Methodology

We took the first study’s 5,983 software brands in 196 categories and asked about them in three ways:

• Best tools for X: “What are the best tools for {category}?”, once per category (196).
• Alternatives to X: “What are the best alternatives to {product} for {category}?” (11,525).
• X vs Y: “How do {product} and {competitor} compare for {category}?” (6,363).

The first study’s questions about a single brand are left out, because they don’t ask a platform to choose between tools.

We reused the first study’s answers, collected over two weeks in August 2026, so every platform answered the same questions in the same fortnight.

• Google search and AI Overviews come from Google’s results page, read through DataForSEO: the top ten results, plus the AI Overview on the 17,522 questions that showed one.
• ChatGPT and Gemini were asked in their apps, also through DataForSEO, so the answers are the ones a user sees.
• Claude (Haiku 4.5) and Perplexity (Sonar) were asked through their APIs with web search on. Claude decides for itself when to search, and didn’t on about one question in five.

Two parts were asked again later. ChatGPT’s searches for X vs Y questions were lost in August, with only 181 of 6,363 kept, so we asked every X vs Y question again in October 2026, and 6,361 came back with their searches. Perplexity’s follow-up questions weren’t requested in August, so we asked 2,000 of the questions again in September, a 10% sample.

Every chart counts citations: a page cited in one answer. If an answer cites the same page twice, it counts once. Addresses are cleaned up first and sites are grouped by their main domain, so learn.g2.com counts as g2.com.

For three platforms we had to decide what counts as cited:

• For Perplexity, only the pages its answer actually marks, about 9 per answer, not the 20 or so its searches found.
• For ChatGPT, the product links its app adds to brand names are left out. They point to homepages, aren’t part of the answer, and were 14% of its pages.
• For Google, links back to Google itself, such as Shopping panels, are left out.

That leaves 686,749 citations of 311,766 different pages.

We downloaded 226,844 of the cited pages once, in September 2026, without running JavaScript, and 181,352 of them had text we could read. The other 84,922 sit on review sites, forums, YouTube and social networks, which mostly block crawlers. Like the 45,492 pages that came back empty, they were labelled from their address instead. That leaves 6,365 pages whose address says nothing, and they have no label.

Each readable page was labelled by TypeSafe’s JEV classifier, which reads the page’s address, title and cleaned-up text and answers a fixed set of questions about it. Two apply to every page, who published it and what kind of page it is, so a listicle on G2 and a listicle on a vendor’s blog stay apart. List pages also get where the publisher’s own product sits and the patterns in “What self-promotional listicles do”. The same classifier labelled the AI answers and ChatGPT’s searches, while simple things like counts, years, tables and “site:” are matched in the text.

Before the full run we tested the questions on 600 texts against a stronger model, OpenAI’s gpt-6-luna, reworded or dropped the weak ones, and read the disagreements by hand. On pages the two agree on 91% of publishers, 92% of page types, 96% of whether the publisher sells in the category and 86% of where its own product sits. Where the classifier is unsure, as with a page’s tone, the label is shown as unclear.

• A page is self-promotional when it is a software vendor’s listicle or comparison, in a category the vendor sells in, that includes the vendor’s own product.
• Pages we couldn’t read count as not self-promotional, so every self-promotional share is a floor.
• Shares count citations, so a page cited in ten answers counts ten times.
• Charts about what answers cite leave out answers that cite nothing: 22% of Claude’s and 29% of Gemini’s.
• The answer patterns leave out “it depends”, which the classifier was too often unsure about, and 2,234 AI Overviews captured without text.