Each connected source, how fresh it is, what it is allowed to be used for and what its numbers do not mean. Calculation index-calc-v1, taxonomy taxonomy-2026-11-v1.
5 OF 6 SOURCES LIVE
Sources collecting
5/6
Product Hunt waiting
Records fetched
6,499
Every version of every listing is kept
AI-relevant records
3,415
53% passed the AI relevance filter
Classification model
Configured
Schema-validated, retried, quarantined on failure
Connected sources
Official APIs only. Every record keeps its original link and fetch time.
Used only to find and verify funding announcements for tracked products. Fetching a page is not a licence to republish it: only a short evidence quote and the link are stored.
Coverage
Searched per product, at most once a week each. A deal is verified only when the article quotes the amount and names the product's own website.
17 searches so far. Each product is searched at most once a week; only deals that pass every check are stored.
What makes a deal verified
How it is measured
The rules every chart on the overview follows.
Building activity
Distinct products whose earliest public launch signal falls in the window: a Uneed launch day, a Product Hunt launch, a Show HN post or a repository creation date. A product listed on several sites counts once. One product can sit in several cells, so cells do not add up to the total.
Attention
Votes, points and stars are never added together. Each product's latest metric becomes a percentile within its own source among products from the same window, then the percentiles it has are averaged. A cell shows the median and needs at least three scored products.
Verified funding
Only announcements stored with a primary source count. Undisclosed amounts stay unknown and are never summed as zero. With no verified deals, funding views stay empty.
Gaps and change
Hatched cells and shaded chart stretches mean the sources were not collecting then. Change against the previous period is shown only when that period was collected too.
AI classification
What the model does, and what it is not allowed to do.
How tags are made
Products are classified by the configured model through an OpenAI-compatible endpoint. Output is schema-validated against the controlled vocabulary and every tag must cite one of the product’s own records. Failed batches are retried with the errors, then held for review. A person’s edit locks that axis from later AI runs.
Suggested use cases
757 new labels proposed by the model are waiting for review and are not counted yet: 3D Exploration (1), 3D Mesh Generation (1), 3D Mockups (1), 3D Model Generation (1), 3D Modeling (2), 3D Museum (1), 3D Rendering (3), AI Companionship (1), AI Content Critique (2), AI Content Filtering (1), AI Cost Tracking (2), AI Detection (2)…
Admin
Correct tags, refresh sources and review held items.
Official GraphQL API v2 (posts by day, topic artificial-intelligence), developer token
Rate limits
Complexity-based limits per 15 minutes. One request every 1.2 s, backoff on 429/5xx.
Use and licence
API terms: not for commercial use without Product Hunt's permission. Votes are Product Hunt community attention, not usage. Maker names are not stored.
Coverage
AI-topic launches per day, newest first. Only active once PRODUCT_HUNT_TOKEN is set.
PRODUCT_HUNT_TOKEN is not set. Create a developer token at producthunt.com/v2/oauth/applications.
Official NSF Award Search API (/services/v1/awards.json) by AI keyword and award date
Rate limits
Public API, no key. One request every 1 s, up to 100 awards per keyword, backoff on 429/5xx.
Use and licence
US federal award records, effectively public domain. The obligated amount is the award as recorded by NSF, not a private funding round. Academic, government and non-profit awardees are excluded so only companies enter the index. People named on an award are not stored.
Coverage
Awards announced in the window for eight AI keywords (artificial intelligence, machine learning, large language models, robotics, computer vision, NLP, autonomous systems, AI safety). Award date is the announcement date, not the project start date.
Public Y Combinator company directory (api.ycombinator.com/v0.1/companies), undocumented
Rate limits
Undocumented public endpoint. One request every 600 ms, capped at 20 pages (500 companies) per pass.
Use and licence
Undocumented endpoint, so it can change without notice. Batch is published as a season and year only; the stored launch date is the first day of that batch (Winter January, Spring April, Summer July, Fall October) and must not be read as a precise launch day. Team size is what YC displays, not headcount we measured.
Coverage
Newest batches first, 25 companies per page, 20 pages per pass. Re-reads the same head pages each pass; known companies are deduplicated, so deep history needs a larger page cap.
Recent collection runs for Y Combinator
Started
Window
Status
Fetched
New
Requests
35 min ago
Nov 18, 2026 to Nov 20, 2026
success
500
0
20
46 min ago
Nov 18, 2026 to Nov 20, 2026
success
500
2
20
1 h ago
Nov 18, 2026 to Nov 20, 2026
success
500
1
20
2 h ago
Nov 18, 2026 to Nov 20, 2026
success
500
0
20
3 h ago
Nov 18, 2026 to Nov 20, 2026
success
500
0
20
Reddit, G2, Capterra and the other candidate directories are not connected; each needs its API terms or licence confirmed first.
A news search finds a headline naming the product together with funding words.
The article is fetched and the AI quotes the sentence announcing the raise.
Code checks the quote appears word for word in the article and states the amount.
The article must also mention the product’s own website, so a different company with the same name is never counted.
Anything that fails a step is kept only in the search log, not on the product.