We counted the names in 26,607 answers
When a buyer asks an engine for the best tool in a category, they get a short list. How short it is decides how many companies get looked at at all.
The sample is every answer from ChatGPT, Claude, Perplexity and Google AI Overviews collected from October 4 to November 1, 2026, for a buyer question that doesn't name the tracked brand: 26,607 answers to 3,470 questions for 187 brands in 116 accounts. We dropped repeated identical answers.
A name is any company or product the answer mentions, as Proofsource extracts it. Every one of the 132,032 extracted names appears word for word in its answer.
Three engines named five. Perplexity named three.
We split off the 3,222 answers to "X vs Y" questions and looked at the other 23,385, the questions that ask for options. The median number of names per answer:
- ChatGPT: 5 (average 6.4). Three or fewer in 17.5% of answers (95% interval 16.6% to 18.5%).
- Claude: 5 (average 5.6). Three or fewer in 12.6% (11.8% to 13.5%).
- Google AI Overviews: 5 (average 5.6). Three or fewer in 15.8% (14.9% to 16.8%).
- Perplexity: 3 (average 3.6). Three or fewer in 41.0% (39.7% to 42.3%).
Perplexity picked one name four times as often as ChatGPT
Perplexity named exactly one company in 8.8% of these answers. ChatGPT did it in 2.2%. At the other end, ChatGPT named more than ten in 9.2% of its answers, and Perplexity in 1.5%.
So a brand that's usually the fourth or fifth name in its category is in the answer on three engines and mostly out on Perplexity. MaxAEO's June 2026 audit of 300 ChatGPT answers to B2B software questions found a median of 4. Our ChatGPT median is 5, across a wider spread of categories, from software to rooftop solar.
Comparison questions got two names on every engine
The 3,222 answers to "X vs Y" questions had a median of two names on all four engines, and 78.0% named three or fewer. The engine answers the question it was asked. You only get into those answers when a buyer types your name next to a competitor's, so don't read them as a shortlist.
And 9.9% of the option questions named nobody: Perplexity 14.3%, Claude 11.4%, ChatGPT 7.8%, AI Overviews 6.3%. Most of them asked what to look for and got criteria back.
Each answer showed five names. Each question used 18.
356 pairs of question and engine were asked at least eight times in the month, 11 times at the median. Over those asks, the median question collected 18 different names: ChatGPT 14, Perplexity 13, Claude 20 and AI Overviews 25.
About four of them appeared in at least half the answers. Seven appeared only once. So the list a buyer sees is a small core plus a few names drawn from the edge, and the edge changes from one ask to the next.
What we'd do
- Ask each of your buyer questions several times on each engine and write down every name, with how often each one appeared.
- If you're in half the answers or more, you're in the core. If you appear once in ten asks, you're on the edge. Read the pages the engine cites when it names the core, because those are the pages you need to be in.
- Score Perplexity on its own. With three names, fourth place there mostly means absent.
- Keep "X vs Y" answers out of your shortlist numbers. They tell you how you compare, not whether you get picked.
How Proofsource handles it
Proofsource asks each buyer question on every engine on a schedule and records every company or product each answer names, beside the competitors you entered. So you can see the whole pool for a question and how often each name made the list.
Where this could be wrong
A mention isn't always a recommendation. An answer that names a company to rule it out still counts here, and one brand's product lines can count as two or three names.
These are brands our users chose to track, many of them young, in categories from software to local services. A category with three real vendors will give short lists everywhere. The questions are mostly English.
The pool figure covers only 356 question and engine pairs, and it grows with every extra ask, so 18 holds for about 11 asks. Engines change, and a model update can lengthen or shorten lists overnight.
