On 14 August, Google swapped the model behind its AI Mode. No press release, no blog. A single post on X by Robby Stein, VP of Product at Google Search, and that was it.
The new engine is called Gemini 3.7 Flash. It is the third time in nine months that Google has swapped the brain behind its search answers: Gemini 3 Pro in November, Gemini 3.5 Flash-Lite in July, and now this.
To you as an employer, that sounds like technology far removed from your reality. It is not.
What AI Mode does when a candidate searches your sector
The model retrieves sources, weighs them, and kneads them into a single answer containing two or three names. You hope that one of those names is yours.
Now for the uncomfortable part. Swap the model, and the answer changes, even if the sources stay exact. The same pages, the same reviews, the same job listings. A different model reads them and summarises them differently. No ranking was overturned and there was no core update. The only thing that changed was the model that kneads the sources into an answer.
Flash vs. Pro: why the model determines who gets mentioned
The reason the answer changes lies within the model itself. A Flash model is built for speed, a Pro model for depth. They weigh sources differently, they summarise more briefly or at length, and on one occasion they dare to mention a name while on another they play it safe. The same facts, a different judgement.
That means something simple and harsh. The employer that was mentioned last month may drop away this month. Without anything having changed on their site. Without them having done anything wrong. Purely because the ground beneath them has shifted.
And you don't hear it. For now, this version is opt-in for paying users and in English only. It is the pattern that matters: the layer that determines who gets mentioned changes every few months, and usually you only notice when you pay attention to it. Google’s own help page doesn't even mention which model powers the default. The layer your visibility depends on is a black box that changes shape without warning.
That same week, Google fully rolled out its third spam update of 2026. The entire search landscape is continuously shifting, both above and beneath the bonnet.
Here lies the difference with how most employers view visibility. They treat it as a task with an end date: the careers site sorted once, the texts rewritten once, ticked off. The model that changes every few months pulls the rug out from under that approach. What you made quotable in May can turn out differently under a new model in August.
Measuring as a habit, not a one-off action
We see that movement in the work. In continuous monitoring, I ask the same questions every week about multiple AI platforms, and the list of mentioned employers shifts, sometimes within a single month. The names move because the underlying models move, while the employers themselves change nothing on their site. Anyone who only measures in the spring thinks they are visible, and does not know that the answer in the autumn names someone else.
For recruitment, that shifts where the battle lies. The candidate arrives with a perception that a model has formed, and that model may already be different since your last measurement. You are managing based on last month's reality, whereas the candidate is living in today's.
What is the answer today might be a footnote after the next model change. What is a footnote today could be at the front after the change after that. And what you set up once and then leave alone, you will lose in three months without anyone warning you.
So you measure it as a trend. Every month the same questions your candidates ask, on every platform they use, captured in a sequence that you can lay side by side. Then you see a drop before it hits your influx, and you know whether it's down to you or a model that has been swapped out.
Measurement tells you where you stand; climbing is the work afterwards. If you see your name dropping after a model change, you go back to the sources: which page did the old model pick up that the new one didn't, and why. That is manual labour, not a button.
The foundation underneath changes less quickly than the model on top. Technology that the crawler can access, structure that any model can summarise, facts that remain the same everywhere: that stands firm, whichever model Google puts underneath it. A strong foundation beats an accidentally favourable model version, every single time.
Google keeps swapping its model, more often than it tells you. The employers that survive that have one habit in common: they keep measuring while the ground beneath the answer moves.
My name is Bryan Peereboom, founder of AI Rebels. We make employers visible in language models.