ChatGPT vs Google, measured on the same software questions, comes out lopsided: 74 percent of ChatGPT’s software recommendations in my 21 teardowns went to brands whose own site was outside Google’s top ten for that question. That’s 5,953 of 7,997 recommended brand mentions, across 2,100 answers.
I ran the teardowns for High Salience between September 8 and 26, 2026, one category at a time, from project management and CRM to password managers, e-signature and AI writing tools. Each category got 100 buyer questions, asked once in ChatGPT, logged out with web search on, with Google’s top ten organic results for the same questions as the control. Every category’s tables and data files are on High Salience’s research page. This is the whole series read as one story.
The 21 software categories in one count
The 21 software categories in one count come to 2,100 answers and 8,218 brand appearances, where an appearance means a brand was named as an option.
| Measure | Count |
|---|---|
| Answers | 2,100 |
| Brand appearances | 8,218 |
| Recommended, tied to a stated fit | 7,997 |
| Picked, the answer’s own verdict | 6,475 |
| Citation events | 6,443 |
Recommended and picked are two different levels. A brand is recommended when the answer ties it to a stated case, like “best for small teams.” It’s picked when the answer makes it the verdict: “my pick,” “choose X if,” a shortlist it tells you to act on, or the top of its own ranking. Of the 8,218 appearances, 7,997 came with a fit and 6,475 were picks. When ChatGPT offered a software brand as an option, it almost always said who the brand was for.
How many ChatGPT recommendations went to brands Google didn’t rank
ChatGPT recommendations went to brands Google didn’t rank 5,953 times out of 7,997, or 74 percent. For picks the share was 72 percent of 6,475. A brand counts as ranking only when its own domain sits in Google’s top ten for that exact question. A “best of” list that mentions the brand doesn’t count, because the question is whether Google sends the buyer to the brand’s own site.
The share moved a lot from one category to the next.
Business intelligence sits at the top of the series and payroll at the bottom; the 13 categories not shown fall between them. Payroll’s low share fits what Google showed for it, in my read: vendors’ own sites ranked often there, and gusto.com alone was in Google’s top ten for 58 of the 100 payroll questions. My CRM breakdown shows what a 70 percent category looks like from the inside.
Ranking in Google didn’t make a brand ChatGPT’s pick
Ranking in Google didn’t make a brand ChatGPT’s pick, and the clearest case is Domo. In business intelligence, domo.com was in Google’s top ten for 62 of the 100 questions, more than any other vendor’s site, mostly with articles from its Learn section. ChatGPT named Domo in 10 answers and picked it in 3, all of them head-to-heads with Domo in the question. In the 96 questions that didn’t name Domo, it was picked in none.
Domo ranked for 62 questions and was picked in 3.
The gap ran the other way too. Google ranked adobe.com for 4 of the 100 e-signature questions, and ChatGPT cited it in 35 answers. In SEO tools, ahrefs.com ranked for 9 questions and was cited in 42 answers. In data warehouses, snowflake.com ranked for 3 questions and was still the most-cited domain in the category.
The sites ChatGPT cited for software
The sites ChatGPT cited for software were mostly not the ones Google ranked. Across the 21 categories, between 18 and 43 percent of ChatGPT’s citations came from domains in Google’s top ten for the same question, with payroll at the top of that range. In the first seven categories, vendors’ own sites took 51 percent of 2,311 citation events.
Reddit is the sharpest case. It ranked in Google’s top ten for 631 of the first 700 questions, and none of the 6,443 citations in these 2,100 answers pointed to reddit.com. That gap gets its own post.
How often the top brand was ChatGPT’s pick
How often the top brand was ChatGPT’s pick shows how settled each category is. The most-picked brand in each category was picked in a median of 61 of 100 answers, from Mailchimp’s 44 in email marketing to Bitwarden’s 78 in password managers. Payment processing had the most answers with no pick at all, 22 of 100, and 84 of its answers asked for the buyer’s own numbers.
How I ran the 21 teardowns
How I ran the 21 teardowns stayed the same in every category, so the categories can sit side by side.
- Queries
- 100 per category in 21 software categories (30 category, 30 comparison, 20 alternative, 20 recommendation)
- Surface
- ChatGPT, logged out, web search forced on, US English, one run per query, via the DataForSEO scraper
- Window
- September 8 to 26, 2026, one collection per category
- Control
- Google’s top ten organic results for the same queries in the same window (October 23 for 481 swapped pages)
- Coding
- Two AI readers (Claude agents) working blind from a written protocol, a third settling disagreements (codebook v2.0)
- Agreement
- 97.8 percent of brand codes, 99.2 percent at the recommended level
- Ranking test
- A brand ranks only when its own domain is in Google’s top ten for that question
I corrected the Google side of this study on October 23, 2026. Since late September, Google has answered some automated searches with swapped pages full of the wrong sites, and 14 of these categories took their Google control on September 26. Of those 1,400 pages, 481 shared under half their sites with Google’s results for the same question on October 23, against 7 of the 699 pulled earlier. I used the October 23 page for those 481, which moved the headline from 78 to 74 percent.
To run the same test on your own category, see how I check which brands ChatGPT recommends.
What the ChatGPT vs Google gap means for software brands
What the ChatGPT vs Google gap means for software brands, in my read, starts with the rank report. A rank report shows the brands whose own sites Google puts in front of buyers, and in these answers those brands got only 26 percent of ChatGPT’s recommended mentions. It needs a second list beside it: a fixed set of buyer questions, asked the same way every month, coded for recommended and picked.
Ranking articles is not the same as being chosen. Domo’s articles put its site in front of searchers; they didn’t put Domo in the verdict. If your traffic plan rests on informational pages, check whether the answers that matter to buyers ever pick you.
The stated fit is where the work is. With 7,997 of 8,218 appearances tied to a case, the job is choosing your case and making sure answers turn it into a pick, and one word in the prompt can change which case wins. Being read is not enough either: a site can be cited in answer after answer and still not be the name the answer gives.
Sources
- AI Search Research: How ChatGPT and Google Pick Brands (High Salience)
- ChatGPT vs Google: How Their Answers Differ (High Salience)
- 100 Business Intelligence Questions: What ChatGPT Recommends (High Salience)
- 100 Payroll Software Questions: What ChatGPT Recommends (High Salience)
- ChatGPT Software Recommendations: 21 Categories, 2,100 Answers (Hugging Face)