Glassdoor Get Company Reviews

Purpose

Given any reference to a company on Glassdoor — full reviews URL, EmployerId (E1234567), bare company name, or name + location disambiguator — return the company's overall rating, sub-rating averages, rating distribution, recommend / business-outlook / CEO-approval percentages, plus a filtered slice of employee reviews (with per-review pros / cons / advice / sub-ratings / employment status / permalink). The skill exposes Glassdoor's full review-page filter surface: star bucket(s), job title, location, employment status, language, employment type (FT/PT/contractor/intern/freelance), sort order (most recent / most helpful / highest / lowest), free-text keyword, and a limit for pagination (Glassdoor paginates by 10/page). Read-only — the skill never clicks "Write a Review", "Add Salary", "Add Interview", "Follow Company", or any other mutation control, and never submits any form.

When to Use

• An agent / workflow needs a structured snapshot of a single company's reviews on Glassdoor (e.g. weekly employer-brand monitor, candidate research, market intelligence).
• A user pastes a Glassdoor reviews URL and asks "summarize the negative reviews from the last 3 months in NYC, full-time only".
• A recruiter wants the top-5 most-helpful current-employee reviews for a specific job title at a target company.
• Any flow that would otherwise scrape Glassdoor HTML — this skill bakes in the auth / Verified / anti-bot detail those flows always get wrong on the first try.

Workflow

The Glassdoor reviews surface has no usable public API today. All four of the obvious shortcuts were probed and confirmed dead during skill generation (2026-05-16):

1. Partner API (api.glassdoor.com) → HTTP 410 Gone, zero-byte body. Officially retired to new partners in 2021; live callers now get 410. Do not attempt.
2. Internal GraphQL gateway (/graph) → HTTP 403, zero-byte body (a real server-side Forbidden, not the Cloudflare interstitial). The endpoint exists but rejects every request that doesn't carry a fresh anti-CSRF + session cookie set issued inside a logged-in browser context.
3. /autocomplete/suggest (name → EmployerId resolver) → HTTP 403 with the same Cloudflare "Security | Glassdoor" challenge body as the reviews HTML page. So even cheap name-to-EmployerId resolution must happen inside the authed session.
4. Bare browse cloud fetch on /Reviews/<Company>-Reviews-E<id>.htm → HTTP 403 Cloudflare challenge page (~241 KB, <title>Security | Glassdoor</title>, __cf_chl script tags embedded). Confirmed with --proxies on and off; the Fetch API path is permanently walled for this domain because no real Chrome runtime is executing the challenge.

The only working path is scripted browsing through Browserbase with --verified --proxies AND a pre-authed Glassdoor session re-attached via a Browserbase Context ("cookie-sync"). Unauthenticated sessions hit Glassdoor's "give-to-get" wall after the first review — the page renders the first review fully, then overlays a modal demanding the visitor either log in or contribute a review/salary/interview before continuing. There is no URL-param bypass for the give-to-get wall.

1. Operator-side one-time login → Browserbase Context

Done by the human operator running this skill, not by the agent. Performed once per Glassdoor account; the resulting CTX_ID is the durable artifact the skill consumes.

# Create a persistent context (cookies + localStorage + IndexedDB)
CTX_ID=$(browse cloud contexts create | jq -r '.id')

# Spin up a one-off session attached to that context, with Verified + proxies
SID=$(browse cloud sessions create \
        --context "$CTX_ID" \
        --verified --proxies --keep-alive \
        --region us-east-1 | jq -r '.id')

# Hand the human operator the live debug URL — they log in normally
browse cloud sessions debug "$SID" | jq -r '.debuggerFullscreenUrl'

# Operator: log in at https://www.glassdoor.com/profile/login_input.htm,
#           satisfy any CAPTCHA / email verification, dismiss the
#           "give-to-get" modal by completing the minimum contribution
#           if the account is brand-new.

# Release the session — cookies are persisted into $CTX_ID
browse cloud sessions update "$SID" --status REQUEST_RELEASE

Cookies live ~30 days. When a skill invocation lands on /profile/login_input.htm instead of the requested reviews page, the operator must re-do this step.

2. Skill invocation — per request

SID=$(browse cloud sessions create \
        --context "$CTX_ID" \
        --verified --proxies --keep-alive \
        --region us-east-1 | jq -r '.id')

--verified and --proxies are both mandatory on top of the context. Bare context + no Verified still gets Cloudflare-flagged because the upstream IP pool used by Browserbase's default egress is fingerprinted; the residential proxy + Verified shim is what clears the JS challenge. Region is operator-pinned (default us-east-1 for US companies, eu-central-1 for European companies — Glassdoor geo-renders some content).

3. Resolve the company reference → canonical reviews URL

The skill accepts four input shapes; pick the cheapest resolution path that's available.

Persist name → EmployerId mappings to a sidecar cache so each company resolves exactly once.

4. Apply filters via URL query string

Glassdoor's filter rail is a thin client over a documented set of query-string params. Set them on the canonical URL directly — opening the URL with the params attached re-renders the filtered list server-side. Do not click the rail buttons unless a filter is not URL-expressible (rare); they trigger client-side state that does not survive page reloads.

Canonical URL skeleton:

https://www.glassdoor.com/Reviews/<Slug>-Reviews-E<EmployerId>.htm
  ?filter.iso3Language=<eng|spa|fre|...>
  &filter.employmentStatus=<REGULAR|PART_TIME|CONTRACT|INTERN|FREELANCE>
  &filter.currentJob=<true|false>            (current-only vs former-only)
  &filter.defaultEmploymentStatuses=true     (both current AND former — default)
  &filter.jobTitleFTS=<urlenc-job-title>     (auto-completed list)
  &filter.countryId=<int>                    (autocomplete-derived)
  &filter.stateId=<int>
  &filter.cityId=<int>
  &filter.metroId=<int>
  &filter.locationId=<int>                   (office-level granularity)
  &filter.ratings=<1|2|3|4|5>                (repeat the param for multi-select)
  &filter.searchKeyword=<urlenc-free-text>   (note: robots-disallowed — see gotchas)
  &sort.sortBy=<DATE|HELPFULNESS|RATING_DESC|RATING_ASC>
  &sort.ascending=<true|false>

Pagination uses a path suffix, not a query param:

.../Reviews/<Slug>-Reviews-E<EmployerId>_P<page>.htm

Page 1 has no suffix; pages 2+ append _P2, _P3, etc. _P* URLs are robots-disallowed — you must be authenticated, and you must throttle (~1 request / 2 s sustained).

Discovery procedure for filter IDs (jobTitleFTS, locationId, etc.):

1. Open the reviews page with no filters.
2. browse cloud browse snapshot and find the filter-rail combobox refs.
3. browse cloud browse click the combobox, then browse cloud browse type <partial> to surface the autocomplete dropdown.
4. Each dropdown entry's data-id attribute is the integer ID; read it via browse cloud browse get html body and a regex over data-job-title-id="\d+" / data-location-id="\d+".
5. Cache the resolved IDs alongside the EmployerId.

5. Scrape rendered reviews

browse cloud browse --connect "$SID" open "$URL"
browse cloud browse --connect "$SID" wait load
browse cloud browse --connect "$SID" wait timeout 2500     # async sub-rating widget
browse cloud browse --connect "$SID" get html body > /tmp/page.html

Per-review block markers in the rendered HTML (stable across 2026 re-skins as of skill generation, but treat as breakable — verify each run):

Permalink: https://www.glassdoor.com/Reviews/Employee-Review-<Slug>-RVW<review_id>.htm.

6. Scrape the company header (overall + distribution + CEO)

The reviews page renders the company aggregates in the header. Locators:

7. Page through results

for page in 1 2 3 ... ; do
  URL_P="${URL_BASE}${page > 1 ? "_P${page}.htm" : ".htm"}${QS}"
  browse cloud browse --connect "$SID" open "$URL_P"
  browse cloud browse --connect "$SID" wait load
  browse cloud browse --connect "$SID" wait timeout 2000
  # extract; concat
  sleep 2    # robots-disallowed path → conservative throttle
done

Stop when one of:

• The requested limit is reached.
• The page renders the "End of reviews" footer (<div data-test="reviews-list-end">).
• The URL after navigation is /profile/login_input.htm (cookies expired — bail out, prompt operator to re-auth).

8. Release the session

browse cloud sessions update "$SID" --status REQUEST_RELEASE

Site-Specific Gotchas

• READ-ONLY. Never click "Write a Review", "Add Salary", "Add Interview", "Apply on Glassdoor", "Follow Company", the helpful 👍 button, or any time-slot / submit / order control. The skill's sole job is to extract; any mutation breaks the contract.
• Cloudflare wall on bare HTTP. browse cloud fetch to any glassdoor.com path returns HTTP 403 with <title>Security | Glassdoor</title> and a __cf_chl interstitial in the body (~241 KB). Confirmed with --proxies on and off, 2026-05-16. The Fetch API surface is permanently unusable for this domain. Real Chrome + Verified + residential proxy is the only path.
• Partner API is gone. https://api.glassdoor.com/api/api.htm returns HTTP 410 with a zero-byte body. The Glassdoor Partner API was deprecated to new partners in 2021 and now actively 410s. Do not waste turns on it.
• /graph is a trap. Glassdoor's internal GraphQL gateway at https://www.glassdoor.com/graph returns HTTP 403 with an empty body (server-side Forbidden, distinct from the Cloudflare interstitial). It requires a fresh anti-CSRF token + login-session cookies issued from the rendered page, which the page rotates per-tab. You can't reuse them out-of-band. Stay in the rendered DOM.
• /autocomplete/suggest is also walled. The JSON autocomplete endpoint that resolves company name → EmployerId returns the same Cloudflare 403 as the HTML pages. Do not assume it's CORS-public; it isn't.
• Give-to-get login wall. Unauthenticated sessions render the first ~1 review fully, then overlay a modal: "To continue reading, contribute a review / salary / interview, or log in." There is no URL-param bypass. The wall is keyed to the visitor's account state, not the IP — so even a fresh Verified session triggers it. Solution: pre-authed Browserbase Context (see Workflow step 1). Brand-new accounts must additionally satisfy the give-to-get minimum (one contribution) before being granted full read access; do this in the operator's one-time login step, not the agent's per-request path.
• filter.searchKeyword= is robots-disallowed. Glassdoor's robots.txt explicitly disallows *filter.searchKeyword=*. Authenticated sessions can still use it, but throttle aggressively (≤ 1 request / 2 s) and expect occasional 429s. If the keyword filter starts returning empty results when the un-keyworded query returns reviews, you've been throttled — back off 60 s.
• Pagination is robots-disallowed. /Reviews/*_P*.htm and /Reviews/*_IP*.htm are both Disallow: in robots.txt. Only the SEO exception /Reviews/*-reviews-SRCH_*_IP2.htm* is Allow:. In practice an authed session can paginate normally, but if you start to see Cloudflare-served 403s mid-paginate, you've crossed the throttle threshold — add sleep 2 between page fetches, never less.
• Filter rail state does not survive a reload. Clicking buttons in the rail mutates a client-side store but not the URL; if the agent reloads or follows pagination after rail-only interaction, the filters are lost. Always express filters as URL query params (filter.ratings=4, sort.sortBy=DATE, etc.) — do not click rail buttons unless capturing autocomplete IDs.
• Multi-select stars uses repeated filter.ratings= params. To filter to 4 + 5 star reviews: &filter.ratings=4&filter.ratings=5. Comma-joining is silently dropped to a no-op filter.
• filter.currentJob is tri-state, not boolean. true = current employees only, false = former employees only, omit the param entirely to get both. Sending filter.currentJob=both 400s.
• EmploymentStatus enum: REGULAR | PART_TIME | CONTRACT | INTERN | FREELANCE. Surface depends on the company's review distribution — Glassdoor hides options with < 5 reviews in that bucket, so the rail may not show every value. URL-set the param directly; an unsupported value renders zero reviews (it's a filter, not an error).
• Sort param is split: sort.sortBy accepts DATE | HELPFULNESS | RATING_DESC | RATING_ASC, plus a separate sort.ascending=<true|false> that is honored only when sortBy=DATE (toggles chronological direction). The "most helpful" sort silently re-ranks per-page, so two requests for the same page can return different ordering — capture the review ID set per page and de-duplicate when assembling the final output.
• datetime attribute is canonical, rendered relative text is not. Always read <time datetime="...">; the human-readable "5 days ago" string is locale-dependent and rounds.
• CEO approval count is in aria-label, not text. Body text says "78%"; the <div aria-label="78% approve · 1,234 ratings"> carries the count. Don't drop the aria-label.
• Sub-rating widget renders async after load. Wait ≥ 2 s after wait load before snapshotting, otherwise the sub-ratings come back as zeros.
• Locale routes affect content. /Reviews/... is the US/global English route. UK reviewers may post on glassdoor.co.uk (same EmployerId, partial overlap of reviews depending on the company). For an exhaustive cross-locale extract, the skill must iterate the glassdoor.{com,co.uk,de,fr,nl,com.au,ca,ie,com.mx,com.br,it,com.hk,jp} TLD set and de-dupe by review_id.
• The skill ships as candidate. The cookie-sync requirement is operator-supplied and was not exercised end-to-end during skill generation: the host sandbox blocked outbound CDP egress to connect.*.browserbase.com, so all four anti-bot walls above were probed via the REST/Fetch API only. The authed-context workflow is sound on paper and matches the documented Glassdoor architecture, but production validation requires (a) a real operator login, (b) one end-to-end run on at least one company per locale-cluster, and (c) tightening the per-review locator table against any 2026+ DOM re-skin.

Expected Output

The skill returns one of three top-level shapes.

Shape 1 — Success (company + filtered reviews)

{
  "success": true,
  "company": {
    "name": "Stripe",
    "employerId": 671932,
    "canonicalUrl": "https://www.glassdoor.com/Reviews/Stripe-Reviews-E671932.htm",
    "logoUrl": "https://media.glassdoor.com/sql/671932/stripe-squarelogo.png",
    "headquarters": "South San Francisco, CA",
    "industry": "Internet & Web Services",
    "sizeBucket": "1001 to 5000 Employees",
    "foundedYear": 2010,
    "website": "https://stripe.com",
    "ceo": {
      "name": "Patrick Collison",
      "approvalPercentage": 94,
      "approvalCount": 1234
    },
    "overallRating": 4.1,
    "totalReviewCount": 612,
    "recommendToFriendPercentage": 78,
    "businessOutlook": { "positive": 71, "neutral": 17, "negative": 12 },
    "subRatingAverages": {
      "careerOpportunities": 4.0,
      "compAndBenefits": 4.3,
      "cultureAndValues": 4.1,
      "seniorManagement": 3.7,
      "workLifeBalance": 3.5,
      "diversityAndInclusion": 4.0
    },
    "ratingDistribution": { "5": 312, "4": 158, "3": 78, "2": 36, "1": 28 }
  },
  "filters_applied": {
    "ratings": [4, 5],
    "jobTitle": "Software Engineer",
    "locationId": 1147401,
    "currentJob": true,
    "language": "eng",
    "employmentStatus": "REGULAR",
    "sortBy": "DATE",
    "limit": 30
  },
  "reviews": [
    {
      "reviewId": "78912345",
      "permalink": "https://www.glassdoor.com/Reviews/Employee-Review-Stripe-RVW78912345.htm",
      "date": "2026-04-22T00:00:00Z",
      "headline": "Best engineering culture I've worked in",
      "reviewerJobTitle": "Software Engineer",
      "reviewerLocation": "San Francisco, CA",
      "employmentStatus": "Current Employee",
      "tenure": "more than 3 years",
      "employmentType": "REGULAR",
      "overallRating": 5,
      "pros": "Smart coworkers, well-funded teams, high autonomy.",
      "cons": "Some processes feel heavy as headcount grows.",
      "adviceToManagement": "Keep investing in IC career ladders.",
      "recommend": "POSITIVE",
      "ceoApproval": "POSITIVE",
      "businessOutlook": "POSITIVE",
      "helpfulCount": 17,
      "subRatings": {
        "careerOpportunities": 5,
        "compAndBenefits": 5,
        "cultureAndValues": 5,
        "seniorManagement": 4,
        "workLifeBalance": 4,
        "diversityAndInclusion": 5
      }
    }
  ]
}

Shape 2 — Ambiguous (multiple top-tier matches for a bare-name input)

{
  "success": false,
  "reason": "ambiguous_name",
  "query": "Aurora",
  "matches": [
    { "name": "Aurora Innovation", "employerId": 2754012, "headquarters": "Pittsburgh, PA", "reviewCount": 132, "url": "https://www.glassdoor.com/Reviews/Aurora-Innovation-Reviews-E2754012.htm" },
    { "name": "Aurora Cannabis", "employerId": 1419533, "headquarters": "Edmonton, Canada", "reviewCount": 287, "url": "https://www.glassdoor.com/Reviews/Aurora-Cannabis-Reviews-E1419533.htm" },
    { "name": "Aurora Health Care", "employerId": 11651, "headquarters": "Milwaukee, WI", "reviewCount": 1842, "url": "https://www.glassdoor.com/Reviews/Aurora-Health-Care-Reviews-E11651.htm" }
  ]
}

Shape 3 — Authentication required (cookies expired or context not attached)

{
  "success": false,
  "reason": "auth_required",
  "detail": "Session landed on /profile/login_input.htm — Browserbase context cookies are stale or missing. Re-run the operator-side one-time login (see SKILL Workflow step 1) and rotate CTX_ID.",
  "lastUrl": "https://www.glassdoor.com/profile/login_input.htm?from=%2FReviews%2FStripe-Reviews-E671932.htm"
}

Shape 4 — Company not found

{
  "success": false,
  "reason": "company_not_found",
  "query": "Not-A-Real-Company-LLC",
  "lastUrl": "https://www.glassdoor.com/Search/results.htm?keyword=Not-A-Real-Company-LLC"
}