Google Maps Reviews & Sentiment Extractor

Correction (2026-09-16): the claim below that free-text place search is closed is wrong. The sibling actor google-maps-lead-generation-scraper searches Google Maps successfully; the pb used here was malformed (it omitted the !7i20!10b1!12m6… result-request tail, so Google returned search metadata with no result list). This actor still takes place identifiers only, but the limitation is a design choice, not a platform restriction.

Pull every public Google Maps review for a place — rating, date, reviewer, like count, photos and the owner's reply — and get each one scored positive, neutral or negative automatically. Export to JSON, CSV, Excel or NDJSON.

No login, no Google API key, no browser. Pure HTTP.

---

What you get per review

Field	Notes
rating	1–5 stars
text / textTranslated	Original text plus Google's translation when it has one
publishedAt	Real UTC timestamp, not just "2 months ago" (relativeTime has that too)
reviewer	Name, profile URL, avatar, total reviews, total photos, Local Guide flag
likesCount	How many people found the review helpful
ownerResponse	The business reply — text and when it was posted
photos	Photos attached to the review
attributes	Google's structured answers, e.g. "When did you visit?" → "Weekday"
sentiment	label, score (−1…1), method, and ratingTextConflict
permalink	Direct link to that single review

Each place also produces one PLACE summary row: sentiment counts and percentages, average rating, rating histogram, owner-response rate, total likes, and a count of rating/text conflicts.

---

Input

Supply places as Google Maps URLs, Place IDs, CIDs, or feature IDs:

{
  "placeUrls": [
    { "url": "https://www.google.com/maps/place/Eiffel+Tower/@48.8583701,2.2944813,17z/data=!4m6!3m5!1s0x47e66e2964e34e2d:0x8ddca9ee380ef7e0!8m2!3d48.8583701!4d2.2944813" },
    { "url": "https://www.google.com/maps?cid=3292831917685799941" },
    { "url": "https://maps.app.goo.gl/xxxxxxxx" }
  ],
  "placeIds": ["ChIJrTLr-GyuEmsRBfy61i59si0", "0x6b12ae37b47f5b37:0x8eaddfcd1b32ca52"],
  "mode": "reviews",
  "sortBy": "newest",
  "maxReviewsPerPlace": 300,
  "exportFormats": ["csv", "excel"]
}

A free-text search query is not accepted. Google closed every HTTP-only place-search surface, so this actor works from place identifiers. Paste the URL from your browser's address bar and you are done. See Limitations.

Modes

Mode	Emits	Use it for
reviews (default)	Every review + a PLACE summary	Full extraction and analysis
summary_only	PLACE summary only, one page fetched	Cheap sentiment snapshot across many places
owner_responses	Only reviews the business replied to	Auditing reply rate and reply quality

Useful options

Option	Default	Notes
sortBy	newest	newest is the only order that reliably walks the whole list. relevant returns a mixed-rating sample.
maxReviewsPerPlace	100	Google serves 60 on page 1, then 20 per page.
minRating / maxRating	—	Set both to 1 and 2 to isolate complaints.
onlyWithOwnerResponse	false	Same filter as owner_responses mode.
language / region	en / US	hl and gl. Also picks the translation target language.
exportFormats	[]	json, csv, excel, ndjson → written to the key-value store.

---

How the sentiment scoring works

Each review is scored by blending VADER text polarity with the star rating, because neither alone is reliable:

• the rating is coarse — 3-star reviews are often strongly opinionated, and 5-star reviews frequently contain real complaints;
• the text lexicon has no opinion at all on rating-only reviews, which are a large share of Google Maps data.

The default blend is 60% text / 40% rating, and method records which signals a given row actually used:

method	Meaning
text+rating	English review text blended with the rating
translated+rating	Non-English review, scored via Google's English translation
rating-only	No text, or a language with no translation available
none	Neither text nor rating — scored neutral, never dropped

Sarcasm handling. A lexicon cannot detect sarcasm: "Highly thrilling experience 😡" on a 1-star cancellation complaint scores +0.94 on text alone. When a confident text polarity contradicts an extreme rating (1, 2 or 5 stars), the rating is weighted to 75% — the reviewer's own star verdict is the better ground truth for satisfaction. That review correctly lands at −0.54, negative. Those rows are flagged with sentiment.ratingTextConflict: true, and counted per place as ratingTextConflicts — they are usually the most interesting reviews in the set, so they are surfaced rather than just corrected.

Scoring runs inside the actor: no API key, no per-review cost, no data sent anywhere.

---

Limitations (read before buying)

These are properties of what Google still exposes publicly, verified rather than assumed. Full evidence in CRAWLING_METHOD.md.

1. No free-text place search. Google's Maps search is JS-only for non-browser clients and every legacy search endpoint now 404s. Supply place identifiers. Unrecognised input gets a clear error, never a wrong place.
2. Place name and coordinates come from your URL, not from Google. Google's resolver returns null for the name and the same default world-view viewport for every place. Paste a full /maps/place/<Name>/... URL and both are captured; supply a bare Place ID and both are null.
3. Geographic filters filter, they do not discover. Because of (1) and (2), boundingBox / radiusKm / areaName narrow the places you supply to those inside an area — they cannot find new places in it. A place with no coordinates is kept with a warning by default, or skipped with a SKIPPED record if you set onMissingCoordinates: "skip". Nothing is dropped silently.
4. Reviews per place are finite. Google stops serving pages well before the lifetime review count of large places; the actor stops when the cursor runs out and reports how many it got.
5. averageRating is the average of what was scraped, not Google's lifetime average for the business.
6. Sentiment is English-centric. Non-English reviews are scored through Google's translation where available, and fall back to rating-only otherwise. method always tells you which happened.

---

Anti-bot and proxy

Reconnaissance found no WAF on this endpoint: ten TLS fingerprints across five Google country hosts all returned 200 from a plain residential connection.

Do not use residential proxy here. This is the opposite of the usual advice and it was measured, not assumed:

Egress	Result
Plain home ISP, no proxy	works
Apify platform, no proxy (the default)	works — 42 records in 7s
Apify platform, RESIDENTIAL	fails — Google answers the place resolver with an interstitial instead of the page

Google polices residential proxy pools far harder than datacenter ranges, because those IPs are widely abused. Since these endpoints have no WAF, the residential IP buys nothing and costs you the run.

The proxy is therefore off by default and the actor runs on the platform's own IP. Enable one only to spread load across IPs at high volume, and leave the group list empty so Apify picks a group your plan actually has — naming a group your account lacks (a free plan has no DATACENTER) fails the run at input validation before it starts:

{ "proxyConfiguration": { "useApifyProxy": true } }

The client impersonates Chrome 150 via curl_cffi, warms the Maps cookie jar once per session, retries with exponential backoff, and rotates TLS profile and cookie jar on 403/429/503.

---

Output shape

Records follow the portfolio envelope — _input, _source, _scrapedAt, recordType — with _error / _errorDetail on failures and _warning on skips. Split a run by recordType: PLACE, REVIEW, SKIPPED, ERROR.

A place that fails to resolve produces an ERROR row rather than vanishing, so you can always reconcile inputs against outputs.

---

Development

pip install -r requirements.txt
python test_errors.py     # offline checks: parsing, geo, sentiment, exporters
python test_local.py      # live end-to-end run against Google

test_local.py reads _input.json when present, otherwise uses its built-in default input.