Employer branding and AI: what ChatGPT tells a candidate, and where it gets it from

21% of candidates use AI to research a company before applying (Insight Global, 2025), and the share grows with seniority and technical skills. The Careers page is still visited, but more and more often after a generative answer has already formed an opinion. This article tells what candidates find and what an HR team can do about it.

The question they ask

It is not "which positions are open". It is "what is it like to work at [Company]?", "is it a good place for a [role]?", "[Company] or [Competitor]: where should I go?". They are experience questions, and models treat them as such: they look for sources in which someone recounts the experience.

What the AI typically answers

In the employer assessments we run, the typical answer has three components: a description of the company (industry, size, headquarters), a judgement on working there ("some reviews mention...") and, if it exists, a relevant dated fact (a restructuring, redundancies, a change of ownership). What is almost always missing is everything the company has done in the last two years: benefits, training, career paths, awards, flexibility.

It is not missing because it does not exist: it is missing because it does not exist in a source the model reads and considers reliable.

Where it gets the information

The sources that weigh in answers to candidates are, in order of frequency:

1. Reviews: Glassdoor and Indeed above all, with one peculiarity: models summarise the whole body, and four-year-old reviews left unanswered weigh as much as recent ones.
2. News: an article from a news outlet on a restructuring or a lawsuit remains a source for years, if no more recent source updates it.
3. LinkedIn: company page, posts, profiles. Read above all by Copilot and used for size, locations and stated culture.
4. Wikipedia and Wikidata, when they exist, for basic facts and dates.
5. The Careers site, but only if it contains facts (numbers, programmes, dates) and not only claims.

What an HR team can do in the coming weeks

• Measure: ten questions as a candidate would ask them, on ChatGPT with web search and on Perplexity. The guide is in the DIY test.
• Reply to reviews, even old ones: a dated, concrete reply changes the way the model summarises the platform.
• Invite recent reviews from those who are satisfied, without forcing: the median year of reviews is one of the indicators that move sentiment most.
• Publish a "working with us: facts and figures" page: training hours per person, share of internal promotions, named benefit programmes, flexibility, with dates. These are facts a model can cite.
• Update Wikipedia and Wikidata where they exist, with the facts after the dated news item.
• Get someone to speak: an interview with the HR director in a trade publication, with data, is worth more than ten posts.

In life sciences

For pharmaceutical companies the topic has an extra dimension: the gap between talent needs and availability in the industry is estimated at 35% (AMS / CSG Talent, 2025), and the most contested profiles (researchers, medical advisors, MSLs) are also the most likely to research with AI. The pharma version of the Employer Assessment works within the regulatory perimeter.