79% of AI citations on Perplexity, Gemini, and Claude go to someone else’s website. On ChatGPT, 74.6% go to yours.
We ran 75 B2B SaaS buyer queries across ChatGPT, Perplexity, Gemini, and Claude, and classified 2,020 citations by format, source and buyer stage.
- ChatGPT goes directly to vendor websites: 74.6% of its product-query citations are first-party, while 79.0% on Perplexity, Gemini and Claude are third-party.
- The cited format changes at every buyer stage: listicles at consideration (50%), comparisons at evaluation (42%), pricing guides at decision (30%), how-tos at implementation (46%).
- Pages with specific numbers win: 80% of decision-stage citations contain pricing, benchmarks or ROI figures.
- Citations show only part of the picture: Reddit, YouTube and LinkedIn shape answers without getting the attribution.
- Research reports were 2% of citations, but original data inside ordinary formats is what makes pages citable.
- AI search validates the existing disciplines: SEO wins on ChatGPT, content and digital PR win on the other 3.
At VisibleIQ, we run AI search audits for B2B SaaS companies (here is our step-by-step guide to running one). This study came from a question our clients keep asking: when a buyer uses AI to research software, where do the citations come from?
We ran 75 B2B SaaS buyer queries across ChatGPT, Perplexity, Gemini, and Claude, and classified every citation by format, source type, and buyer stage: 2,020 citations in total. The product-specific queries told the clearest story: 79% of citations on Perplexity, Gemini, and Claude link to someone else’s website. On ChatGPT GPT-5.4, 74.6% link to the product’s own website.
ChatGPT now goes directly to vendor websites using hidden search operators, bypassing the blogs and comparison sites the other 3 platforms still rely on. In other words, ChatGPT is rewarding what good SEO has always built: a well-structured website with clear product information, pricing, and documentation. The other 3 platforms are rewarding what content marketing and digital PR have always built: other people writing about you.
Methodology and caveats
We queried 75 B2B SaaS buyer prompts across 4 AI platforms on their current production models and manually classified every citation URL by format, source type, and buyer stage, producing a dataset of 2,020 citations collected on 28 March 2026 via the engines’ production APIs.
We split the 75 queries evenly across 5 buyer stages: awareness, consideration, evaluation, decision, and implementation, with 15 queries per stage. The queries ranged from “What is product-led growth” to “HubSpot vs Salesforce total cost of ownership” to “How to set up HubSpot for a SaaS company,” covering the full spectrum of what a B2B SaaS buyer might ask AI during a purchase decision.
We tested 4 AI platforms, each on their web-search-enabled model: Perplexity sonar-pro, Gemini 2.5-flash, Claude haiku-4.5, and ChatGPT GPT-5.4 (released 5 March 2026). Every query ran on all 4 engines, and where one returned zero citations because it answered from training data instead of searching the web, we recorded that gap. All data was collected on 28 March 2026, one pass per engine on the same day, and the raw dataset is available for download at the bottom of this post.
An important caveat. This study tracks what AI models cite in their visible responses. What they consult goes further. Models draw on training data, fan-out queries and internal retrieval that never appears in citations: Reddit is heavily consulted behind the scenes yet almost never cited, and video and professional networks surface in answers far more than citation counts suggest. Our data captures the visible layer. The real influence is broader, and we address it in the section on what AI uses versus what it cites.
79% of Perplexity, Gemini, and Claude Citations Link to Someone Else’s Website
Across 45 product queries on Perplexity, Gemini, and Claude, 79.0% of citations pointed to third-party domains. Only 21.0% cited the vendor’s own website. Even when a buyer asks about a specific product by name, the AI overwhelmingly cites what other people say about that product.
Of our 2,020 total citations, 45 of the 75 queries were product-specific (consideration, evaluation, and decision stages where specific brands are discussed). These are the queries where the distinction between “the product’s own website” and “someone else’s website” applies. The remaining 30 queries were educational and how-to queries with no specific brand being evaluated, so we excluded them from this analysis. That leaves 1,424 product query citations across all 4 platforms, or 1,243 on Perplexity, Gemini, and Claude alone (we separated ChatGPT GPT-5.4 because its behavior is different enough to warrant its own analysis). The split on those 3 platforms: 261 citations to the product’s own website, 982 to someone else’s.
But that overall number masks how much the ratio shifts by query type.
| Query type | Third-party % | First-party % | Citations counted |
|---|---|---|---|
| “Best X” recommendations | 65.9% | 34.1% | 361 |
| Head-to-head comparison (“X vs Y”) | 81.7% | 18.3% | 487 |
| Decision/pricing (“Is X worth it”, “How much does X cost”) | 93.7% | 6.3% | 158 |
| Topic/educational (“What is X”, “How to X”)* | 99.6% | 0.4% | 1,021 |
| Product queries only (weighted)** | 79.0% | 21.0% | 1,243 |
* Topic/educational queries excluded from headline figure as they contain no product references.
** This is the headline figure: 79.0% third-party across the 3 product query types on Perplexity, Gemini, and Claude.
* Includes queries 51-75 (additional consideration, evaluation, and decision queries) which add 237 citations to the 3 core product query types shown above.
The decision stage surprised us. When a buyer asks “Is HubSpot worth it for a Series B SaaS,” you would expect HubSpot’s own site to dominate the citations, but it doesn’t. 93.7% of citations for decision-stage queries pointed to agency blogs, comparison sites, and independent reviews.
The number only tells part of the story, though. What format that third-party content takes changes completely depending on what the buyer asks.
Which Content Formats Get Cited Most in AI Search? It Depends on the Buyer Stage
Blog content accounts for 66.9% of all AI citations, but the specific format shifts at each buyer stage. Listicles dominate at consideration (50%), comparison pages at evaluation (42%), pricing guides at decision (30%), and how-to guides at both awareness (34%) and implementation (46%). Publishing the wrong format for the stage means zero visibility.
First, the overall picture. Here is every content format we classified across the full dataset:
| Content format | % of 2,020 citations |
|---|---|
| Blog post (how-to / guide / strategy) | 30.9% |
| Listicle / “Best X” / “Top X” | 20.0% |
| Comparison page / “X vs Y” | 16.0% |
| Vendor blog / product resource page | 10.0% |
| Pricing / cost guide | 5.0% |
| Glossary / definition page | 4.0% |
| Research report / original data | 2.0% |
| Video (YouTube) | 1.6% |
| Review aggregator (G2 / Gartner / Capterra) | 1.5% |
| Vendor documentation / knowledge base | 1.1% |
| Community / Q&A (SaaStr, HubSpot Community) | 1.0% |
| Template / downloadable resource | 0.9% |
| News / editorial (Forbes, TechRadar) | 0.8% |
| Reddit / forum | 0.8% |
| Job description / hiring template | 0.6% |
| Other (Medium, newsletters, PDFs) | 3.8% |
The top 3 formats (blog posts, listicles, and comparison pages) account for 66.9% of all citations. G2 and Gartner combined sit at just 1.5%. Reddit and YouTube together make up 2.4%. The long tail is wide but thin.
Those overall numbers are useful, but they hide the real story. The format that dominates changes at each buyer stage:
Awareness: glossary pages still work here
At the awareness stage, how-to guides make up 34% of citations and glossary pages account for 18%. Research reports hit their peak at 6%. Educational institutions (Harvard Business School, Coursera, Northwestern) only appeared at this stage.
Strategy content rounds out the picture at 15%. This is where the “What is X” and “How to X” content earns its citations. If you are writing educational content, this is where it lands.
Consideration: the listicle stage
Listicles dominate at 50%. Half of all citations at the consideration stage came from “Best X for Y” and “Top 10” format posts. Blog guides follow at 20%, vendor comparison pages at 7%, and YouTube at 3% (its peak across all stages).
This is the stage where vendor-authored “best of” lists are cited most. Vendors writing listicles that include their own product alongside competitors appeared repeatedly. If you sell project management software, writing “Best Project Management Tools for SaaS Teams” and including yourself feels a bit cheeky. But it works. The AI picks it up every time.
YouTube citations peak at this stage too, at 3%. Perplexity in particular cited video reviews and comparison walkthroughs. Agency blogs also performed well at 6%, mostly from agencies publishing “best of” lists in their client’s category.
Evaluation: comparison content is king
Third-party comparison pages account for 42% of evaluation-stage citations. Competitor-authored comparisons add another 16%. Vendor-owned “us vs them” pages contribute 15%.
We expected vendor comparison pages to perform better than 15%, but they didn’t. Third-party independent comparisons were cited at nearly 3x the rate of vendor-published ones. When Avoma writes “Gong vs Chorus,” it gets cited. When Gong writes “Why Gong beats Chorus,” it gets cited less often. The implication: what other people write about your product carries more weight at evaluation than what you write about yourself.
Review aggregators (G2, Capterra, Gartner Peer Insights) account for only 3% of evaluation-stage citations. We expected higher. AI models prefer individual comparison articles over aggregator pages with summary scores. G2’s influence appears to be more indirect: models reference G2 data in their reasoning without citing the page.
Decision: pricing data wins
Pricing and cost guides make up 30% of decision-stage citations. ROI analysis and benchmark content follows at 20%. Comparison content persists at 14%.
This is the stage with the highest data density in cited content. 80% of pages cited at the decision stage contained specific numbers: pricing tiers, implementation cost ranges, budget benchmarks, or ROI percentages. Case studies appeared at only 1%, and community content (SaaStr, Reddit) at 3%.
One Series B project management SaaS company we audited had published 14 blog posts. All how-to guides. They appeared in zero AI responses for their consideration-stage target queries because the format was wrong for the buyer stage they were trying to reach. Their competitors, who had published pricing comparison content and “X vs Y” pages, were cited repeatedly.
Implementation: back to how-to
How-to guides reclaim the top spot at 46%. Vendor-published implementation resources account for 17%, and agency guides for 14%. Templates appear here at 5%, the only stage where they make a meaningful contribution.
Data and statistics are almost absent at this stage. Under 15% of implementation-stage citations contained specific numbers. The content that gets cited here is procedural and step-by-step rather than data-driven. Job descriptions and hiring templates also appear uniquely at this stage, at 3%.
How ChatGPT, Perplexity, Gemini, and Claude Cite Different Sources in AI Search
Perplexity cites on 100% of queries (8.0 avg), Gemini cites the most per query (11.9 avg), Claude skips 25% of queries, and ChatGPT GPT-5.4 only cites on 65% of queries. But the real story is what ChatGPT cites: 74.6% vendor websites, compared to 79.0% third-party on the other 3 platforms (for product queries).
Perplexity is the most consistent. 100% citation rate, 6 to 10 sources on every single query, no exceptions. It is the most balanced across source types: it cites YouTube at 3% (highest of the 4), includes research reports more often than the others, and rarely over-indexes on any single domain. If you want to know whether your content is citable, Perplexity is the most reliable test.
Gemini cites the most: 11.9 sources per query on average. It hits 100% citation rate but has a tendency to cite the same domain 5 to 7 times in one response. It also generates 3.7 fan-out sub-queries per prompt on average, with comparison queries generating up to 7. Vendor product pages appear in 21% of Gemini citations, the highest of the 4.
Claude skips 25% of queries entirely, usually educational and definitional ones. When it does cite, it averages 5.5 sources per query overall, rising to 7.4 when it does choose to search the web. Claude has a noticeable preference for statistics roundup posts, often citing a single data-rich source 5 or more times within one response. It never cites YouTube or Reddit.
ChatGPT GPT-5.4 is where the study gets interesting, because we expected it to behave like the other 3 platforms with minor differences, but it didn’t.
This is not a new playbook. Writesonic’s comparison of GPT-5.3 and GPT-5.4 found a similar pattern: vendor website citations jumped from 8% to 56% between model versions. ChatGPT using site: operators to go directly to vendor pricing pages is traditional SEO being rewarded by AI search. The websites that have clear pricing, strong product pages, and well-structured documentation are the ones ChatGPT finds and cites. That has always been the goal of good SEO.
GPT-5.4 cites vendor websites in 74.6% of its product and comparison queries. Blog citations collapsed to around 15%. When it searches for “Klaviyo vs Mailchimp,” it does not search for comparison blog posts. It runs fan-out queries like site:klaviyo.com pricing and site:mailchimp.com pricing, pulling data directly from each vendor’s pages.
We saw site: operators in approximately 50% of GPT-5.4’s fan-out queries, always targeting vendor domains. Examples from our dataset: site:hubspot.com marketing hub pricing email marketing official, site:gong.io platform revenue intelligence, site:apollo.io pricing.
G2 appeared in GPT-5.4’s reasoning text as a trust verification signal (“maybe including G2 for reliable insights,” “G2 or Capterra could be helpful”), but it rarely made it into the final citations. G2 was directly cited in only 3 of 75 queries. The model uses G2 as a cross-reference rather than a primary source.
We spent days designing this study expecting to find broadly similar patterns across all 4 platforms. ChatGPT threw that assumption out. It is building responses from a fundamentally different source hierarchy than the other 3.
Why Pages with Specific Data Get Cited More in AI Search
AI models prefer pages with concrete numbers they can extract and quote. The data density varies dramatically by buyer stage, and the type of data that gets cited is highly specific: pricing tables, G2 scores, benchmark percentages, and ROI figures.
The pattern here is straightforward: pages with specific numbers get cited, and pages without them don’t, especially at the decision stage. At the decision stage, 80% of cited pages contained quantitative data like pricing tiers, benchmark percentages, G2 scores, or ROI figures. At the implementation stage, that dropped to just 15% because how-to content is procedural rather than numerical.
We saw this play out consistently across our dataset. When Avoma published G2 scores in a comparison post, the AI quoted those exact numbers and cited the page. When Moosend included deliverability statistics in an email marketing comparison, the AI extracted and cited those figures. The pattern is specific: if you include a concrete number that answers a buyer’s question, AI will quote it and link to your page as the source.
What AI Search Uses vs What It Cites: The Hidden Influence of Reddit, YouTube, and LinkedIn
Citations capture what users see. The AI consults far more than it shows. Gemini generates 3.7 sub-queries per prompt. GPT-5.4 uses site: operators in its fan-out queries. Reddit occupies a significant share of ChatGPT’s internal search slots but less than 1% of visible citations. The full influence map is wider than the citation layer.
Our Gemini data showed 3.7 fan-out sub-queries per prompt on average. For comparison queries, that number climbed to 5.0, with a single “X vs Y” query generating up to 7 sub-queries covering features, pricing, and reviews for each product separately. Every sub-query consults different sources, most of which never appear in the final response.
GPT-5.4’s fan-out queries contain site: operators targeting vendor domains. We can see it going directly to specific websites before composing its answer. It searched for Accenture’s Salesforce implementation cost page, Grow and Convert’s pricing data, and a16z’s hiring advice, but none of these appeared in the final citations because the consultation happened invisibly.
The gap between what AI consults and what it cites is vast. In our dataset, Reddit accounted for under 1% of visible citations. But OpenAI pays for licensed access to Reddit data, and ChatGPT’s own search activity leans on Reddit far more than its citations admit. YouTube is among the most-cited domains in Google AI Overviews, and professional networks surface constantly in answers. The visible citation layer undersells all 3.
Our citation data captures what users see. The real influence includes training data, fan-out consultation, and internal retrieval that never surfaces in the response. This is why we build client presence on the platforms AI learns from as well as the ones it cites. Reddit, LinkedIn, YouTube, G2: these platforms shape the answer even when they don’t get the attribution. Any AI visibility strategy that ignores them is working with half the picture.
What B2B SaaS Content to Publish for AI Search Visibility at Each Buyer Stage
Each buyer stage demands a different content format, data density, and engine focus. For ChatGPT, optimize your own product and pricing pages. For Perplexity, Gemini, and Claude, third-party blog content still dominates. You need both strategies running in parallel.
- Awareness
How-to guides and glossary pages. Educational content about the problem your product solves; glossary pages peak here at 18%, and all 4 platforms cite this stage.
- Consideration
Listicles with comparison tables. Best-X-for-Y content is 50% of citations here; include your product alongside competitors, with pricing and integration data.
- Evaluation
X-vs-Y comparison pages. Third-party comparisons get cited at 3x the rate of vendor us-vs-them pages; specific numbers throughout.
- Decision
Pricing and cost guides. 80% of cited pages here carry data. For ChatGPT, your own pricing pages; for the other 3, third-party cost analysis.
- Implementation
Step-by-step how-to guides with templates. Perplexity and Gemini cite this stage reliably; vendor documentation performs at 17%.
This is what our service lines look like in practice. SEO & Content is blog content matched to buyer stage, with the right format and data density. AI Search Optimization covers your own product and pricing pages, especially for ChatGPT, which now goes there directly. Digital PR builds genuine presence on the platforms AI consults behind the scenes.
Why Original Data Drives AI Search Visibility Across Every Content Format
Formal research reports made up just 2% of citations in our dataset, but that figure is misleading on its own. Original data does not only live in research reports: it shows up inside blog posts, comparison pages, pricing guides, and listicles, which collectively account for 66.9% of all citations. The pages AI cites most are the ones with specific, original numbers that no one else has published.
There is an important distinction between original data and research reports. A research report is a format: a standalone study with methodology and findings. Original data is information that appears in any format. When Cleanlist.ai tested 15 data enrichment providers and published accuracy scores inside a listicle, that was original data in a listicle format, and AI cited it. When Avoma published G2 scores inside a comparison blog post, that was original data in a comparison format, and AI quoted those exact figures. The format was a blog post, but the reason it got cited was the original numbers inside it.
This is central to how we think about content at VisibleIQ. We do not create standalone research reports and hope AI cites them. We embed original data into the content formats AI already cites most: comparison pages with real pricing we have compiled, buyer guides with audit findings from our own client work, listicles with test results and benchmarks we have collected ourselves. The original data is what makes the content citable. The format is what gets it in front of the AI in the first place.
On top of that, original data has a compounding effect. This blog post is itself original research, and if publications link to it, that builds our domain authority, which means our service pages, comparison content, and buyer guides all become more visible in AI search over time. Brand search volume is the strongest predictor of AI citation, and original data is what builds brand search volume. We have seen this compounding effect first-hand: one B2B SaaS client added $1.1M in ARR from organic and AI search in 8 months.
Findings summary
| Finding | What it means | What to do |
|---|---|---|
| 79.0% of citations go to someone else’s website | What others say about you matters more than what you say about yourself. | Build third-party presence: comparisons, reviews, and guest content on independent sites. This is what our Digital PR pillar delivers. |
| ChatGPT GPT-5.4: 74.6% vendor citations | ChatGPT goes directly to vendor sites for product queries. | Optimize your product and pricing pages. This is what our AI Search Visibility audits identify. |
| Format changes by buyer stage | A how-to guide will not get cited for a consideration query. A listicle will not get cited at the decision stage. | Map your content library to buyer stages. Fill format gaps per stage. |
| site: operators in GPT-5.4 fan-outs (~50%) | ChatGPT goes directly to vendor domains, bypassing third-party content. | Make your key product and pricing pages easily crawlable with clear structured data. |
| Data table effect: 80% at decision stage | Pages with specific numbers get cited at the decision stage; pages without them do not. | Add pricing tables, G2 scores, benchmark figures, and ROI data to decision-stage content. |
| G2: trust verification rather than citation | GPT-5.4 references G2 in its reasoning but rarely cites it directly (3 of 75 queries). | Maintain strong G2 ratings. They influence the AI trust model even without direct citation. |
| Blog citations dropping on ChatGPT (~55% to ~15%) | Blog posts that dominate Perplexity, Gemini, and Claude are barely cited by ChatGPT. | Diversify beyond blog content. Product pages and documentation matter on ChatGPT. |
| Fan-out queries: 3.7 avg on Gemini, site: on GPT-5.4 | AI models research extensively before responding. Most sources consulted are never cited. | Build presence on platforms AI consults behind the scenes: Reddit, YouTube, LinkedIn, G2. This is our Digital PR pillar in action. |
| Research reports = 2%, original data in 66.9% | Original data is not a format, it is what makes any format citable. Pages with specific numbers get cited regardless of format. | Embed original data into every content format you publish. This is central to our content strategy at VisibleIQ. |
| Perplexity: most reliable for third-party content | 100% citation rate, 8.0 avg citations, most balanced source profile. | Use Perplexity as your benchmark for AI citability. If your content gets cited there, it is likely citable elsewhere. |
ChatGPT rewarding vendor websites is traditional SEO winning. Perplexity, Gemini, and Claude rewarding third-party content is content marketing and digital PR winning.
AI search did not invent new work. It made the work we already do more measurable and more urgent. Your own website needs to be ready for ChatGPT to visit directly, and other people need to be writing about you for the other 3 platforms. Both have always mattered. Now we can prove it with data.
That is what we do at VisibleIQ.
Download the dataset
We are publishing the full dataset: 2,020 citations classified by engine, buyer stage, content format, and source type. Download it, run your own analysis, cite us if you find something we missed.
2,020 citations. 75 queries. 4 platforms. 5 buyer stages. Raw data in CSV.
Frequently asked questions
Which content format gets cited most by AI for B2B SaaS queries?
Blog posts (guides, listicles, and comparisons) account for 66.9% of all citations in our dataset. But the dominant format changes by buyer stage. Listicles make up 50% at consideration, comparison pages 42% at evaluation, pricing guides 30% at decision, and how-to guides 46% at implementation. There is no single “best” format. The right format depends on which buyer stage you are targeting.
Does ChatGPT cite the same sources as Perplexity?
No. In our study, ChatGPT GPT-5.4 cited vendor websites in 74.6% of its product and comparison queries. Perplexity, Gemini, and Claude cited third-party content 79.0% of the time for product queries. ChatGPT uses site: operators in its fan-out queries to go directly to vendor pricing and product pages. Perplexity and the other platforms rely on independent blogs, comparison sites, and listicles. A content strategy built only for Perplexity will underperform on ChatGPT, and vice versa.
Do you need high domain authority to get cited by AI?
Not necessarily. Niche sites like saashero.net and emailvendorselection.com were cited across all 4 platforms in our dataset. What matters more is topical authority (consistent content in a specific category), data density (pages with pricing tables and specific statistics), and recency (current-year dates in titles and URLs). Small domains with strong category focus regularly outperformed larger publications that covered the topic more broadly.
What are fan-out queries and why do they matter?
Fan-out queries are sub-queries that AI models generate before composing a response. In our data, Gemini generates 3.7 fan-out queries on average, with comparison queries generating up to 7. ChatGPT GPT-5.4 uses site: operators in roughly 50% of its fan-out queries. They matter because they determine which sources the AI consults, even if users never see the sub-queries. A vendor domain that appears in a fan-out query influences the response even when it does not appear in the final citations.
Is it better to optimize for ChatGPT or Perplexity?
You need different strategies for each. For ChatGPT visibility: optimize your own product pages, pricing pages, and documentation, because GPT-5.4 goes directly to vendor sites. For Perplexity, Gemini, and Claude: invest in third-party content like comparison articles, listicles, and data-rich guides. These are not separate strategies. They are parts of the same approach: content and digital PR that covers every engine at once rather than one. This is the approach we take at VisibleIQ.
Methodology: 75 B2B SaaS buyer queries across Perplexity sonar-pro, Gemini 2.5-flash, Claude haiku-4.5, and ChatGPT GPT-5.4. Data collected 28 March 2026, one pass per engine on the same day. All citations manually classified. GPT-5.4 analysis based on 75 queries with web search enabled (49 triggered web search, 26 answered from training data). Total citations: 2,020 across all platforms. Contact rafael@bevisibleiq.com for questions about the methodology.
