Does AI always pick the same name?

We asked four AI models to recommend a bank, a hotel chain, a law firm, a CRM, a PR agency and an electric car, ten times each. In some categories one name wins every time. In others it depends on which assistant you ask.

This is a study of how AI assistants behave. It is not a ranking of companies and it is not advice. The names below are the assistants' own words, reported as measured. Being named often says nothing about quality.

What we found

1. In two categories, one name wins on every assistant. Asked to pick a single CRM for a small business, all four models chose HubSpot in all 40 answers. For a law firm for a startup, Cooley was the pick in 37 of 40 and was named in every one of the 40 open answers.
2. In three more, one name wins on three assistants out of four. For a hotel chain, Gemini, Claude and Perplexity chose Marriott in 29 of 30 answers, while ChatGPT chose Hyatt in 7 of 10. For a bank, ChatGPT, Gemini and Perplexity chose Capital One in all 30, while Claude leaned to Ally. For an electric car, three chose Hyundai most often and ChatGPT chose Tesla in 8 of 10.
3. PR agencies were the most open category. Each model had a favourite and repeated it, but the favourites differed: LaunchSquad on two models, Bospar on one, Treble PR on one. A model named about a dozen different agencies in those ten answers, counting the alternatives it mentioned beside its pick.
4. Stocks, where we started, were the least concentrated of all. On money questions some models hedge or decline to pick; Claude never singled out one stock. Our first study understated how narrow AI answers become elsewhere.
5. Which sources get cited changes with the category. For cars, about a third of the cited sources were news and another third were rankings and review sites. For PR agencies, law firms and CRM software, about half were business-owned websites, and news was 9%, 2% and 13%.
6. In business services, many cited pages are written by businesses in the field. Many of the most-cited pages were “best agencies” or “best firms” lists published by an agency, a firm or a vendor on its own site. We did not test whether a company's own pages are the reason it gets recommended.
7. An app can still answer differently from its model. Gemini's app gave the same single pick as the Gemini model in four categories and a different one in two: Highwire as its one PR agency in 10 of 10 answers, where the model chose LaunchSquad in 8 of 10, and Tesla as its one electric car in 6 of 10, where the model chose Hyundai in all 10. ChatGPT's app and model agreed on the name for both a bank and a PR agency, though the app was less consistent.

1. Forced to pick one, who comes first?

The question was “If you had to pick just one …, which one would it be?”. Each cell is the name a model put first most often, and in how many of its ten answers.

The bar shows how often each model repeated its own first pick, averaged over the four. It can be high while the models disagree with each other, so the last column counts how many of the four had the same first pick.

2. Asked openly, how often is one name in the answer?

The question was “What is a good … right now?”. An open question gets a list, so nearly every answer contains each assistant's leading name. What differs is how long the list is and how much of it goes to three names.

“Leads on 4 of 4” means the same name was the most-named on all four models.

3. Which sources the answers cite

Every website an answer cited, counted once per answer, by kind of site. Read across a row: it adds up to 100%.

“Not placed” is mostly other businesses' blogs and sites we could not identify with confidence. The kinds were assigned site by site with a model and spot-checked by hand; treat the shares as approximate. In the PR-agency row more than a quarter of the citations are not placed, so that row is the least certain.

4. The apps beside the models behind them

Most people meet these models through an app. We also asked the apps, logged out, the same questions: Gemini's app in all six categories and ChatGPT's in two. Each cell is the first pick when forced to choose one, and in how many answers.

An app is a product built around a model, with its own instructions and its own search, so the two can differ. A brand's standing in one says little about the other.

What this means for a brand

• Where one name wins, second place is nearly invisible. In a list of ten links a runner-up is still on the page. In a single-pick answer it is not there at all.
• Where assistants disagree, measure each one. A brand can be the first pick on one assistant and absent on another. An average across assistants hides both facts.
• In business services, look at what businesses publish, yours included. About half of the sources cited for PR agencies and law firms were pages on agencies', firms' and vendors' own sites, far more than press coverage. This study does not show that changing a page changes the answer; that takes a before-and-after test.

About this study

What we asked

Two questions per category, each framed “in the US”: an open one (“What is a good … right now?”) and a forced one (“If you had to pick just one …”). Each was asked ten times of each model on 6 October 2026.

How we counted

Names that are the same company count once: a car model under its maker, a hotel brand under its parent company, a bank's product under the bank. The stock row uses the questions of our first study; its ChatGPT answers were collected again for this report, so that row differs from the first report.

An exploratory study. Limits: one day, ten runs per question and model, US framing, four models asked directly. The apps were asked logged out, up to ten answers per question: Gemini's in all six categories, ChatGPT's in two. Shares by kind of website are approximate. Nothing here describes other countries or languages, and nothing here says which company is better.

Download the report as a PDF