TL;DR
- Ask a different AI engine and you get a different shortlist.
- Reruns on one engine shared 63.6% of brands; two engines shared 43.6%.
- Half of all recommended brands (49.8%) came from only one engine.
- Cited sources overlap even less, about 3% to 10% between engines.
- Track each engine separately and treat one run as a sample.
Ask an AI engine the same buying question twice and it mostly gives you the same shortlist. Ask a different engine and you get a different one. We sent 160 buyer prompts to ChatGPT, Microsoft Copilot, Google Gemini and Google AI Mode three times each, 1,920 answers in total. A rerun on the same engine shared 63.6% of its recommended brands with the earlier answer. Two different engines, given the same prompt, shared 43.6%.
The split runs through every layer of the answer. Half of all recommended brands (49.8%) were named by only one of the four engines, and 20.9% were named by all four. The engines agreed on the #1 pick for 40 of the 160 prompts. ChatGPT and AI Mode repeat themselves most, with about 74% rerun overlap, while Copilot and Gemini reshuffle about half their list each time. The sources behind the answers overlap least of all: two engines share only 3% to 10% of the domains they cite.
Each engine agrees with itself more than with the others
All four engines were more consistent with their own reruns than with any other engine. The closest pair, ChatGPT and AI Mode, shared 51.1% of brands on the same prompt. That is still well below the 63.6% an engine shares with itself on a rerun.
Brand overlap between each pair of engines, same prompt:
- ChatGPT and AI Mode: 51.1%
- Gemini and AI Mode: 45.1%
- ChatGPT and Gemini: 43.5%
- ChatGPT and Copilot: 43.1%
- Copilot and AI Mode: 41.1%
- Copilot and Gemini: 37.8%
Rerun overlap is how much of its brand list an engine keeps when asked again. The last column counts recommendations that no other engine made.
| Engine | Rerun brand overlap | Same #1 in all 3 runs | Brands per answer | Cited domains per answer | Answers with a shopping module | Brands only this engine named |
|---|---|---|---|---|---|---|
| ChatGPT | 73.5% | 110 of 160 | 5.49 | 2.88 | 8.5% | 132 |
| Google AI Mode | 73.9% | 110 of 160 | 6.13 | 5.21 | 7.5% | 192 |
| Google Gemini | 54.5% | 77 of 160 | 6.28 | 2.86 | 40.8% | 491 |
| Microsoft Copilot | 52.4% | 80 of 160 | 6.86 | 4.31 | 57.5% | 546 |
ChatGPT and AI Mode hold their lists, Copilot and Gemini reshuffle
ChatGPT and AI Mode are the stable pair. Their reruns shared 73.5% and 73.9% of brands. Each kept the same #1 pick across all three runs on 110 of the 160 prompts. Most of what they recommend comes back every time: 59.6% of ChatGPT’s brands and 58.6% of AI Mode’s appeared in all three runs.
Copilot and Gemini move much more. A rerun shared 52.4% of brands on Copilot and 54.5% on Gemini, so about half the list changes between answers. Copilot kept its #1 pick through all three runs on 80 prompts, Gemini on 77. Only about a third of their brands showed up in every run (34.6% on Copilot, 36.0% on Gemini).
These two engines also name the most brands per answer: 6.86 on Copilot and 6.28 on Gemini, against 6.13 on AI Mode and 5.49 on ChatGPT. On either of them, a single check can put a brand on the list or leave it off by chance.
Half the recommended brands come from a single engine
Pool all three runs per engine and the 160 prompts produce 2,733 distinct recommendations, where one recommendation is one brand on one prompt. Of those, 1,361 (49.8%) came from one engine only. Another 455 (16.6%) came from two engines and 345 (12.6%) from three. Only 572 (20.9%) came from all four.
On a typical prompt, four brands made all four lists. Everything past those four depended on which engine you asked. Two prompts had no shared brand at all. “Best field service management software” drew 30 brands and “best payroll service” drew 15, and none of them appeared on all four engines.
Copilot and Gemini supply most of the one-engine brands, 546 and 491. AI Mode added 192 that no one else named, and ChatGPT 132. By segment, business services had the thinnest agreement, with a median of 2 brands on all four engines. B2B software and developer tools had 4, and consumer products had 3.
The #1 pick splits the same way. We took each engine’s most frequent first pick per prompt. All four engines had the same #1 on 40 of 160 prompts (25%). They split two ways on 60 prompts and three ways on 51. On 9 prompts, every engine put a different brand first.
Five prompts, four answers
Five prompts show the split clearly.
- Best CRM software. HubSpot is the #1 pick on ChatGPT, Copilot and Gemini. AI Mode puts Salesforce first.
- Best running shoes. ASICS leads on ChatGPT and AI Mode. Copilot picks Salomon and Gemini picks Nike.
- Best smartwatch. Apple Watch leads on three engines. Copilot puts Garmin first.
- Best answering service. The engines named 28 brands between them. Smith.ai was the only one all four named.
- Best JavaScript framework. Six of the 14 brands named (React, Next.js, Vue.js, Svelte, Nuxt and Angular) appeared on all four engines, one of the highest agreement rates in the study.
Some categories do settle on one answer everywhere. Sony was the #1 pick for wireless earbuds on all four engines. Power BI led business intelligence tools on all four, and HubSpot led marketing automation software. The low-agreement categories look like the answering service example: best translation service (28 brands, only Lionbridge on all four), best logo maker (24 brands, only Looka) and best database for startups (23 brands, only PostgreSQL).
The sources split far more than the brands
Two engines share about 44% of their brands on the same prompt and only 3% to 10% of the domains they cite. The same brand can reach two engines through completely different pages. Gemini and AI Mode share the most sources, 9.5%. ChatGPT and AI Mode share 4.4%, ChatGPT and Gemini 3.0%, and every pairing with Copilot sits at 3.2% or below.
AI Mode cites the most, 5.21 domains per answer, against 4.31 for Copilot, 2.88 for ChatGPT and 2.86 for Gemini. It cited Reddit in 50.4% of its answers and YouTube in 52.9%. ChatGPT cited neither in any answer, which is consistent with the collapse in ChatGPT’s Reddit citations since August. Gemini cited Reddit in 4.6% of answers and YouTube in 1.3%. Copilot cited Reddit in none and YouTube in 0.6%.
Copilot has its own favorite. It cited toolradar.com in 31.7% of its answers, while ChatGPT cited it in 3.7% and Gemini and AI Mode never did. Sources also move between runs on the same engine. A rerun kept 57.3% of AI Mode’s cited domains, 54.7% of ChatGPT’s, 49.8% of Copilot’s and 33.7% of Gemini’s.
Google’s AI Overview reads from yet another set. We had already mapped AI Overview citations for the same 160 buyer queries, and 158 of those queries joined cleanly with this study. AI Overview shares 25.5% of its cited domains with AI Mode, 8.9% with Gemini, 5.6% with ChatGPT and 2.9% with Copilot. Google’s two AI surfaces sit closest to each other and still draw most of their sources from different sites.
Copilot and Gemini answer with a shopping shelf
Copilot showed a shopping or product module in 57.5% of its answers (276 of 480). Gemini showed one in 40.8%, ChatGPT in 8.5% and AI Mode in 7.5%.
Copilot’s product carousel appeared on 79.2% of consumer answers and on 61.1% of B2B software answers. It also showed on 42.2% of services answers and 36.7% of developer tool answers. ChatGPT’s shopping cards stayed on consumer prompts: 34.2% of those answers had one, and no other segment did.
AI Mode showed sponsored ads in 77 of its 480 answers (16%). We dropped the advertisers from the brand counts, so every brand in this study is an organic pick. In those answers, a paid placement sits right next to the organic shortlist, and ads in Google AI Mode compete for the same attention as the brands that do not advertise.
What this means for brands
- Track each engine on its own. With 43.6% brand overlap between engines, an average across four engines describes none of them. Half of all recommended brands come from a single engine. Visibility rate and citation share per model show where the gaps are, and share of voice measured engine by engine shows who fills them.
- Treat one run as a sample. On Copilot and Gemini only about a third of brands appear in all three runs, so one answer can make a brand look present or absent by chance. ChatGPT’s cited sources shift from run to run as well. A free manual visibility audit holds up once you repeat each prompt several times and count how often you appear.
- Earn the sources each engine reads. AI Mode cites Reddit and YouTube in about half its answers, while ChatGPT cites neither. Reddit work for AI citations and YouTube videos built to be cited pay off mainly on Google AI Mode. Copilot leans on toolradar.com, so getting cited by Copilot starts with the pages it already reads.
How the study was run
The 160 prompts are plain buyer questions such as “best crm software” and “best running shoes”. They are the same set used in the AI Overview study. They cover 60 B2B software categories, 40 consumer products, 30 developer tools and 30 business services.
Each prompt went to ChatGPT, Microsoft Copilot, Google Gemini and Google AI Mode three times from the US in September 2026. All 1,920 answers were usable; ChatGPT returned 4 with no brand named and AI Mode returned 1. We pulled the recommended brands from each answer and grouped them by maker within a prompt, so Hoka Clifton 11 counts as Hoka. Overlap is the number of brands two answers share, divided by all the brands either one named. A second, independent extraction of 42 answers agreed with the first 94% of the time.
Download the dataset
The full dataset is free to use under CC BY 4.0. It has 11,889 rows, one per brand per answer, in JSON and CSV, and it is published on GitHub alongside our other AI search datasets. Filter it by prompt to see which engines name your brand, or by engine to see where a competitor shows up that you do not.