CKR-001v1.0.0Data collected October 2026CC BY 4.0
Mentioned Isn't Recommended: How ChatGPT and Gemini Recommend HR Software
QueryWhich HR software do ChatGPT and Gemini recommend to a 50-person US startup?Summary
In a CobaltKite Research test of 60 answers from ChatGPT and Gemini, collected in October 2026, Deel was named in 68% of answers but placed in the top three recommendations only once. Rippling was ranked first in 87%. Being named by an AI assistant is not the same as being recommended.
Key findings
- Finding 01StrongDeel was named in 41 of 60 answers (68%) but placed in the top three recommendations in only 1 of 60.
- Finding 02StrongRippling was ranked first in 52 of 60 answers (87%) and appeared in the top three of every answer.
- Finding 03StrongFour vendors, Rippling, Gusto, BambooHR and Justworks, held 99% of all top-three recommendation slots.
- Finding 04ModerateWorkday was named in 13 of 60 answers and never recommended; every one of those mentions advised against it for a 50-person company.
- Finding 05ModerateHiBob was named in 20 of 30 ChatGPT answers but only 2 of 30 Gemini answers.
- Finding 06ModerateChatGPT named more HR vendors per answer than Gemini: 6.1 on average against 4.8.
- Finding 07TentativeWhen the buyer described their problem instead of asking for a list, BambooHR was ranked first in 5 of 20 answers, against 0 of 40 for the other prompts.
Strong: holds when any one prompt or assistant is left out. Moderate: clear numbers, but the cause is uncertain or rests on my reading of the answers. Tentative: driven mostly by one assistant or a few runs.
Named vs. recommended
Rippling, Gusto and BambooHR were named in every answer and usually recommended near the top. Deel, HiBob and Workday were named often but almost never placed in the top three.
Data table
| Category | Named in answer | In top three |
|---|---|---|
| Rippling | 100% | 100% |
| Gusto | 100% | 88% |
| BambooHR | 100% | 80% |
| Justworks | 75% | 28% |
| Deel | 68% | 2% |
| HiBob | 37% | 2% |
| Workday | 22% | 0% |
Source: CobaltKite Research 001, 60 answers from ChatGPT and Gemini, October 2026. CC BY 4.0.
What it means
For brands. If you measure AI visibility by counting mentions, you can look healthy while losing almost every recommendation. Deel shows up in two out of three of these answers. It almost never shows up where a buyer actually looks, which is the first three names.
The assistants are not negative about Deel. They file it under a condition: “if you hire internationally”. A buyer who doesn’t meet that condition moves past it. That is a positioning problem, not an awareness problem, and a mention count will never show it to you.
Workday makes the same point from the other side. It was named in 13 answers, and every one of them told a 50-person company not to buy it. In most AI visibility trackers, each of those still counts as a mention.
For anyone measuring AI visibility. Earlier this year, SparkToro and Gumshoe ran 2,961 prompts and found that AI assistants almost never return the same list of brands twice, let alone in the same order. Their advice was to stop tracking rank and track how often a brand appears across many runs instead.
This study is the next step. Mention rate is more stable than rank, but on its own it can still mislead. A useful measure looks at three things together: whether the brand was named, where it was placed, and what the answer said about it.
Method
I asked the same three questions to ChatGPT and Gemini, ten times each, for 60 answers in total.
| Prompt | Exact text |
|---|---|
| P1 · Quick search | best HR software for a 50 person startup in the US |
| P2 · Problem-led | We’re a 50 person SaaS startup in the US. Onboarding, time off and employee records are all in spreadsheets and it’s getting messy. What HR software should we switch to? |
| P3 · Shortlist | I’m evaluating HR platforms for our 50 person US SaaS startup. Which ones should be on my shortlist, and what is each one best for? |
- Assistants: ChatGPT (GPT-5.6 Sol, Thinking set to High) and Gemini (Gemini 3.1 Pro, extended thinking off).
- Setup: a fresh temporary chat for every run, personalization off, no follow-up messages, prompts rotated so no single prompt was bunched into one time of day. Web search was left at each assistant’s default.
- Location: runs were made from outside the US without a VPN.
- When: October 2026.
- Coding: I wrote down every vendor each answer put forward as an option, in the order it ranked them. Mentions anywhere in the answer were detected automatically from the full text.
- Statistics: rates use Wilson 95% confidence intervals. Differences use a bootstrap that resamples runs within each assistant and prompt. Every headline finding was recomputed with each prompt and each assistant left out.
Definitions: Named means the brand appears anywhere in the answer. Top three means it was one of the first three recommendations in the answer’s list.
The full method, column definitions and every check are in the README that ships with the data.
Limitations
- One category, one buyer, three prompts. These results describe HR software for a 50-person US SaaS startup. Other categories and phrasings may behave differently.
- Small samples. Ten runs per prompt per assistant. The intervals are wide, and small gaps between groups are noise.
- A snapshot. AI assistants change their models and answers constantly. Expect these numbers to drift. I plan to rerun the same protocol every quarter.
- Location. Runs were made from outside the US without a VPN, which can affect what an assistant searches and cites.
- Different settings per assistant. ChatGPT ran with Thinking on High and searched the web in all 30 runs. Gemini ran with extended thinking off and searched in 25 of 30. Differences between the two reflect these setups, not only the models.
- One coder. I coded the recommendation lists myself. Every raw answer is published so anyone can check.
- Two assistants. This version covers ChatGPT and Gemini. Claude is planned for version 2.0.
Data and code
Every answer, every coded list and the analysis script are public under CC BY 4.0. The permanent copy lives on Zenodo (DOI 10.5281/zenodo.23361807). Copies are also hosted here.
To cite every version of this dataset, including future quarterly updates, use the concept DOI 10.5281/zenodo.23361806.
- runs.csvAll 60 answers word for word, with settings and the coded recommendation list
- mentions.csvOne row per brand per answer: named, listed, position, top three
- prompts.csvThe three prompts, verbatim
- summary.csvEvery number on this page, with confidence intervals
- analysis.pyRebuilds mentions.csv and summary.csv from runs.csv
- README.mdFull method, column definitions and limitations
How to cite
Jaiprakash, T. D. (2026). Mentioned Isn't Recommended: How ChatGPT and Gemini Recommend HR Software (CobaltKite Research 001) (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23361807
Free to reuse with credit under CC BY 4.0. If you quote a number, please link this page.
Changelog
- v1.0.0 · 14 Oct 2026First release. ChatGPT and Gemini, 60 runs.
Want this run for your category?
I run the same test for your market: your buyers' questions, your competitors, every answer logged, and the raw data handed over.