highsalience/chatgpt-software-recommendations

ChatGPT Software Recommendations: 21 Categories, 2,100 Answers

This dataset records which software brands ChatGPT names, recommends and picks when buyers ask about 21 software categories, and which websites it cites. It covers 2,100 ChatGPT answers (100 buyer questions in each of 21 categories), coded brand by brand, with Google's organic top 10 for the same questions as the control.

It is published by High Salience, an AI search and SEO agency. Every file here is also published on the matching research page at highsalience.com/research, where each category has a written teardown.

What is in the dataset

Measure	Count
Software categories	21
Buyer questions (one ChatGPT answer each)	2,100
Brand appearances (recommended or listed)	8,218
Recommended (the answer assigns the brand to a case or fit)	7,997 of 8,218
Picked (the answer itself chooses the brand)	6,475 of 7,997
Cited source links	6,443
Brands in the category dictionaries	1,074

How the data was collected

• Surface: ChatGPT with web search on, United States, English, one run per question, collected through the DataForSEO scraper.
• Dates: September 8 to September 26, 2026. The collection date of each category is in teardowns.csv.
• Questions: 100 per category in four intents: category ("best CRM software"), comparison ("salesforce vs hubspot"), alternative ("salesforce alternatives") and recommendation ("which CRM should I use for a 5 person sales team").
• Control: Google's organic top 10 for the same question on the same date.
• Brands: only brands in each category's dictionary (brands.csv) are coded. Aliases and domains are listed per brand.

How the answers were coded

Two independent AI readers (Claude agents, not people) coded every answer blind from a written protocol, reader-protocol.md. A third reader settled the 204 codes on which they disagreed (2.2 percent of 9,252 brand codes). The two readers gave the same one of six codes 97.8 percent of the time, and agreed 99.2 percent of the time at the recommended level. The full rulebook and its dated amendments are in codebook.md.

The codes, strongest first:

• picked: the answer itself chooses the brand, overall or for a case ("my pick", "choose X if").
• recommended: the answer assigns the brand to a stated case or fit ("best for small teams"). In mentions.csv a picked brand has mention_type = recommended and picked = 1.
• listed: presented as an option with no case or fit.
• passing: named, but not offered to the buyer as an option.
• anchor: the brand named in an "alternatives to X" question. It is excluded from every share.

Files

File	Rows	What one row is
teardowns.csv	21	One category: collection date, counts and the teardown page
queries.csv	2,100	One buyer question
mentions.csv	9,211	One brand named in one answer
citations.csv	6,443	One source link cited in one answer
observations.csv	2,100	One answer: counts of brands, recommendations and sources
brands.csv	1,074	One brand in one category dictionary
reddit-presence.csv	700	One question from categories 1 to 7: Reddit in Google, in the AI Overview and in the ChatGPT answer
youtube-presence.csv	700	One question from categories 1 to 7: YouTube in Google, in the AI Overview and in ChatGPT's citations
industry-queries.csv	100	One buyer question in fintech, healthcare, legal or ecommerce (25 each), collected September 26, 2026
industry-per-question.csv	100	Google top 10, AI Overview sources and ChatGPT cited domains for each industry question
industry-chatgpt-citations.csv	554	One ChatGPT citation on an industry question
codebook.md		Coding rulebook, version 1.0 through amendment 2.0
reader-protocol.md		The protocol the AI readers followed

Every teardown table starts with teardown (1 to 21) and category. The id column is the question number inside that category (q001 to q100), so join tables on teardown plus id.

mentions.csv

Column	Meaning
intent	category, comparison, alternative or recommendation
query	The buyer question
brand	Brand name from the dictionary
position	Order of first appearance in the answer, counting from 1
mention_type	recommended, listed, passing or anchor
sentiment	positive, neutral or negative
in_query	1 if the question names the brand
brand_domain_in_google_top10	1 if the brand's own site was in Google's organic top 10 for the same question
bolded	1 if the answer bolds the brand name
in_table	1 if the brand appears in a table
picked	1 if the answer itself chooses the brand

citations.csv

Column	Meaning
domain	Registrable domain of the cited page
class	first-party (the site of a brand in that category's dictionary), review-or-media, or other-third-party
url	The cited address as ChatGPT returned it
domain_in_google_top10	1 if that domain was in Google's organic top 10 for the same question

observations.csv

Column	Meaning
status	ok for every answer in this release
n_brands	Dictionary brands named in the answer
n_recs	Brands coded recommended
n_listed	Brands coded listed
n_sources	Sources cited
google_top10_brand_sites	Google top 10 results that are sites of dictionary brands
google_top10_third_party	Google top 10 results that are any other site

Load it

from datasets import load_dataset

mentions = load_dataset("highsalience/chatgpt-software-recommendations", "mentions", split="train")
citations = load_dataset("highsalience/chatgpt-software-recommendations", "citations", split="train")

The same files are on GitHub at High-Salience/chatgpt-software-recommendations.

Limits to keep in mind

• One run per question. ChatGPT's answers vary between runs, so a brand missing from one answer is "not observed in this run", not proof that it is never named.
• The readers are AI models. Their agreement with each other is high, but it is not the same as agreement with a human panel.
• Only dictionary brands are coded. An answer can recommend a brand outside the dictionary, and that brand does not appear in mentions.csv.
• United States, English, September 2026. Answers change over time and by location.

Citation and license

Please cite as: High Salience (2026). ChatGPT Software Recommendations: 21 Categories, 2,100 Answers. https://highsalience.com/research/

The data is released under Creative Commons Attribution 4.0. You may use and adapt it, including commercially, with a credit and a link to High Salience.

Questions and corrections: hello@highsalience.com