Last week, ChatGPT hit 1 billion weekly users. In that same week, OpenAI launched a new model, GPT-5.6, which, according to its own measurements, makes 68 percent fewer factual errors than the version before it.
Add those two together. 1 billion people receive an answer every week that sounds more certain and misses the mark less often. Even when that answer is about you as an employer.
The comfort of the mistake disappears
That is bigger news for employer branding than it seems.
For two years, you could still comfort yourself with the mistakes. The model mixed up your branches, attached an old reorganisation to you, and invented an employment condition you never had. Annoying, but visibly nonsense, and therefore harmless. The candidate smelled the mistake and looked elsewhere.
That margin is now disappearing. A model that is 68 per cent less wrong sounds reliable across the board, and a reader who agrees with it on nine points will believe it on the tenth as well. Right there, on that tenth point, is often your story: the point that the model still gets wrong, but now packaged in a tone that leaves no room for doubt.
A model that thinks more deeply states its conclusion more assertively
The new version also reasons more deeply. Paid users get a slider that lets them make the model think for longer; free users get a button that does the same thing. The answer about you is therefore no longer a reflex. The model thinks for longer, compares sources, and states its conclusion more assertively.
And the model gets those sources from somewhere. ChatGPT relies on Wikipedia for almost half of its answers and cites an average of more than ten sources at a time. If you do not consistently appear among them, it builds its picture of you from what is actually there: an expired job posting, a review site, a forum thread from three years ago.
What is missing will be filled in. What is unclear will be simplified. And what is nowhere stated does not exist.
The story that you no longer write yourself
I spent fifteen years in recruitment. The whole time was about having a grip on the narrative: what the job description says, what the hiring manager says, what is on the careers site. That narrative had an owner, and that owner was you. Now a model reads everything about you that is floating around, weighs it longer than ever, and squeezes it into a single paragraph that sounds like it is right. That paragraph is the first impression of your next candidate, and you didn't write it.
Take your employment conditions. If they appear differently on your jobs page than on Glassdoor, and differently again in an old press release, the model used to guess one version. Now, it chooses with more conviction, and presents that choice as established fact. Where the candidate used to sense doubt, they now read a fact.
If your culture is only in an atmosphere video and nowhere in readable text, the crawler reads nothing, no matter how good the model has become. Even the best model only reads what is in words. Clarity beats volume, and that difference grows with every version.
It doesn't stop at your own site either. A model only truly believes you when others confirm you. A consistent mention in a trade publication, an employee explaining how things work on a forum, a review site saying the same thing as your page: that carries more weight than the finest sentence on your careers site. You no longer control the conversation alone; you feed it, or you let others feed it.
Visible or not visible as an employer
For a language model, there are only two types of employer left. The employer with a clear, consistent story that the model can retell, and which features in the answer. And the employer whose information cannot be found anywhere, and who is therefore second-guessed by the model for lack of guidance. That second-guessing is now happening with more confidence, and therefore with more damage.
The reflex is to wait. It won't come to that, surely the candidate will still turn up on your site. Some of them still do. But 1 billion people a week are already orientating themselves at the machine, and that number is growing every month. You can wait until a confident model states something definitively wrong about you to someone you never see passing by. Or you can decide today what it gets to read.
What you can do this week
What that asks for is less work than it sounds.
Put the answers to the real questions about salary, career progression and what an interview is really like into plain sentences on your own domain. Make sure your story says the same thing everywhere, so the model doesn't have to choose. And check what the models are saying about you today, literally, by asking them in ChatGPT, Gemini and Perplexity. There is a strong chance you won't see yourself reflected, even with a strong brand.
What is still an advantage today will be the standard in six months' time. What still feels optional today will be the baseline in a year. And what you haven't made readable for a model today will cost you the candidates you never even saw passing by.
1 billion people are asking AI. We make sure you are the answer.
My name is Bryan Peereboom, founder of AI Rebels. We make employers visible in language models.