Where AI Engines Get Their Answers: We Classified the Sources Behind 90 Days of AI Citations

Published July 29, 2026
The short answer

We aggregated the source domains AI engines cited across the buyer questions we track for live client brands over a 90-day window (2026-04-12 to 2026-07-11), covering three unrelated verticals: auto insurance, beauty/skincare, and cannabis retail. We classified the top roughly 1,000 cited domains per project into seven source types. The headline finding: a brand's own website accounts for only 2 to 6 percent of the sources AI engines cite when its category comes up (auto insurance 4.0%, beauty 5.8%, cannabis retail 2.0%). Third-party editorial, community, and reference sources make up 26 to 59 percent, and the mix varies sharply by vertical. You cannot win AI search from your own site alone.

The short answer

We classified the sources AI engines cited across 90 days of live tracking for client brands in three unrelated verticals: auto insurance, beauty/skincare, and cannabis retail. The finding that held in every vertical: a brand's own website is only 2 to 6 percent of the sources AI engines cite when its category comes up. Auto insurance came in at 4.0 percent, beauty/skincare at 5.8 percent, cannabis retail at 2.0 percent. The third-party bucket (editorial, UGC, and reference sources combined) ranged from 26 to 59 percent. The rest is corporate sites, competitors, institutional sources, and a small residual. The practical conclusion is blunt: you cannot win AI search from your own site alone, because your own site is a sliver of the citation pool the engines are drawing from.

2-6%
Share of the sources AI engines cite that belong to the brand's own website, measured across live brands in three unrelated verticals over a 90-day window. Third-party editorial, community, and reference sources took 26 to 59 percent of the same pool.
Source: AEO Labs, aggregated AI citation tracking, 2026-04-12 to 2026-07-11

Key findings

  1. A brand's own website is only 2 to 6 percent of the sources AI engines cite in its category. Auto insurance 4.0 percent, beauty/skincare 5.8 percent, cannabis retail 2.0 percent, an average of roughly 3.9 percent. In no vertical did the brand's own domain reach even a sixteenth of the citation pool.

  2. Third-party surfaces take 26 to 59 percent of citations, several times the brand's own share in every vertical. Editorial, UGC, and reference sources combined reached 34.9 percent in auto insurance, 59.0 percent in beauty/skincare, and 26.0 percent in cannabis retail. These are the surfaces a brand can influence through earned placements but does not own.

  3. The source mix is vertical-specific, so a single AEO playbook does not transfer between industries. Editorial dominates beauty/skincare at 41.4 percent of citations, while corporate sites dominate cannabis retail at 65.6 percent and auto insurance at 52.2 percent. The same engines, tracked the same way, build answers from structurally different source pools depending on the category.

  4. Reference sites are a real citation surface in insurance but nearly absent in beauty. Encyclopedias and documentation-style sources took 13.8 percent of auto insurance citations, against 0.9 percent in beauty/skincare and 1.9 percent in cannabis retail. Regulated, definition-heavy categories lean on reference material; taste-driven consumer categories do not.

  5. Community carries far more weight in consumer categories than in financial ones. UGC sources (forums, social, community sites) took 16.6 percent of beauty/skincare citations versus 4.3 percent in auto insurance and 9.9 percent in cannabis retail. Where buyers ask peers, engines cite peers.

Own site vs third-party share of AI citations 0 15% 30% 45% 60% Own site Third-party AUTO INSURANCE Own site 4.0% Third-party 34.9% BEAUTY & SKINCARE Own site 5.8% Third-party 59.0% CANNABIS RETAIL Own site 2.0% Third-party 26.0%
The headline finding. In every vertical we measured, third-party editorial, community, and reference sources out-cited the brand's own website by a wide multiple. Source: AEO Labs, aggregated AI citation tracking, 90-day window, 2026.

Methodology

We run our client tracking on Peec AI daily. For this study we aggregated the source domains that AI engines (ChatGPT, Perplexity, Gemini, Google AI Overviews, Copilot, and others) cited across the buyer questions we track for live client brands, over a 90-day window from 2026-04-12 to 2026-07-11.

The sample covers three unrelated verticals: auto insurance, beauty/skincare, and cannabis retail. We chose the word "unrelated" deliberately. These categories differ in regulation, purchase journey, price point, and audience, which is exactly why a pattern that holds across all three is worth reporting.

For each project, we took the top roughly 1,000 cited domains and classified every one into seven source types:

Domains that did not fit any category are reported as a residual "Other" bucket. Percentages represent each type's share of total inline citations across the top cited domains per project.

A note on why we measured it this way. Most public commentary about AI search visibility is built on one-off screenshots: someone asks an engine a question, notices who gets named, and generalizes. That approach cannot separate signal from noise, because AI answers vary run to run, engine to engine, and week to week. Aggregating inline citations across a full quarter of daily tracking smooths that variance out. It also shifts the question from "who got mentioned today" to the more durable one: what kinds of sources do these engines structurally rely on when they build category answers? That structural pattern is what a marketing team can actually plan against.

Limitations, stated plainly

Research is only useful if you know what it does not show, so here are the limits.

The full source mix, vertical by vertical

Here is the complete seven-way classification for each vertical, plus the residual bucket. Every number is the share of total inline citations across the top roughly 1,000 cited domains for that project.

Source type Auto insurance Beauty/skincare Cannabis retail
Corporate 52.2% 25.5% 65.6%
Editorial 16.8% 41.4% 14.1%
Reference 13.8% 0.9% 1.9%
UGC 4.3% 16.6% 9.9%
You (brand's own site) 4.0% 5.8% 2.0%
Competitor 6.2% not broken out 1.1%
Institutional 2.0% 2.3% 2.6%
Other 0.8% 7.3% 2.9%

Competitor share was not broken out as a separate class in the beauty/skincare project; competitor domains there fall inside the corporate and other buckets.

Three patterns in this table deserve a closer look.

Corporate is the largest single class everywhere except beauty. In auto insurance (52.2 percent) and cannabis retail (65.6 percent), engines lean heavily on company websites: not the brand's own, but the wider population of vendor, aggregator, and industry company domains. In beauty/skincare, editorial takes that crown at 41.4 percent, with corporate at 25.5 percent.

The You row is the smallest meaningful class in every vertical. At 4.0, 5.8, and 2.0 percent, the brand's own site is out-cited by editorial in all three verticals, by UGC in all three, and by the corporate bucket in all three. If you have read our piece on why a brand goes missing from AI search, this is the data behind the third-party citation gap we call the most underestimated cause.

Reference and UGC trade places depending on the category. Insurance buyers get answers built on reference material (13.8 percent); beauty buyers get answers built on community discussion (16.6 percent). The engines mirror where trust lives in each category. That mirroring is worth sitting with, because it means the engines are not imposing a uniform sourcing standard on the web. They are amplifying whatever evidence ecosystem already surrounds a purchase decision, and a brand's earned-media strategy has to meet that ecosystem where it is rather than where the brand wishes it were.

The full source mix behind AI citations, by vertical Share of inline citations across the top ~1,000 cited domains per project. Each bar sums to 100%. AUTO INSURANCE 52.2 16.8 13.8 BEAUTY & SKINCARE 25.5 41.4 16.6 CANNABIS RETAIL 65.6 14.1 9.9 You (own site) Corporate Editorial Reference UGC Institutional Competitor Other Competitor share was not broken out separately in the beauty/skincare project. Source: AEO Labs, 90-day window, 2026.
The full mix. The lime segment on the far left of each bar is the brand's own site. Everything to its right is a source the brand does not control, and the composition of that territory changes completely between verticals. Source: AEO Labs, aggregated AI citation tracking, 2026-04-12 to 2026-07-11.

What this means if you run marketing

The instinct when a brand is invisible in AI answers is to fix the website: more content, better structure, more schema. That work matters, and we walk through it in how to rank in ChatGPT. But this data says the website is, at best, a 2 to 6 percent play. The other 94-plus percent of the citation pool is territory you influence through earned presence, or not at all.

Three implications follow directly from the numbers.

The battleground is third-party. Editorial, UGC, and reference sources took 26 to 59 percent of citations in our sample, several times the own-site share in every vertical. Getting your brand into the roundups, comparison articles, community threads, and reference pages the engines already cite moves you into the pool that actually gets quoted. This lines up with the broader industry evidence: Ahrefs found that roughly 43.8 percent of pages cited by ChatGPT are listicles ("best X" and "top X" formats), based on over 1 billion data points. If almost half of ChatGPT's citation diet is listicles, being absent from the listicles in your category is a structural handicap no amount of on-site optimization repairs.

The buyers are already there. This is not a future problem. G2 reports that roughly 48 percent of buyers now use AI in the buying process. The sources AI engines cite are, increasingly, the shortlist your buyers see. Citation share in those sources is market presence.

We can see this in our own analytics. Over the 33 days ending 2026-07-29, chatgpt.com was the second largest traffic source to aeolabs.ai at 169 sessions, nearly double the 90 sessions Google search sent us, and 47 percent of those ChatGPT visits were engaged sessions. A site that practices AEO gets found through AI answers. That is the mechanism this study describes, visible in our own referral logs.

Measure citation share, not just rankings or traffic. None of this shows up in a normal analytics stack. AI-influenced buyers tend to arrive as direct or branded visits, long after the answer that shaped their shortlist, and your Google rankings say little about which sources the engines quote. The metric that maps to this study is citation share: of the sources cited when your buyer prompts run, what fraction mention or belong to you, and how is that trending against competitors? That is a number you can baseline, report monthly, and hold a program accountable to, the same way search teams once held themselves to rankings.

Your playbook must be vertical-specific. A beauty brand should be fighting for editorial (41.4 percent of citations) and community presence (16.6 percent). An insurance brand needs reference and editorial coverage on top of a strong corporate footprint. A cannabis retailer operates in a pool where corporate domains take 65.6 percent, which makes directory-style and industry corporate placements disproportionately valuable. Copying another industry's AEO strategy means fighting on the wrong surface.

The operational loop we run for clients is the same one we used to produce this study: track the prompts, classify the citations, find the gap between where the engines look and where the brand appears, then close it surface by surface. If you want to run the tracking side yourself, start with our guide to the best AEO tracking tools. If you want the baseline done for you, our AI visibility audit measures your citation share across engines and shows you exactly which third-party surfaces are citing your competitors instead of you.

For a primer on the discipline behind all of this, this short Ahrefs explainer is a solid starting point.

Key takeaways

Cite this study

You are welcome to reference these findings in articles, newsletters, decks, and roundups. All we ask is attribution to AEO Labs and a link to this page so readers can check the method and the limits for themselves.

Suggested citation: AEO Labs, "Where AI Engines Get Their Answers: We Classified the Sources Behind 90 Days of AI Citations" (2026), aeolabs.ai/blog/ai-citation-sources-study.

Journalists and researchers who want the full anonymized breakdown behind the vertical percentages can email aidan@aeolabs.ai and we will share it. If your brand's own citation picture looks like the one in this data, our piece on why a brand goes missing from AI search is the diagnostic to run next.

Frequently asked questions

How many citations did you analyze?

We classified the top roughly 1,000 cited domains per project across three projects, one per vertical, drawn from every inline citation AI engines produced for our tracked buyer prompts over a 90-day window (2026-04-12 to 2026-07-11). Percentages represent each source type's share of total inline citations across those top cited domains, not a count of individual answer sessions.

Which AI engines were included in the study?

The citations come from the engines we track in Peec AI for live client brands, including ChatGPT, Perplexity, Gemini, Google AI Overviews, and Copilot, among others. We aggregated citations across engines rather than splitting them out per engine, so the percentages describe the combined citation pool for each vertical.

What counts as a third-party source in this study?

For the headline finding we grouped editorial (news sites, blogs, magazines), UGC (forums, social, community sites), and reference (encyclopedias and documentation) into a third-party bucket, because those are the surfaces a brand can influence but does not own. Corporate sites, competitor domains, institutional sources, and the brand's own site are reported separately in the full mix.

Why is a brand's own website cited so rarely?

Answer engines assemble responses to category questions from the sources that compare, review, and discuss brands, not primarily from the brands themselves. A vendor site answering "best X for Y" is one voice with an obvious interest; an editorial roundup or a forum thread is corroboration. Across our three verticals, own-site share never rose above 5.8 percent of cited sources.

Can I cite this study in my own article or newsletter?

Yes. You are welcome to reference any finding here with attribution to AEO Labs and a link to this page. A suggested citation is included in the Cite this study section. If you want the full anonymized breakdown behind the percentages, email aidan@aeolabs.ai and we will share it.

Does this mean publishing on my own site is a waste of time?

No. Your own site still earns the 2 to 6 percent of citations that go direct, and it feeds the entity and answer-quality signals engines rely on. The finding is about proportion: if most cited sources are third-party editorial, community, and reference pages, then a strategy that only optimizes your own domain is competing for a small slice of the citation pool.

See where you stand in AI search

Free audit. No commitment.

Book a call

Keep reading