The AI reads your awards. It can't read your reviews.

On 17 August 2026 I read a text file, and it describes the next five years of employer brand more precisely than anything I heard said out loud this year. Glassdoor’s robots.txt asks the AI crawlers to stay out. The named agents include GPTBot, ClaudeBot, Claude-Web, anthropic-ai, Google-Extended, Applebot-Extended, Perplexity, Cohere, Amazonbot, GoogleOther and Google-CloudVertexBot, each disallowed in general terms; CCBot and Bytespider are shut out entirely. And then, for the ones merely disallowed, three doors are left explicitly open: /blog/, /About/ and /Award/.

The Best Places to Work registry sits under /Award/.

So the arrangement, stated flatly. The machines that are increasingly asked what it is like to work somewhere are permitted to read the awards and refused the reviews. The promotional surface is legible. The lived one is not.

The temptation is to call this censorship, and calling it that would waste the finding. Nothing editorial happened here. A review corpus is the asset a review company owns, letting a crawler copy it wholesale is giving the business away, and blocking the crawler is the ordinary defence of property. Awards pages are marketing, and marketing wants to be indexed. Not one person in that decision was ruling on which half of the record is true. The asymmetry is a side effect — which is exactly why it will persist. Side effects have no opposition.

The industry has already named this and is already selling the remedy. The trade press defines generative engine optimization for employer brand as “shaping content and reputation so that AI tools such as ChatGPT, Gemini, Perplexity, and Google’s AI Overviews accurately represent an organisation when answering employer-related queries”. It also asserts, without quantifying it, that earned and third-party sources outweigh brand-owned content in AI summaries, and that many queries now end without a click.

Hold those two claims next to the robots file. If third-party sources carry the most weight in a machine’s account of your workplace, and the largest third-party corpus in employer reputation is closed to the machine, then the weight lands where it still can: the awards page, the blog, the About page. Yours. All of them yours.

This is the part worth sitting with. The answer layer does not flatter employers because it likes them. It flatters them because the flattering documents are the ones it is allowed to read.

Meanwhile, the interior. Gallup’s 2025 figures have 12% of employees strongly agreeing that AI has transformed how work gets done in their organization, against 65% reporting that AI positively impacted their personal productivity. A workforce getting quietly faster inside organizations that have visibly not changed. That is a specific texture of working life right now, and it is precisely the kind of thing people put in reviews and almost nowhere a crawler can reach.

Deloitte’s 2026 Human Capital Trends is built as seven questions, and the second one is “How do we know what is true about people and work?” It is a fair question to find at board level, and this is one of the mechanisms that makes it hard to answer: the record is being split by access rather than by accuracy. In the same report, 7 in 10 business leaders say their primary competitive strategy over the next three years is to be fast and nimble — and the fastest available move on a machine-read reputation is an awards submission, not a better workplace.

The same report carries the number that should slow the purchase down. Those taking a tech-focused approach are 1.6x more likely to not realize returns on AI investments that exceed expectations, compared with those taking a human-centric approach.

And before anyone rebuilds a function around the answer layer, the counterweight. Randstad’s employer-brand research runs twenty-five years deep, more than 166,000 respondents, more than 6,400 companies, 34 markets. The top five drivers of employer choice, in order: salary and benefits, work-life balance, job security, work atmosphere, career progression. That list has the stubbornness of something true. Alongside it, 51% of workers consider a positive day-to-day work environment the top enabler of a healthy work-life balance.

Notice what that research does not say. It says nothing about AI and nothing about how candidates go looking. That is why it belongs here. It is the record of what people choose on, not of how they find out. Discovery moved. Choice did not. Before you buy generative engine optimization from the layer that sold you search engine optimization, establish which of the two has actually changed: the thing, or the invoice.

The limits, since they matter. A robots file is a request rather than a wall, it governs polite retrieval now, and it says nothing about what was ingested before the file said this, nothing about licensed feeds, and nothing about the many other places your workplace gets discussed. Glassdoor is one host. But it is the host whose name became the verb, and the direction of the asymmetry is the finding, not its magnitude. The surfaces you author are the surfaces most reliably legible to the machine.

Which leaves two responses. The first is arbitrage. The machine’s account of you is more generous than the human one, so feed the readable half: submit for more awards, publish more of the blog the crawler is welcome to, let the gap work for you. It fails on a mechanism too boring to argue with. The candidate reads both. The machine answer is the first impression, the reviews are the second, they are one click away, and they are still open to people. A first impression that the second impression contradicts is worth less than no first impression at all, because you have spent real credibility to be disbelieved on a specific point.

The second response is duller and it is the one that holds. Since the machine-legible surface is the one you write, write it as something checkable. Dated. Particular. Falsifiable against the reviews it is standing in front of. Not that you value your people, a sentence no employee has ever repeated back to anyone, but the specific arrangement a current employee would confirm without being asked to. The test is one question and it is cheap to run: could someone read your careers page, then your review page, and find the same company twice?

The distance between what a brand sells and what it is lived as is the oldest problem in this work, and none of this is new to it. What is new is that the sold side now has a machine reading it aloud, and the lived side has a text file in the way. That does not change the remedy. It changes the cost of not applying it.