Your AI employer brand is the description an AI assistant generates when a candidate asks what it’s like to work at your company. It is not your careers page and not your Glassdoor score. It is a paragraph assembled at query time from whatever sources the model can reach — and that stack is narrower and stranger than most employer brand teams assume.
We pulled the robots.txt files of the major employer-review destinations in July 2026 to find out which ones AI crawlers are permitted to read. The results reorder the standard playbook: the review corpus most teams spend their budget managing is largely walled off from the crawlers that build AI answers, while a few overlooked directories on those same domains are wide open.
This piece maps the source stack layer by layer, shows which layers a careers site can actually influence, and gives you an audit you can run this week.
What is an AI employer brand?
An AI employer brand is how ChatGPT, Gemini, Perplexity, Claude, Copilot and Google’s AI Overviews describe your company as a workplace when someone asks. It is generated fresh per query from retrieved sources plus model memory, so it varies by phrasing, by engine, and by week — unlike a static Glassdoor rating.
Two properties make it behave differently from a traditional employer brand asset:
- It has no fixed surface. There is no page to edit. The answer is composed at query time, so the only durable lever is what the engine can retrieve.
- It is comparative by default. Ask "should I take a job at X or Y" and the model builds a shortlist. Your company is scored against peers whether or not you chose to compete.
That second property is why employer brand now belongs in the same operating rhythm as buyer-side answer engine optimization. The mechanics that decide whether a vendor makes an AI shortlist decide whether an employer does. The employer-side question set specifically — "is it a good place to work?" — is broken down further in how AI answers employer-brand questions about your company.
What candidates actually ask AI before they apply
Candidate prompts cluster into six intents, and each pulls from a different part of the source stack. This is the prompt set we use when auditing an AI employer brand — worth running verbatim against your own company name.
| Intent | Representative prompt | Sources it usually pulls |
|---|---|---|
| Culture read | "What is it really like to work at [Company]?" | Editorial profiles, forums, careers content |
| Stability check | "Is [Company] financially stable? Any layoffs?" | News, funding coverage, layoff trackers |
| Compensation | "Does [Company] pay competitively for senior engineers?" | Salary aggregators, job listings, forums |
| Management | "What do employees say about leadership at [Company]?" | Review summaries, press, forum threads |
| Comparison | "[Company] vs [Competitor] — which is a better place to work?" | Everything above, merged and ranked |
| Red flags | "Are there any red flags about working at [Company]?" | News, litigation coverage, negative threads |
The last row is where most damage happens. Objection-shaped prompts trigger a retrieval pattern that actively seeks contrast — the same behavior documented on the buyer side in late-funnel objection queries. A model asked for red flags will find something to say. The only question is whether your own material is in the retrieval set to balance it.
Candidate adoption is not marginal. Career-services and recruiting research through 2025 consistently found majority AI usage among job seekers for company research, resume tailoring and interview prep — and rising year over year. Treat the answer as a first-round interview you are not present for.
The employer-reputation source stack AI can actually read
Access to the review corpus is not uniform, and it is not binary — it splits by crawler type.
Training crawlers versus live-fetch agents
AI engines send two distinct classes of bot, and sites treat them differently:
- Index and training crawlers — GPTBot, ClaudeBot, Google-Extended, CCBot, anthropic-ai. These build the corpora behind model weights and search indexes. Blocking them removes content from the general knowledge layer.
- Live-fetch agents — ChatGPT-User, OAI-SearchBot, Claude-User, PerplexityBot. These fetch a page in response to a specific user prompt, in real time.
A site can block one class and permit the other. That distinction explains contradictions employer brand teams report constantly: the model "knows" your rating but never quotes a review; or it cites a page one week and not the next. OpenAI documents the split between GPTBot, OAI-SearchBot and ChatGPT-User in its crawler reference; Google documents Google-Extended in its crawler list. Which index feeds which engine is its own tangle, mapped in which search index powers each AI engine.
What we found in the robots.txt files
We requested robots.txt directly from each domain in July 2026 and read the user-agent blocks. This is a snapshot of stated crawl policy, not a guarantee of crawler behavior — but stated policy is the ceiling on what a compliant crawler retrieves. Re-check before acting on it; these files change without notice.
| Source | Training crawlers | Live-fetch agents | Practical read |
|---|---|---|---|
| Glassdoor | Disallow: / for GPTBot, Google-Extended, anthropic-ai, Claude-Web, ClaudeBot, Perplexity — with Allow: carve-outs for /blog/, /Award/, /About/. CCBot fully blocked |
ChatGPT-User, OAI-SearchBot and PerplexityBot are not named, so they inherit the wildcard rules | Review pages closed to training crawlers; three directories deliberately left open |
| Indeed | GPTBot, ClaudeBot, anthropic-ai and CCBot disallowed from /jobs, /career/, /companies/, /cmp/ |
ChatGPT-User, OAI-SearchBot, Claude-User, PerplexityBot and Google-Extended permitted under Allow: / with selective path blocks |
Company profiles blocked from training, reachable on live fetch |
| Comparably | Only CCBot carries Disallow: /; no other AI crawler named |
Not restricted | Open — the most AI-readable major review source we checked |
| Blind (teamblind.com) | robots.txt returned HTTP 403 to our fetcher | Server-level refusal, not a robots directive | Effectively unreadable to compliant crawlers |
| robots.txt not retrievable by our fetcher; Reddit has publicly announced content-licensing deals with Google and OpenAI | Governed by those agreements rather than open crawl | Reachable in some engines by contract, not by crawling |
The headline: the two review sites with the highest candidate mindshare are the two least available to training crawlers, and the least-known one is wide open. If your AI employer brand reads oddly favorable or oddly thin, this table is usually why.
Two corollaries worth acting on:
- Comparably is underweighted. A complete, current Comparably profile is one of the few review-site assets a training crawler can actually ingest. Most teams treat it as a tertiary platform.
- Blind’s influence is indirect. Blind threads rarely enter answers as citations, but they get quoted in press and recruiting blogs — which are open. The thread reaches the model through journalism, not through Blind.
The Glassdoor award loophole
Look again at Glassdoor’s allowlist: /blog/, /Award/, /About/.
Reviews are closed. Award pages are open. That is a specific carve-out with a direct consequence:
- A Best Places to Work listing is a crawlable, structured, third-party endorsement that AI training crawlers are explicitly permitted to read.
- Ten thousand words of review text on the same domain are not.
- Coverage of your company in Glassdoor’s own editorial blog is likewise readable.
Most employer brand budgets are allocated in the exact inverse of that access map — heavy on review response and rating management, light on award submissions and earned editorial. Reallocating even modestly toward the crawlable side costs less than a review-response program and reaches a surface the review program cannot.
Why your Glassdoor rating still appears when reviews are blocked
Because the number travels without the page. A blocked review corpus does not stop a rating from being repeated — it stops the reasoning behind the rating from being retrievable.
Ratings leak into AI answers through at least four crawlable paths:
- Press coverage. Business journalism routinely quotes "a 3.4 on Glassdoor" in a paragraph on a fully open news domain.
- Recruiting content. Agency blogs, listicles and salary guides restate ratings freely.
- Your own materials. Careers pages that display a rating badge publish the number themselves.
- Model memory. Content crawled before a robots.txt change persists in weights.
The result is a detail asymmetry: the model states your score confidently but cannot substantiate it, so it fills the gap with whatever narrative material is retrievable — usually news. That is how one layoff article ends up doing the explanatory work a thousand balanced reviews should be doing.
When a number in an AI answer has no traceable page behind it, treat it as a provenance problem rather than a sentiment problem. The method for tracing a claim back to its source is in tracing brand misinformation in AI answers.
How layoffs and bad news enter the answer
News is the most retrievable layer in the stack — no paywall on headlines, high domain authority, dense internal linking, and syndication that multiplies one story across dozens of readable URLs.
That creates a structural imbalance in every AI employer brand:
- Negative events are documented in the most crawlable format available. A layoff generates coverage across news wires, trade press and aggregator trackers within hours.
- Positive equivalents are not. Promotion rates, internal mobility and retention improvements generate no external coverage at all unless you publish them.
- Recency weighting compounds it. Retrieval favors fresh documents, so a two-year-old restructuring can still dominate an answer if nothing newer exists on the same topic.
The counter is not suppression, which does not work and reads as evasion when the model surfaces the original story anyway. The counter is volume and recency on the same subject from your own domain: a dated post explaining what changed after the event, the current headcount trajectory, and what the team looks like now. That gives retrieval something contemporaneous to weigh.
What the post-event page needs, minimally:
- A visible date in the page copy, not only in metadata.
- The event named plainly. A page that avoids the word "layoff" will not be retrieved for layoff prompts.
- A specific after-state. Current headcount, teams rebuilt, roles reopened — numbers a model can lift.
- No argument with the coverage. Contradicting a news source gets you a "reports conflict" hedge, not a correction.
Multi-entity confusion: when the model answers about a different company
Underrated failure mode. Before you fix sentiment, confirm the model is describing your company.
Three cases produce a wrong-company answer:
- Name collisions. A common noun or shared brand name pulls in a same-named firm in another sector. Symptom: accurate-sounding details you don’t recognize.
- Post-acquisition drift. After an acquisition or rebrand, answers merge the old and new entities, often citing the acquired company’s pre-deal reviews as current.
- Script and transliteration gaps. For companies operating across languages, the localized name and the English name can resolve to different entities entirely — a mechanism covered in brand name transliteration and AI confusion.
The fix is entity disambiguation, not content: consistent legal name, sameAs links to authoritative profiles, and one canonical about page. Organization schema is the primary tool — with real limits on what it can assert, worked through in Organization schema for AI search.
What a careers site can actually change — and what it can’t
Be honest about use. Some interventions move the answer; most do not.
| Intervention | use | Why |
|---|---|---|
| Publishing specific, dated facts (headcount, remote policy, review cycle, salary bands) | High | Unique retrievable claims with no competing source |
JobPosting and Organization structured data |
High | Machine-readable entity facts; per Google’s JobPosting structured data documentation, also feeds Google’s jobs experience |
| Named employee content with author bios | High | First-hand experience signals; attributable to a person |
| Award submissions and third-party editorial | High | Crawlable endorsements on domains that block reviews |
| A post-event explainer page after layoffs or restructuring | Medium | Competes on recency with news coverage |
| Complete profiles on open review platforms | Medium | Readable where the closed platforms are not |
| Responding to individual reviews | Low | Response text sits behind the same crawler blocks as the review |
| Rating badges on the careers page | Low | Republishes a number you cannot substantiate |
| Video-first culture pages with no transcript | Very low | Nothing to retrieve |
| Careers content behind an ATS subdomain with thin HTML | Very low | Frequently unrendered or unindexed |
The pattern is consistent: retrievable, specific, attributable claims win; sentiment management loses. A sentence reading "engineering runs a two-week cycle with no on-call for the first six months" is worth more to your AI employer brand than a page of adjectives, because it is the kind of claim a model can lift and attribute.
One technical check most teams skip: if your careers content lives on a JavaScript-heavy ATS subdomain, verify what a live-fetch agent actually receives. Browsing agents read the rendered page, and what they see often differs from what your CMS previews — a gap detailed in how AI browsers read your site live. Fetch the page with JavaScript disabled; whatever is missing is missing for retrieval too.
How to run an AI employer brand audit in one afternoon
Run this before changing anything. It takes roughly three hours and produces a baseline you can defend in a budget conversation.
- Pick six prompts from the intent table above, using your exact legal and common company names.
- Run each prompt on four engines — ChatGPT, Gemini, Perplexity and Claude — in a logged-out or temporary session so personalization does not contaminate results.
- Screenshot every answer. Capture the citation list, not just the prose. The cited URLs are the actual intervention targets.
- Log every claim in a sheet: claim, sentiment, engine, cited source, and whether the source is accurate.
- Classify each cited domain as owned, earned, review-site or forum. This is your real source stack, and it will differ from the one you assumed.
- Flag every unsourced claim. Any assertion with no citation is model memory — the hardest category to correct and the one to prioritize.
- Repeat the comparison prompt against your two closest talent competitors and record who appears first and why.
- Re-run in 30 days. One snapshot tells you the state; two tell you the direction.
Score each engine on a 0–5 rubric so results stay comparable over time:
| Score | Condition |
|---|---|
| 5 | Accurate, specific, cites your owned content |
| 4 | Accurate and specific, cites third parties only |
| 3 | Accurate but generic — no differentiating detail |
| 2 | Mostly accurate with one material error |
| 1 | Dominated by a single negative event |
| 0 | Wrong company, or refuses to answer |
Most companies running this for the first time land at 3: accurate, generic, indistinguishable from competitors. That is not a reputation crisis. It is an absence of retrievable specifics — the cheapest problem on this list to fix.
A 90-day sequence for improving your AI employer brand
Ordered by use per unit of effort, drawn from the access map above.
Days 1–30 — make yourself retrievable.
Publish three specific, dated careers pages: how hiring actually works end to end, what the first 90 days look like by function, and current compensation philosophy with real bands. Add JobPosting markup to every open role and Organization markup to the careers hub. Confirm your ATS subdomain serves crawlable HTML.
Days 31–60 — build crawlable third-party proof.
Submit to award programs whose result pages are crawler-accessible — Glassdoor’s /Award/ directory being the clearest example. Pitch two employee-authored pieces to trade publications in your sector. Complete every profile on open review platforms, since those are readable where the closed ones are not.
Days 61–90 — close the narrative gaps the audit found.
For each material negative event surfaced in step 6, publish one dated, factual update on your own domain. Do not argue with the coverage; supply what happened next. Then re-run the full prompt set and compare rubric scores.
Track the delta on three measures: citation share (how often your owned domains appear in the citation list), claim accuracy (share of factual assertions that are correct), and comparative placement (whether you appear first, second or not at all in head-to-head prompts).
Expect lag. Live-fetch surfaces can reflect a new page within days of indexing; model-memory claims persist until the next training refresh, which no publishing schedule can accelerate. Judge the 90 days on citation share, not on whether an old claim disappeared.
How to measure it continuously without checking by hand
Manual audits establish a baseline; they cannot detect drift. AI answers change when a source is republished, an index refreshes or a model updates — none of which produce a notification.
What continuous ai search monitoring needs to capture for employer-side prompts:
- Answer text over time, so a sentiment shift is visible the day it happens rather than at the next quarterly review.
- Citation-level tracking, because the cited URL list is the only reliable map of where to intervene.
- Cross-engine coverage, since engines weight structured owned content and distributed external mentions differently — the same company can score 4 on one and 2 on another.
- Comparative placement, tracking share of voice against named talent competitors on shortlist-shaped prompts.
This is the same llm brand tracking infrastructure marketing teams use for buyer-side visibility, pointed at a different prompt set — which is also the argument for running both from one system rather than buying a separate employer-brand tool. If you are evaluating platforms, best tools to track brand visibility in AI search covers what each one actually measures, and the 12 best AI brand monitoring tools covers the wider category.
Frequently asked questions
Can I stop AI from mentioning negative reviews about my company?
No, and attempting it usually backfires. Blocking crawlers on your own domain removes your side of the story while leaving news coverage and third-party content fully readable. The workable response is publishing specific, dated, retrievable material that gives the model something contemporaneous to weigh against older negative coverage.
Does responding to Glassdoor reviews improve my AI employer brand?
Barely, in retrieval terms. Glassdoor’s robots.txt disallows the main AI training crawlers from review pages, so your responses sit behind the same wall as the reviews. Responding still matters for humans who visit Glassdoor directly — it is simply not an AI visibility lever and should not be funded as one.
Why do different AI engines describe my company so differently?
Because they retrieve from different indexes and weight source types differently. One engine may favor structured content from your own domain while another composes from distributed third-party mentions. Auditing a single engine gives you roughly a quarter of the picture; run all four and score them separately.
How often should I check what AI says about working at my company?
Monthly for the core prompt set, and immediately after any funding round, layoff, acquisition or executive change. Those events generate dense, highly crawlable news coverage that can reshape the answer within days.
Is this the same as SEO for the careers page?
Overlapping but distinct. Traditional search ranks pages, so a strong careers page can win a position. AI search composes an answer from many sources, so your careers page is one input among news, forums and review sites. Generative engine optimization for employer brand means managing the whole retrievable source stack, not one URL.
How long does it take to change what AI says about us?
Live-fetch surfaces like Perplexity and ChatGPT web search can pick up a new page within days of it being indexed. Claims held in model memory only shift at the next training refresh, on the provider’s schedule. Plan in quarters, and measure citation share rather than waiting for a specific sentence to disappear.
Should small companies with few reviews worry about this?
More, not less. With a thin source stack, a single Reddit thread or one news article can carry the entire answer, and models are more likely to hedge or confuse you with a same-named company. Publishing a handful of specific, dated careers pages moves the answer far more for a 50-person firm than for a 50,000-person one.
The short version
Your AI employer brand is already being generated, several times a day, by people deciding whether to apply. The audit above tells you what it currently says. The access map tells you where intervention is possible: specific, dated, retrievable claims on your own domain, plus third-party proof on surfaces AI crawlers are permitted to read.
The review sites you have been optimizing for a decade are, for the largest AI engines, mostly dark. The award page next to them is not. Spend accordingly.