B2B SaaS AI Citation Study: We Classified 2,020 Citations Across ChatGPT, Perplexity, Gemini, and Claude

AI citation sources change with the buyer’s question and the platform answering it. Our March 2026 B2B SaaS study examined those differences across content formats, domains, and buyer stages. It provides a historical view of the sources cited in a defined sample, rather than a current ranking of the whole web.

At VisibleIQ, we use research like this to inform AI Search Optimization for B2B SaaS and AI-native companies. The useful question is specific: when a buyer compares software in your category, which sources does the answer cite, and what do those sources say?

How did we collect and classify the AI citations?

We tested 75 B2B SaaS buyer queries across four platforms, with 15 queries at each of five buyer stages. The original analysis, compiled on March 28, 2026, estimated approximately 2,020 citations. We keep that original sample separate from the later GPT-5.4 product-query analysis when presenting the findings below.

The original runs used Perplexity sonar-pro, Gemini 2.5-flash, and Claude haiku-4.5. The original ChatGPT collection combined GPT-4o API responses and a ChatGPT scraper. It was not a GPT-5.4-only dataset. The separate product-query calculation was compiled on March 30 and compared a GPT-5.4 check with the three other platforms.

Queries covered awareness, consideration, evaluation, decision, and implementation. Examples included “What is product-led growth,” “HubSpot vs Salesforce total cost of ownership,” and “How to set up HubSpot for a SaaS company.” We classified displayed citation URLs by content format, source type, and buyer stage.

Counts represent recorded citation appearances, which can include repeated domains across answers. Domain counts and format shares from the aggregate review are estimates. The product-query tables use the separate recorded classification counts and show their denominators. Their percentages describe citations, not the proportion of answers recommending a brand.

This was a small, selected sample. It can’t establish citation patterns for every category, a trend over time, or the cause of a citation. Model versions, prompts, collection methods, and dates matter. No new AI responses were collected for this October 6 article revision.

How often did AI cite vendor websites versus third-party sources?

Across the 45 product queries on Perplexity plus Gemini and Claude, 79.0% of recorded citations pointed to third-party sources and 21.0% to the product’s own website. The separate GPT-5.4 analysis showed a different split. These are citation shares for the selected product questions, with each collection’s counts kept distinct.

We defined the product’s own website as a domain belonging to a product being recommended, compared, or evaluated in the answer. Other domains were classified as third-party sources, including comparison publishers, agencies, review sites, and other companies. Awareness and implementation questions were excluded from this product-query calculation.

The stage breakdown below applies to Perplexity plus Gemini and Claude only. Its three rows reconcile to the 1,243-citation total above.

That decision-stage result is useful for planning. A vendor’s own pricing page is one source a buyer can encounter, but independent cost explanations also appeared in the sample. We’d inspect both when checking whether AI is giving buyers accurate pricing information. Our guide to reviews and third-party mentions covers the wider source picture.

Which content formats got cited at each buyer stage?

In the original aggregate analysis, the leading format changed with the question. Listicles were prominent at consideration, third-party comparisons at evaluation, pricing guides at decision, and how-to guides at implementation. The table shows the archived estimates for those stages; it doesn’t imply that other formats can’t appear.

Across the original collection, the estimated shares for how-to or strategy blog posts, listicles, and comparison pages were 30.9%, 20.0%, and 16.0%. Together they account for 66.9%. That is a classification of cited content. It doesn’t measure how likely a newly published page of each type is to get cited.

A team choosing software needs selection criteria and tradeoffs. A team setting it up needs instructions. We use the buyer’s question to choose the page’s job before deciding its format. The page-type guide applies this distinction to a content plan.

Which domains were most cited in our B2B SaaS sample?

HubSpot, Salesforce, and saashero.net led the archived domain table, followed by a mixture of software vendors and specialist publishers. The counts below are approximate citation appearances from the original March sample. They exclude the separate GPT-5.4 check and should not be read as a current ranking across AI search.

The prompt list included named vendors and SaaS categories, which affects which domains can appear. HubSpot’s position here doesn’t establish that it leads for an unrelated market. For your own category, inspect a fixed set of buyer questions and retain the cited pages, rather than importing this list as a target list for outreach.

We’ve retained the archive’s approximate counts and omitted percentage shares or per-platform counts that this domain summary doesn’t substantiate. The research notes are available by request below.

How did the four AI platforms differ?

The original collection returned different citation volumes and source mixes across the four platforms. Perplexity and Gemini returned citations for all sampled queries, while Claude and the original ChatGPT collection had gaps. The later GPT-5.4 product check belongs to a separate collection, so it should not replace the original ChatGPT row.

The figures describe the collection methods as well as the models. They are not a controlled test proving one consumer product cites more than another. A higher citation count also doesn’t establish that the answer recommended a vendor more strongly or described it more accurately.

The separate GPT-5.4 analysis is useful because vendor-owned pages accounted for 74.6% of its recorded product-query citations. Keep your product information accurate and make it easy to inspect. The observation doesn’t justify abandoning third-party coverage or assuming future versions will behave identically.

What did cited pages contain at the decision stage?

The archived analysis found decision-stage pages frequently contained pricing, ROI, or benchmark data, with estimates ranging from 73% to 90% across the reviewed groups. This is an observation about the cited pages in our sample. We didn’t test whether adding a data table caused a page to be cited.

Specific information is useful when a buyer asks about cost or suitability. Publish what the buyer needs to evaluate the claim: current pricing assumptions, what is included, and the source or method behind a calculation. Avoid inserting unsupported numbers to make a page look more authoritative.

Can citations show everything that shaped an AI answer?

Displayed citations show the sources attached to an answer. They don’t expose every influence on the response or establish the complete retrieval process. Our study records visible citations, so it can’t quantify the contribution of training data, unlinked mentions, or sources that never appeared in the displayed answer.

That limit matters for measurement. A company can be mentioned or recommended without its website being cited, and a cited source can appear in an unfavorable answer. We assess recommendation, accuracy, sentiment, and share of voice alongside the sources. AI search is an influence channel whose commercial effects need separate evidence.

Our AI search measurement guide shows how to connect recorded referrals and buyer-reported discovery to pipeline while preserving that distinction.

What should a B2B SaaS team publish from these findings?

Start with the buying questions your existing pages don’t answer well. Use the study’s format patterns to choose a useful response, then supply evidence specific to your product or category. A comparison needs fair selection criteria; a cost page needs stated assumptions; an implementation guide needs instructions a reader can follow.

Review existing pages before commissioning new ones. Pricing, comparisons, product documentation, and practical guides serve different needs, but two thin articles answering the same question rarely help the reader. The audit walkthrough gives a starting point for examining the sources buyers encounter.

Our SEO & Content work improves the pages buyers need. Social Marketing, Review Management, and Digital PR develop the wider evidence around a company. AI Search Optimization connects that work to what AI says and recommends.

How does original research fit into cited content?

Research reports were an estimated 2.0% of citations in the original format classification. That category doesn’t measure every instance of original data inside a guide or comparison page. We distinguish primary research from the format carrying it, and from third-party statistics that a publisher has collected or quoted.

A team’s own product test can add evidence to a comparison. Quoting a review score adds sourced information, but it doesn’t turn that score into the publisher’s original research. The study can’t tell us that every cited page contained original data, or prove that originality caused the citation.

There is open territory for research into specific buyer decisions. Publish a method, retain the evidence, and explain the limits so other people can use the result. Distribution then gives relevant readers a chance to examine and reference that work.

What can you apply from this AI citation study?

The practical lesson is to inspect the buyer question, the answer, and the sources together. Our sample showed differences between stages and platform collections, which makes a single universal citation tactic hard to defend. Use those differences to decide what to investigate in your category, then track the effect of your changes.

We’d check whether your own product information answers commercial questions, whether independent sources describe you accurately, and whether the content around your category helps buyers compare options. Citations are evidence of where a source appeared. They don’t substitute for recommendation quality or revenue measurement.

How can you access the research notes?

Contact us to request the March research notes behind these tables. Specify whether you need the original aggregate analysis or the separate GPT-5.4 product-query calculation. The figures on this page identify which collection they describe, so you can cite the relevant finding with its sample, date, and limitations.

Frequently asked questions

The study is most useful when its scope travels with the finding. Keep the collection date, query set, and model or method attached to any number you reuse. These answers cover the distinctions that change how the results should be interpreted or applied to a different B2B SaaS category.

Is this a current list of the most-cited domains in AI?

No. It’s an approximate domain summary from our original March 2026 B2B SaaS sample. The prompt set included particular vendors and categories. It can’t establish a current, web-wide domain ranking.

Does 79.0% apply to every AI citation?

No. It applies to the 1,243 classified citations for the selected product queries on Perplexity plus Gemini and Claude. The separate GPT-5.4 product-query analysis had 181 citations and a different source split.

Does the format table predict whether my page will get cited?

No. It describes the formats of pages already cited in the sample. We didn’t measure all eligible pages of each type, so the table can’t provide a probability that a new page will be cited.

Do citations measure recommendation rate?

No. Recommendation rate requires examining whether the answer recommends the company, with a defined prompt set and denominator. A citation identifies a displayed source. Track these separately.

Research compiled in March 2026. Article revised October 6, 2026 to clarify the separate collections, correct the stage breakdown, and add the archived domain table. The revision doesn’t represent a new data collection.

Rafael De Jesus

Founder

Organic growth and AI search optimization specialist. Rafael has added seven figures in ARR to B2B SaaS and AI-native companies, built a 200K+ following, and generated 3 billion+ impressions across search and social. He writes about how to shape what AI says about your brand.