Across 212 AI visibility audits and roughly 508,000 citations from ChatGPT, Perplexity, and Google AI Overviews, one pattern appeared consistently: the sources AI cites depend on the intent behind the query.
Buyers researching a category and buyers choosing between vendors are shown different mixes of sources. Across the dataset, only one source remained consistently represented in both query types. Almost every other citation signal shifted, in some cases by an order of magnitude. That explains why a brand can appear highly visible during research yet become far less visible when buyers begin evaluating vendors, or vice versa.
Key findings
- Query intent changes what AI cites: Commercial-intent and informational-intent queries rely on different source classes, even within the same category.
- Commercial-intent visibility depends heavily on third-party authority: Review platforms, comparison sites, directories, and entity-validation sources play a much larger role when buyers are evaluating vendors.
- Informational queries favour authoritative educational content: Regulatory sources, public authorities, documentation, explainer content, and industry publications become significantly more prominent.
- The same domain can earn citations in different ways: Vendor and business websites account for roughly 68% of citations in both intent types, but the pages AI cites are fundamentally different.
- AI engines do not rely on the same sources: ChatGPT, Perplexity, and Google AI Overviews each show different citation preferences, making engine-specific visibility an important part of AI search strategy.
- Citation strategy should follow buyer intent: Measuring commercial-intent and informational-intent visibility separately provides a more accurate picture than relying on a single overall AI visibility score.
The intent-driven citation shifts we measured
The findings, at a glance:
- Reddit remains consistently cited: It accounts for roughly 12% of citations in both commercial-intent and informational-intent queries and appears in every one of the 212 audits.
- Wikipedia is primarily an entity-validation source: It represents 7.9% of commercial-intent citations but only 3.6% of informational citations, a relative decline of roughly 55%.
- Authority sources dominate informational queries: Regulatory and public-authority sources (measured by government TLD share) increase from roughly 1.2% of commercial citations to 10.0% of informational citations, an increase of around eight times.
- Review platforms are concentrated in commercial intent: Review sites and directories account for a small but meaningful share of commercial citations, but fall to almost zero in informational queries.
- Vendor websites remain equally important, but different pages earn the citations: Vendor and business websites account for roughly 68% of citations in both intent classes, but the pages AI cites change almost completely.
Two intent classes, two citation profiles
Each of the 212 audits included two types of prompts.
- Commercial-intent prompts captured buying-consideration queries such as “best category for use-case,” “top tool alternatives,” and “Brand A vs Brand B.”
- Informational-intent prompts captured learning-focused queries such as “how does thing work,” “risks of decision,” and “how to process.”
Both prompt sets were run across ChatGPT, Perplexity, and Google AI Overviews. Every cited domain was captured, normalized, and classified by page type.
Comparing the two datasets reveals two distinct citation profiles rather than the same pattern at different volumes. Only 16,048 domains appear in both. Another 71,032 appear only in commercial-intent citations, while 39,960 appear only in informational citations.
The highest-volume sources still overlap, accounting for roughly 61% of citation volume in both datasets. Beyond that shared core, however, the citation landscape changes considerably. Mid-tier sources, long-tail domains, and the mix of page types all shift depending on the intent behind the query.
What each regime is actually made of
At the buying-consideration stage, the citation pattern skews toward validation, comparison, and vendor identity. Wikipedia is the second most-cited source in commercial queries at 7.9% share, doing entity-check work: confirming that a named brand exists as a recognized entity with a documented description. Review platforms, professional-services rating sites, and category directories appear in this universe with small but real shares and virtually disappear from the informational one. Named vendor domains dominate the top of the commercial vendor list. In the HR-technology audits we ran, the highest-cited individual vendor domains were the homepages and pricing pages of the category’s own competitors, each appearing across 20 to 24 distinct audits with around 1,000 citations apiece.
The practical read for a brand competing here: your commercial-intent visibility rides on entity readiness (are you a documented entity in Wikipedia and adjacent reference sources), review-platform presence, and how accurately your competitors’ product pages describe the category. Two of those three sit outside your own domain, which is why commercial-intent optimization is largely a third-party citation problem.
At the learning stage, the citation pattern shifts to authoritative-explainer sources. Regulatory and public-authority sources combine to represent roughly 10% of informational citation volume, up from roughly 1.2% commercial. The class includes tax authorities, sector regulators, court and legal systems, and public authority research databases. Academic content increases slightly. YouTube’s share nearly doubles, from ~1.2% to ~2.1%. LinkedIn’s share more than doubles. Wikipedia’s share falls by more than half; it does far more work as an entity check than as a tutor. Review and directory clusters, prominent at the buying stage, nearly vanish from informational citations. Their replacement is not more UGC. It is more authority-tier content.
The practical read for a brand competing here: informational-intent visibility rewards a completely different asset class. Explainer and documentation depth on your own site, contributed authority content on regulator-adjacent and industry publications, video walkthroughs, and long-form professional-network content are the surfaces that get cited. Review-site rating and Wikipedia legitimacy carry marginal weight. A source gap analysis tuned only to commercial prompts will miss all of it.
The finding that surprised us: same vendor share, different pages
The single most stable statistic across both regimes is vendor and business-website share. Roughly 68% of commercial citations point to a vendor or business website; the informational share sits at roughly 70%. The absolute proportion of “AI cites a company’s own site” barely moves between the two intent classes.
What changes completely is which pages of those vendor sites get cited. In commercial queries, the top-cited vendor domains are product homepages, pricing pages, and named category-leader competitor pages. In informational queries, the top-cited vendor domains are learning portals, product documentation, industry-explainer content, sector-specialist blog articles, and knowledge-base pages. The list of names in the top-15 vendor domains barely overlaps between the two universes.
The practical read: the aggregate framing “AI cites vendor sites about 68% of the time” is technically true and strategically incomplete. The 68% at the top of the funnel and the 70% at the bottom are two different citation surfaces on the same vendor domains. Optimizing for one does not automatically build presence in the other.
The implication for content strategy is granular. A product page optimized for commercial-intent citation, with structured comparison and named-alternative references, does different work than a documentation or explainer page tuned for informational-intent citation, with step-by-step how-to content, embedded regulatory context, and cross-referenced industry standards. Both live on the same vendor domain. Both are subject to different citation-earning mechanics.
Where the engines quietly disagree
Looking only at the overall citation share hides another important pattern. Within the informational dataset, each AI engine shows clear preferences for different source types.
- ChatGPT effectively did not cite YouTube.
- Perplexity and Google AI Overviews together accounted for all 3,440 YouTube citations recorded.
- Reddit citations were overwhelmingly ChatGPT-driven, with 96.5% attributed to ChatGPT.
- Regulatory and public-authority sources also skewed heavily toward ChatGPT.
These differences show that AI engines do more than recommend different brands. They rely on different source classes even when answering the same type of query.
The complete engine-divergence analysis is covered in a separate piece, while the query-side impact is explored in one question, many answers. The key takeaway is that “Which AI engines mention my brand?” and “Which AI engines cite the sources that establish my authority?” are related questions, but they should be measured separately.
A three-part frame you can actually apply
Three questions cover the whole decision:
- Which regime does your buyer meet AI in?
Some categories cluster their AI-answer competition at buying-consideration. SaaS purchase decisions, most B2C considered purchases, and vendor-comparison-heavy service categories tend to sit here. Others cluster earlier, in the learning stage. Categories with regulatory complexity, technical categories where evaluation begins with an education phase, and considered purchases where the buyer researches for weeks before naming vendors tend to sit here. The rough sort is not perfect, but it is where the strategy conversation should start.
- What does the regime look like in your specific vertical?
In our HR-technology audits, informational citations concentrated in workplace-safety and payroll-regulator explainers alongside sector-specialist blog content. In our real-estate audits, they came from county appraisal offices, local property-market blogs, and mortgage-regulator explainers. In legal-services audits, court websites, state-bar publications, and legal-explainer Q&A sites dominated. The specific regulator and explainer surfaces that light up depend on the vertical. The structural pattern (regulator + explainer + long-form + video for engines that lean video) holds across all of them. Schema and entity signals matter regardless of regime because both regimes need to recognize you as an entity before citing you.
- How do you find out where you actually sit?
Pull twenty commercial-intent prompts for your category and twenty informational prompts covering the same topical surface. Fire each at ChatGPT, Perplexity, and Google AI Overviews. Log the cited domains. Compare the source-class distribution you already earn against the two regimes described above. The gap between your current footprint and the regime you are targeting is your work list. The Princeton Generative Engine Optimization paper published at KDD 2024 (available via arXiv) showed that AI citation rewards a substantially different signal set than organic ranking; the matched-pair audit is how you find out which specific signals your gap has.
Citation share is now a segmented KPI, not a single number. The regime you compete in and the vertical you sit in shape which sources matter. The one test that isolates the answer for your specific brand is the matched-pair audit above.
How this analysis was run
The findings are based on ReSO’s proprietary AI visibility dataset covering 212 brand audits conducted between March and June 2025. Each audit included approximately 35 commercial-intent prompts and 15 informational-intent prompts, run across ChatGPT, Perplexity, and Google AI Overviews.
Every cited domain was captured, normalized to a canonical format, and classified using ReSO’s page-type taxonomy. The analysis includes 347,880 citations from commercial-intent prompts and 161,076 citations from informational prompts, representing a combined universe of roughly 87,080 unique domains. The audited brands span multiple industries, including SaaS, professional services, e-commerce, manufacturing, and media, across markets such as the US, UK, Australia, India, Canada, Germany, France, and Singapore.
A few methodology notes:
- Relative differences between commercial-intent and informational-intent citation patterns are considered reliable.
- Detailed page-type breakdowns should be treated as directional while the page classifier continues to evolve.
- Per-engine citation data was available only for informational queries, so all engine-specific findings in this analysis relate to informational prompts.
As Google AI Overviews continue to appear across a growing share of informational searches, understanding which sources AI cites at the learning stage becomes increasingly important. Recent research from Semrush’s AI overviews study points to the same trend: informational queries are becoming a larger part of how users interact with AI-generated search experiences.
Get your intent-segmented AI search visibility audit
Commercial-intent and informational-intent queries rely on different citation sources. Measuring overall AI visibility cannot tell you which of those two citation regimes your brand is winning or missing.
ReSO’s AI Visibility Audit measures your visibility separately for commercial-intent and informational-intent queries across ChatGPT, Perplexity, and Google AI. It shows which source classes AI cites for your category, where your brand is represented, and the opportunities to strengthen your presence at each stage of the buyer journey.
Every audit includes an intent-by-intent visibility breakdown and a prioritised action plan tailored to your category, helping you focus on the citation gaps that matter most. Request your AI Visibility Audit to see how your brand performs across both AI citation regimes.
FAQs
Does the citation-source shift hold across all verticals?
The aggregate numbers are cross-vertical. Vertical-specific splits of the intent shift are computable but not published in this analysis. The 212-audit composition (76 professional services, 68 SaaS, 48 mixed, 16 e-commerce, 4 manufacturing) makes SaaS and services findings the reliable slices; smaller-vertical hints are directional.
Why does Wikipedia lose share when the query becomes informational?
Wikipedia carries entity-check weight: it validates that a named brand exists as a recognized entity with a documented description. That work matters most in commercial queries. Informational queries ask how something works, where regulator explainers and topic-specific tutorials tend to outrank encyclopedic summaries.
How is the informational-versus-commercial split defined here?
Each audit separates commercial-intent prompts (such as “best X” or “X vs Y”) from informational prompts (such as “how to X” or “how does X work?”). Citation patterns are then measured separately for each intent group.
Should a brand optimize for one intent regime or both?
It depends on how buyers make decisions in your category. Some industries rely more on commercial-intent queries, while others build trust through informational content first. Auditing both intent types shows where your visibility matters most.
