How we score
The score comes from fixed rules applied to live public signals, and every rule is on this page. The code imports the same numbers you see here, so the page and the scorer can't drift apart.
1. From your sentence to a search term
We remove filler words (“a”, “app”, “platform”, “that helps”…) and common verbs from your sentence. The longest remaining run of words, up to three, becomes the search term: “A mobile coffee van that serves office parks” gives “mobile coffee van”. You can set your own term under Options, and you should for non-English markets. Only this search term goes to Google and Hacker News. Your full sentence never does. It goes only to the AI summary, and only when AI is on.
2. The signals
| Tag | Source | What we fetch |
|---|---|---|
| D | Google autocomplete (public suggest endpoint) | Suggestions for 5 searches: the term itself, “best …”, “how …”, “… for” and “… vs”. We keep distinct suggestions that contain a word of the term. How Google autocomplete works. |
| C | Hacker News “Show HN” via the HN Search API, plus comparison suggestions from D | Launches from the past 5 years whose title contains the term's words (checked among the 30 most relevant), and suggestions like “… vs …”, “… alternative”, “… reviews”. |
| H | Hacker News stories via the HN Search API | Stories from the past 3 years whose title contains the term's words (checked among the 30 most relevant). |
| N | RDAP, the registry's public lookup (ICANN). For .com the IANA bootstrap file points to Verisign. | 3–5 name ideas built from the term, with well-known brand names left out. “No record” (HTTP 404) means probably available. “Registered” means taken. |
3. Sub-scores (0–100 each)
Demand (weight 35)
- Distinct, relevant autocomplete suggestions: 2 points each, up to 30 suggestions (60 points).
- Question-style suggestions (how, what, is, can…): 5 points each, up to 5 (25 points).
- Buying or solution-seeking suggestions (price, cost, near me, hire, best, app, generator, template…): 3 points each, up to 5 (15 points).
Room to compete (weight 25)
- “Show HN” launches from the past 5 years whose title contains the search term’s words (checked among the 30 most relevant): each takes 6 points off, up to 10 launches (60 points).
- Comparison suggestions (vs, alternative, reviews, pricing): each takes 8 points off, up to 5 (40 points).
Room to compete starts at 100, and each visible competitor takes points off. When only one of the two inputs is available, we scale the score to that input's points.
Community interest (weight 25)
- Hacker News stories from the past 3 years whose title contains the search term’s words (checked among the 30 most relevant): 10 points each, up to 10 stories (100 points).
Community interest comes from Hacker News only, at full weight. It's a tech-leaning crowd, so a low score for a local or consumer idea means “not discussed there”, not “nobody cares”. We deliberately don't collect Reddit or other forum data automatically: Reddit's robots.txt disallows it and its Data API terms require a separate agreement for commercial use. Each report links to a Reddit search you can run yourself.
Name availability (weight 15)
- Share of the suggested .com names with no registration record in RDAP (e.g. 2 of 5 = 40).
4. The overall score
The overall score is the weighted average of the sub-scores that had data. A sub-score with partial data counts at half weight. A missing sub-score is left out, never set to zero. With fewer than two sub-scores we don't show an overall score at all.
| Score | Label |
|---|---|
| 80–100 | Strong signals |
| 60–79 | Promising signals |
| 40–59 | Mixed signals |
| 0–39 | Weak signals |
Confidence is higher when all four sub-scores have full data and demand returned at least 10 suggestions. It is medium with three or more sub-scores and low otherwise.
5. The written summary
When AI is switched on, Cloudflare Workers AI (@cf/meta/llama-3.1-8b-instruct, Built with Llama) writes three short paragraphs: what the signals say, the biggest uncertainty, and the cheapest next test. The model sees your idea sentence, marked as untrusted data, plus the evidence listed above. It is told not to give financial, investment, legal or tax advice and not to predict success. We then check its answer and throw it away if:
- it cites fewer than two different signal tags, or skips any of the three paragraphs;
- it contains any number, quantity word (“thousands”, “millions”) or money or percent sign that isn't in the evidence (a guard against invented statistics);
- it names a company or product that isn't in the evidence or your sentence;
- it gives advice or predictions about investment, loans, funding, profit, revenue, legal or tax matters, or promises success;
- it runs too long, contains links or formatting, or takes longer than 20 seconds.
When the AI answer is rejected, or AI is off, you get a rules-based summary built from the same numbers. The report always says which kind you're reading, and AI-written text is also marked in the page's code (data-ai-generated). Cloudflare says it does not use Workers AI inputs or outputs to train models (Workers AI data usage). The Llama 3.1 model is used under the Llama 3.1 Community License.
6. What the score can't tell you
- Search volumes. Autocomplete shows what people type, not how many people type it.
- Who ranks on Google. There is no free way to get Google results within Google's terms, so we link you to the search instead of scraping it.
- Offline competitors. Hacker News covers software launches and a tech-leaning audience. Local shops, marketplace sellers and established brands mostly won't appear.
- Willingness to pay, margins, regulation, your skills. We don't score monetisation, because we have no live, reliable signal for it yet.
- The future. A high score means strong signals today, not a successful business. The score is informational only. It is not investment or financial advice.
7. Limits and fair use
Ideas can be 10–280 characters long. Each network gets a small number of reports per 10 minutes and per day. We count them under a salted, daily-rotating hash of the IP address and never store the address itself. Identical requests within a few hours come from a short-lived cache, so nobody's sources get hammered. We don't score ideas that plainly describe illegal activity.