← Radar Labs
How we choose what to build
Before writing code for a product, we measured the market. Here is the raw output, including the parts that talked us out of ideas.
The method
Every product we consider has to clear three gates, in this order:
- Can I get the data without a browser, a proxy, or an account? If a source needs a logged-in session, a captcha, or residential IPs, we skip it — those tools break on day 30 and take the product with them.
- Do new entrants in this niche actually get users? Star counts mean nothing. We take every actor in a search result, read its creation date from the public API, and compute users per month of age. A niche where the best newcomer gets 2 users/month is a graveyard no matter how big the head is.
- Is what we would sell the thing people already pay for? A niche can be busy and still want something else — see the salary-benchmark story below.
What 55 niches looked like (measured 2026-10-22)
“Supply” = actors matching the search term. “Newcomer u/mo” = median users-per-month-of-age among actors launched in the last 12 months.
| Niche | Supply | Newcomer u/mo | Our verdict |
|---|
| douyin / xiaohongshu | 213 / 194 | 119 / 67 | Best numbers — but needs browsers & anti-bot work. Skipped. |
| trustpilot | 470 | 38.5 | Attractive, but the API is 403 and the HTML blocks us. |
| seo / backlink | 7,051 | 30 | Crowded but alive; needs paid data sources. |
| salary / jobs | 6,510 | 27.8 | This is what we built. |
| flight prices | 3,806 | 20.6 | Needs a paid flight API to be useful. |
| product hunt | 3,009 | 11.6 | Small but keyless. Candidate for a third product. |
| mcp server | 1,297 | 8.5 | Hyped, but nobody pays. |
| ai agent / rag / prompt | 12,173 | 2–6 | Hyped, but nobody pays. |
| llm pricing / openrouter | 748 / 154 | 0.7 / 1.2 | Worst numbers in the whole scan. |
The uncomfortable finding. Everything adjacent to AI had terrible newcomer traction
(LLM pricing 0.7, OpenRouter 1.2, prompts 2.1, MCP servers 8.5), while “boring” data with a clear buyer —
job listings, salaries, flight prices — scored 10–40× better. We built the boring thing.
The idea we killed after measuring it
A salary-benchmark product looked obvious: we already had normalised salary data. Then we looked at who offers one:
| Actor | Users | Runs |
|---|
hirebase/salary-benchmark | 3 | 24 |
thebee/SalaryBenchmark | 2 | 21 |
northbell/linkedin-salary-benchmark | 2 | 24 |
studio-amba/eu-salary-benchmark | 1 | 212 |
| …while the biggest job-board scraper has | 27,860 | — |
Four people had already tried the same idea and none of them got traction. What sells is the raw listing data,
not our opinion about it. So instead of a fifth benchmark we added connectors — which is why the ATS Job Feed
covers six systems and ships a list of 259 verified companies.
What that produced
- ATS Job Feed — 259 companies, 18,825 open jobs, 8,071 with a published pay range, 6,848 after de-duplication. On Apify →
- Freelance Lead Radar — five sources, noise removed, ranked by USD/hour. On Apify →
中文摘要
选品不是靠感觉，是先量市场：
① 数据源能不能「无浏览器、无代理、无账号」拿到——要登录态/验证码/住宅 IP 的一律跳过，那种工具第 30 天就坏；
② 这个赛道里新上线的产品拿不拿得到用户——不看星标，只看「用户数 ÷ 上线月数」，最好的新人也只有 2 用户/月的赛道直接放弃；
③ 大家真正在付钱买的是不是我要卖的东西——薪资基准（salary benchmark）有 4 个人做过，用户数分别是 3 / 2 / 2 / 1，而纯招聘数据抓取有 27,860 用户。所以我们没做第 5 个基准，而是把连接器做多。
顺便一句实话：AI / LLM 周边的数据产品在 Apify 上几乎没人买（OpenRouter 1.2、LLM 定价 0.7），
「无聊」的岗位与薪资数据反而是 10–40 倍。