Mapping local search demand with free autocomplete data: 12,782 suggestions for three towns in Hokkaido, and what we kept

Yuki Nakazawa · originally published in Japanese on 4 October 2026 · Japanese original · The same method on San Jose, Santa Clara and Sunnyvale

TL;DR: You don’t need a paid SEO tool to see what people in a small town search for. We pulled the Google and Bing autocomplete suggestions for “town + real-estate topic” in Tomakomai, Shiraoi and Chitose (Hokkaido, Japan) through the Egosaru API: 5,194 unique suggestions from Google alone. After cleaning, 1,372 were about property. They showed that people who want to sell don’t search for “sell”, and Bing surfaced selling and commercial-land searches that Google never showed.

A local real-estate agency that wants inquiries from the web first has to answer a basic question: what words do people actually type? National keyword tools tend to drop small-town terms as “no volume”. So for Banana House Co., Ltd., a real-estate agency in Tomakomai, we ran the whole research step, from collection to clustering, through the new API of Egosaru, the autocomplete tool we operate. Here are the steps and the real numbers.

Paid tools and free autocomplete are a combination, not a choice

To be clear: I use Ahrefs every day. For search volume, keyword difficulty and backlinks of the ranking pages it is indispensable, and this article is not against paid tools.

But when a local business needs to enumerate thousands of “town + something” phrases, checking all of them in a paid tool gets expensive, and many of them are too small to exist in the tool’s database at all. So we split the work: cast a wide net with free autocomplete, then measure only the shortlisted candidates in the paid tool.

Paid tools also have usage limits. Once, after opening a large number of reports in Ahrefs in a short time, I got a “suspicious activity detected” notice and report access was temporarily restricted. Paid tools are not meant for brute-forcing thousands of phrases. Leave the brute force to free autocomplete and use Ahrefs for the words you really need to measure; that respects both the tool’s terms and the budget.

The usage-limit notice Ahrefs showed after I opened a large number of reports (account details masked).

Autocomplete does not tell you how often something is searched. It does draw the map of which word combinations people use, for free. Draw the map first, then measure: that order alone cuts the number of terms you need to send to the paid tool.

The five steps. The costly one is step 3, cleaning: only about a quarter of the terms survived.

Why an API

Egosaru started as a browser tool. I added the API after asking Claude Code for help on another project (autocomplete research for about 500 terms for an affiliate benchmark site). If you fetch Google autocomplete directly in bulk, your IP address gets blocked quickly. The conclusion was to keep the queue and the fetching on Egosaru’s side, behind a single endpoint, so that large and daily jobs run reliably.

Talking it through with Claude Code: why an API, and the minimal spec it proposed.

Claude Code reading the finished API spec and planning around the asynchronous 202 → retry flow and the free-tier limits.

About 500 terms being collected in the background.

Step 1: town names alone return no real-estate terms

Autocomplete for just “苫小牧” (Tomakomai), “白老” (Shiraoi) or “千歳” (Chitose) is all ferries, sightseeing, weather, hotels and Upopoy (the national Ainu museum). Not one real-estate term in the top suggestions. Seed with town names only and you throw most of the harvest away.

So the seeds were “town + real-estate axis”. We used 25 axes (sell, land, used house, single-storey house, condominium, rent, warehouse, shop and so on) × 3 towns = 75 seeds.

Step 2: collect all 75 seeds through the API

For each keyword the Egosaru API returns its autocomplete plus the expansions with every Japanese kana, A–Z and 0–9 appended, so one seed means dozens of upstream queries that the API handles for you. We collected Google, and the same 75 seeds on Bing, because more people now reach Bing through the Edge and Copilot search boxes, and Bing phrases things differently.

Egosaru, Google, “千歳 地価” (Chitose land value). Tokyo’s Chitose-Karasuyama and Chitose-Funabashi are mixed in.

The API response: one JSON array of suggestions per keyword.

curl -H "Authorization: Bearer $KEY" \
  "https://gstudio1.com/api/v1/google/suggest?q=苫小牧%20不動産&flat=1"

A keyword nobody has fetched yet does not come back immediately. You get HTTP 202, it joins the queue, and the same request sent again after retry_after_seconds returns 200 with the result. Results are cached for 7 days. All 75 seeds were done in about 40 minutes; the script only had to repeat “202: wait and resend, 200: save”. The spec is in the Suggest API v1 documentation.

Results varied a lot by seed:

Step 3: from 5,194 terms to 1,372, by throwing away noise

You can’t turn the raw list into articles. The 5,194 unique terms were full of things unrelated to local property. Four kinds of noise dominated:

• “Buyout” (買取) collisions: cars, books, stamps, appliances, Pokémon cards. Of 304 terms containing 買取, 2 were about buying out property.
• Same-name places elsewhere: Tokyo’s Chitose-Karasuyama and Chitose-Funabashi, Chitose in Oita, and so on. 462 of the 2,184 terms containing 千歳 (about a fifth).
• Store openings: “Tomakomai + brand + new store” is retail news, not shop property.
• Competitors and property portals: terms with other agencies’ or portals’ names. Not something to target with our own articles.

The rules were simple: (1) keep terms with a property or transaction word (land, house, condo, rental, vacant house…); (2) keep shop and warehouse terms only together with property / rent / sale / vacant; (3) drop anything matching other-region place names or the exclusion list. That left 1,372 terms. Rules miss things, so a person reviews the final list.

Of 5,194 terms, 1,372 were usable. 462 were “Chitose” places outside Hokkaido and 302 were “buyout” terms not about property.

Step 4: split by intent, then group into article units

We sorted what remained by what the searcher wants to do next: selling, commercial land, prices, buying, renting, choosing an agency. Selling and commercial land get a higher weight because they are where an agency earns most, and the priority score multiplies “how many seeds a term appeared under” by “its rank within the expansion”. Autocomplete has no monthly volume, so this is the stand-in.

Terms by town and intent. Chitose has 1.6× Tomakomai’s price-related terms.

Terms with the same town, the same intent and the same leading word form a cluster, and one cluster becomes one article. For example “苫小牧 平屋 中古 / 苫小牧 平屋 新築 / 苫小牧 平屋 800万円” (Tomakomai single-storey used / new / 8 million yen) is one article. Several near-identical articles would only compete with each other in the rankings.

Bing shows a different face

The same “苫小牧 空き家” (Tomakomai vacant house) on Bing returned terms about the SoftBank data center being built in Tomakomai, plus lodging and furnished rentals, none of which Google showed. It fits a city where a large data center is under construction, and you would miss that demand looking at Google only.

Egosaru, Bing, “苫小牧 空き家” (Tomakomai vacant house): different terms from Google.

Same 75 seeds, side by side:

Bing returns more terms, but many are only loosely related to the seed (“どうして” — “why”, “zozo” — a fashion retailer). Only about 5% survived as property terms. Google for volume, Bing to fill Google’s gaps.

And the 129 Bing-only terms included ones we could not afford to miss:

• Selling: “苫小牧 土地売却” (Tomakomai land sale), “白老町土地売却” (Shiraoi land sale), “苫小牧市マンション売却” (Tomakomai condo sale). On Google, “land + sell” combinations returned almost nothing; on Bing they appear.
• Commercial: “苫小牧東部産業用地” (Tomakomai East industrial land), “千歳市工業用地” (Chitose industrial land), “苫小牧市収益物件” (Tomakomai income property), “千歳市貸倉庫” (Chitose warehouse for rent). High-value deals showed up relatively often on Bing.

Bing is the engine behind Edge, Windows’ default browser, and behind Copilot. Many people search it on a work PC, which may be why business terms are more common there (that is my guess, not a finding). Either way, the biggest lesson was that Google alone missed selling and commercial-land terms, the two that matter most to an agency.

What the data said

On Google, people who want to sell don’t search for “sell”. “苫小牧 土地 売却” (Tomakomai sell land) returned 0 suggestions and “苫小牧 不動産 売却” (Tomakomai sell property) 2. The terms gather around “苫小牧 空き家” (Tomakomai vacant house), “空き家 無料” (vacant house free), “空き家バンク” (vacant-house bank, a municipal listing service) and “不動産 査定” (property valuation) instead. The selling inquiries an agency wants most arrive through the vacant-house topic, so the selling articles should start there. Because Bing does show “land sale” and “condo sale”, we still prepare one article that deals with selling head-on.

Chitose cares about prices. Chitose has 72 price-related terms, 1.6× Tomakomai. “地価 上昇” (land value rising), “地価 高騰” (land value soaring), “土地 値上がり” (land getting more expensive), “ラピダス 千歳 土地” (Rapidus Chitose land): the semiconductor plant moving in is visible in the searches. Official land-price and transaction data can back these articles, which makes them easy to write with evidence.

Shiraoi is small: one article. Shiraoi accounts for about 6% of the three towns’ terms, but has distinctive ones the others don’t: “土地 0円” (land for 0 yen), “温泉付き” (with hot spring), “空き家バンク”. Rather than splitting by district, one hub article for Shiraoi is the right size.

District names are searched as-is. Tomakomai’s Numanohata, Tokai-cho, Utonai and Uenae; Chitose’s Koyodai, Midoridai and Komasato; Shiraoi’s Takeura. That justifies two levels: a town article with district articles under it.

Step 5: rules decided before writing

“Town × something” can generate hundreds of pages mechanically. But thin pages that only swap the town name are treated by Google as doorway pages or scaled content, and they can drag down the rest of the site. So before writing anything we decided:

• Every article carries at least one piece of first-party information only that area has (official transaction prices or land values, the agency’s own deals, photos taken on site).
• Every number has a source and a date. No home-made price estimates, no invented example listings.
• AI-written text is a draft only. A licensed real-estate broker (宅地建物取引士) reviews it before publishing, including compliance with Japan’s Real Estate Brokerage Act ban on exaggerated advertising and the industry’s fair-competition code for property advertising.
• Start with a pilot of about 10 articles, then watch Search Console and inquiries for 2–3 months before deciding to scale.

Summary

Autocomplete research takes more effort in throwing away and grouping than in collecting. Our flow had five steps: seeds from axes → bulk collection through the API → rule-based cleaning → intent classes → clusters as article units. About a quarter of what we collected was usable. But that quarter changed the plan: “the entry point for selling is vacant houses”, “Chitose is about land prices”, “fill selling and commercial land from Bing”.

Paid tools measure; free autocomplete finds. Do the finding widely and for free, and hand only the terms worth measuring to the paid tool. For local businesses, I think that combination gives the best return on keyword research.

Update, October 2026: the Egosaru API now covers 20 languages and adds Baidu, Yandex, Naver and Daum alongside Google, Bing, YouTube, Amazon, Rakuten, Yahoo! JAPAN and DuckDuckGo. The free API key is still free; see the API documentation and pricing. The web tool needs no sign-up: Egosaru.

更新履歴

• 第１稿投稿　2026年10月24日 1時14分（記事コンテンツアップ）