82 per cent of applicants who use AI have changed their mind about an employer after a language model answered their question. Before they opened the job vacancy, before a recruiter was involved, before you knew they existed.
That figure comes from research by PerceptionX, May 2026, among job applicants in seven countries. The rest of the figures are equally solid.
96 per cent of jobseekers who use AI deploy it to research employers, understand roles or prepare for an interview. 72 per cent do so before deciding whether to apply at all. Your candidate is therefore forming their judgement during a stage when you think you have no influence whatsoever.
And then the figure that hurts: 58 per cent has caught a language model giving incorrect information about an employer.
The candidate decides on the basis of AI, and the AI gets it wrong more than half the time. About you.
Your employer story without you in it
A language model does not write its story based on your careers site. It reads what exists about you in the places the model visits: review sites, Reddit, old job listings, news articles, forums. Your employer brand is compiled there by a system that has never spoken to you. What is missing is filled in. What is unclear is simplified. And what is nowhere to be found, does not exist.
That is the heart of the matter. Most employers have a carefully crafted employer brand on their own channels, and a completely different story in the sources that the model actually reads. The candidate gets to see the latter, summarised as fact, with no source referenced for them to check.
Your newest story is your weakest source
What makes it extra galling is the timeline. A language model attaches importance to what is frequently repeated and what has been online for a long time. A review thread from two years ago therefore carries more weight than the campaign you launched last month. Your newest story is the model's weakest source, simply because it is young and comes from yourself. The candidate notices nothing of the sort. They receive a fluent, confident answer and assume it is correct.
This plays out across multiple models at the same time. ChatGPT stands at 89 per cent among these users, Gemini at 65, Claude at 39, Google AI Overviews at 27, Perplexity at 18. There is not a single place that covers you and leaves you done. It is the entire layer from which people now get their first impression.
What candidates question the model about
What are they asking about? Salary, 53 percent. Career progression, 52 percent. The job application experience itself, 51 percent. Culture, 41 percent. Job security, 34 percent. Those are precisely the topics employers are vaguest about on their own websites. The candidate asks the honest question; the model gives the smartest answer it can distill from the available mess.
It's getting louder, not quieter.
What is still a search query today will be a conversation in six months' time. What is still a conversation today will be an agent making the initial selection on behalf of the candidate in a year's time. 77 percent say they want to leave their search partly or entirely up to such an AI agent. 65 percent expect to lean more on AI for employer research within twelve months.
The reflex is to shout louder on your own channels. More content, more employer branding, more campaign. The model does not read that as truth; it reads it as the party's own marketing, and weights it accordingly.
What the model does take into account is whether your story is the same everywhere. Whether you, as an employer, are recognisable as a single entity, with one consistent story, in the places where the model looks. Whether your vacancies are structured in such a way that a language model can read and quote them. Whether what is written about you corresponds with what you say yourself.
Measure before you repair
It starts with measuring what the models are currently saying about you, in the questions candidates are actually asking. Only once you have that in black and white do you know whether you have a visibility problem or a reputation problem. Those are two different fixes, and most employers are guessing the wrong one.
That is the job. We call it GEO: generative engine optimisation. For employers, it means one thing: being citable in the answer a candidate gets before Google enters the picture.
The difference compared to classical labour market communication lies in the playing field. In the past, you competed for the click on your page. Now you compete for the sentence that the model speaks about you, in a window where your page does not even appear. Whoever determines that sentence determines whether the candidate presses send.
The employers who are tackling this now are building an advantage in precisely the system where the choice is made: the model's answer. The rest will find out in a year's time that they have been losing candidates for months to a summary they never even saw.
Your candidate has already asked the question. The only question that remains is whether your answer was there.
My name is Bryan Peereboom, founder of AI Rebels. We make employers visible in language models.