DEV Community

Cover image for Keyword research for several countries with Google Autocomplete, without translating keywords
Joshua Smith
Joshua Smith

Posted on

Keyword research for several countries with Google Autocomplete, without translating keywords

The usual way to localise a keyword list is to send the English list to a translator. The result is grammatical and often wrong: it is what an English speaker would say in German, not what a German searcher types. Google Autocomplete is a cheap way to check, because it shows phrases real people in each market type. This post runs one topic through four markets, compares the suggestion sets, and pulls out what each market searches for that the others do not.

Why translated keywords miss

Three things go wrong, and one product category, running shoes, shows all three:

  • The head term differs, even within one language. Americans say "running sneakers", British searchers say "running trainers".
  • Literal translations pick up the wrong meaning. German laufen means both "run" and "walk", so the literal "schuhe zum laufen" (shoes for running) also returns schuhe zum laufen lernen, shoes for toddlers learning to walk.
  • Borrowed words and regional usage. Spanish searchers type the English loanword: "zapatillas running" sits beside the textbook "zapatillas para correr". Zapatos (shoes) versus zapatillas (trainers) also varies between Spain and Latin America.

None of this shows up in a translated spreadsheet. It shows up when you ask each market's search box.

The method

  1. Pick two or three seed phrasings per market. One is the direct translation; the others come from a native speaker, local retailers' category names and local forums. Keep the direct translation on purpose: it is your control.
  2. Query each market in its own country and language. Autocomplete takes a country code (gl) and an interface language (hl). Set both; gl=de with an English interface is a different audience.
  3. Use the same expansion and the same cap for every seed, so suggestion counts are comparable.
  4. Compare sets, not lists. Measure overlap between markets, then list the words only one market uses.

A note on expansions. The a-z expansion works in any language. The question and preposition expansions add English words (who, what, for, with), so for German or Spanish they produce mixed-language queries such as who laufschuhe. Leave them off for non-English markets and add local question forms as seeds instead, such as welche laufschuhe or qué zapatillas. The a-z letters are ASCII, so continuations starting with ä, ö, ü or ñ get no letter of their own.

Collect the suggestions

Google Autocomplete Scraper is an open-source Apify Actor that sends the expansions, throttles the requests and de-duplicates suggestions per seed. country and language apply to the whole run, so run it once per market. The German input:

{
    "keywords": ["laufschuhe", "joggingschuhe", "schuhe zum laufen"],
    "country": "de",
    "language": "de",
    "expandAlphabet": true,
    "maxSuggestionsPerKeyword": 300
}
Enter fullscreen mode Exit fullscreen mode

In the Console, enter the seeds in Seed keywords, set Country and Language, tick Expand with a-z and click Start; repeat per market. From a shell, save the inputs as us.json, gb.json, de.json and es.json and run:

for m in us gb de es; do
  curl -s -X POST "https://api.apify.com/v2/acts/josh99smith~google-autocomplete-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
    -H "Content-Type: application/json" -d @$m.json -o suggestions-$m.json
done
Enter fullscreen mode Exit fullscreen mode

Each record carries seedKeyword, suggestion, country and language, so the four files can be analysed together without extra labelling.

Compare the markets with a script

The script groups suggestions by country/language, lower-cases them and strips accents (so dämpfung and dampfung match), then prints:

  • suggestions per seed, which shows which phrasings run out early;
  • two overlap matrices using Jaccard similarity (shared items divided by all items in either market): one on exact suggestions, one on words with seed words and common stop words removed;
  • the words common to every market;
  • per market, the words no other market uses, ranked by how many suggestions contain them, with an example.

A word must appear in at least three suggestions in a market to count, which filters one-off noise. Save it as markets.py; Python 3.7+, no packages.

import json
import re
import sys
import unicodedata
from collections import Counter, defaultdict

sys.stdout.reconfigure(encoding="utf-8")  # suggestions contain accents; Windows consoles default to cp1252

MIN_SUGGESTIONS = 3  # a word must appear in this many suggestions to count for a market
STOP = set("""a an the and or of for in on to with by is are de la el los las y en para con del
der die das und für mit von im zum zur ein eine""".split())


def fold(text):  # lower case and strip accents, so "für" and "fur" match
    text = unicodedata.normalize("NFKD", text.casefold())
    return "".join(c for c in text if not unicodedata.combining(c))


STOP = {fold(w) for w in STOP}


def words(text):
    return re.findall(r"[a-z0-9]+", fold(text))


markets = defaultdict(dict)  # "gb/en" -> {folded suggestion: original}
seeds = defaultdict(Counter)
failed = Counter()
for path in sys.argv[1:]:
    for r in json.load(open(path, encoding="utf-8")):
        m = f"{r['country']}/{r['language']}" if r.get("country") else "?"
        if not r["success"]:
            failed[m] += 1
            continue
        markets[m].setdefault(" ".join(words(r["suggestion"])), r["suggestion"])
        seeds[m][r["seedKeyword"]] += 1

vocab = {}  # market -> {word: suggestions containing it}, seed words and stop words removed
for m, sugg in markets.items():
    seed_words = {w for s in seeds[m] for w in words(s)}
    count = Counter(w for s in sugg for w in set(s.split()) - seed_words - STOP if len(w) > 1)
    vocab[m] = {w: n for w, n in count.items() if n >= MIN_SUGGESTIONS}

names = sorted(markets)
print("Suggestions per seed (same cap for every seed):")
for m in names:
    print(f"  {m:6}", ", ".join(f"{s!r} {n}" for s, n in seeds[m].most_common()),
          f"| failed requests: {failed[m]}")


def matrix(title, sets):
    print(f"\n{title}")
    print(" " * 7 + "".join(f"{m:>8}" for m in names))
    for a in names:
        cells = []
        for b in names:
            inter, union = len(sets[a] & sets[b]), len(sets[a] | sets[b])
            cells.append(f"{100 * inter / union if union else 0:>7.0f}%")
        print(f"{a:7}" + "".join(cells))


matrix("Shared suggestions (Jaccard, exact text after folding):", {m: set(s) for m, s in markets.items()})
matrix(f"Shared words (Jaccard, words in {MIN_SUGGESTIONS}+ suggestions, seed words removed):",
       {m: set(v) for m, v in vocab.items()})

shared = set.intersection(*(set(v) for v in vocab.values()))
shared = sorted(shared, key=lambda w: (-sum(v[w] for v in vocab.values()), w))
print(f"\nWords common to every market ({len(shared)}):", ", ".join(shared[:20]))

for m in names:
    others = set().union(*(set(w for s in markets[o] for w in s.split()) for o in names if o != m))
    unique = [(n, w) for w, n in vocab[m].items() if w not in others]
    print(f"\n{m}: {len(unique)} words no other market uses, by suggestions containing them")
    for n, w in sorted(unique, reverse=True)[:10]:
        example = next(orig for s, orig in markets[m].items() if w in s.split())
        print(f"  {w:16} {n:>4}  e.g. {example}")
Enter fullscreen mode Exit fullscreen mode

Run python markets.py suggestions-*.json.

Read the result

Here is trimmed output from one sample run in November 2026 with the inputs above (US seeds: "running shoes", "running sneakers"; UK: "running shoes", "running trainers"; Spain: "zapatillas running", "zapatillas para correr", "zapatos para correr"). Treat it as illustrative output from sample data: suggestions change from day to day, and this is one topic.

Suggestions per seed (same cap for every seed):
  de/de  'laufschuhe' 300, 'joggingschuhe' 300, 'schuhe zum laufen' 160 | failed requests: 0
  es/es  'zapatillas running' 300, 'zapatillas para correr' 300, 'zapatos para correr' 293 | failed requests: 0
  gb/en  'running shoes' 300, 'running trainers' 300 | failed requests: 0
  us/en  'running shoes' 300, 'running sneakers' 300 | failed requests: 0

Shared suggestions (Jaccard, exact text after folding):
          de/de   es/es   gb/en   us/en
de/de      100%      0%      0%      0%
es/es        0%    100%      0%      0%
gb/en        0%      0%    100%     17%
us/en        0%      0%     17%    100%

Shared words (Jaccard, words in 3+ suggestions, seed words removed):
          de/de   es/es   gb/en   us/en
de/de      100%      8%     10%      9%
es/es        8%    100%      9%      8%
gb/en       10%      9%    100%     48%
us/en        9%      8%     48%    100%

Words common to every market (10): nike, adidas, asics, new, balance, hoka, brooks, trail, puma, cloud

de/de: 74 words no other market uses, by suggestions containing them
  damen              89  e.g. laufschuhe adidas damen
  herren             78  e.g. laufschuhe asics herren
  lernen             23  e.g. schuhe zum laufen lernen
  dampfung           10  e.g. laufschuhe gute dämpfung
  breite             10  e.g. laufschuhe breite füße

us/en: 9 words no other market uses, by suggestions containing them
  similar             9  e.g. running shoes similar to brooks ghost
  nearby              6  e.g. running shoes nearby
  hsa                 3  e.g. running shoes eligible for hsa
Enter fullscreen mode Exit fullscreen mode

How to read it:

  • Same language, different market. The US and UK share 48% of their words but only 17% of their exact suggestions. A UK page built from the US list targets mostly phrases Britons do not type.
  • Across languages, the overlap is brands. The 8 to 10% of words shared between languages are brand and product names, plus "trail". They are the only keywords you can reuse untranslated.
  • A seed that runs out early is a hint. "schuhe zum laufen" stopped at 160 suggestions, under the cap, and much of what it found was about walking. Autocomplete counts are not search volume: a seed that fills the cap means "people type this", not "most people type this".
  • Unique words mix translation with real differences. Damen and herren are just German for women and men. The useful rows are concepts with no counterpart elsewhere: dämpfung (cushioning) and breite füße (wide feet) in Germany, asfalto and placa de carbono (carbon plate) in the Spanish list, HSA-eligible in the US. Read each list with a native speaker before you plan content.

Cost and limits

The Actor charges $0.003 per seed keyword, whatever the expansions and however many suggestions come back. The sample above used 10 seeds across four runs: 10 x $0.003 = $0.03. A fuller study of 25 seeds in each of four markets is 100 x $0.003 = $0.30. A seed with any failed request (rate limiting, network) is not charged, and its failures come back with an errorType. Each run also stops at the maximum cost you set.

Limits to plan around:

  • One country and language per run, so one run per market.
  • With a-z on, each seed sends up to 27 requests and stops early once it reaches maxSuggestionsPerKeyword (default 50). Raise the cap for a comparison like this, or every seed hits 50 and the counts tell you nothing.
  • Suggestions change over time, so date your results and repeat the run before drawing conclusions.
  • Autocomplete shows what people type, not how often. Pair it with a volume source before you prioritise.
  • If Google starts rate-limiting, enable Apify Proxy in the input; proxy traffic is billed separately by Apify.

If this saves you time, a review on the Store page helps other people find it, and a star on GitHub helps too.

Disclosure: I built this Actor. Source: github.com/josh99smith/google-autocomplete-scraper. Google is a trademark of Google LLC, and the brand and retailer names in the sample output belong to their owners; the Actor is not affiliated with Google or any of them.

Top comments (0)