Every month hundreds of products launch on Hacker News. Their founders can rank on Google and still be invisible to ChatGPT, Perplexity and Claude, because assistants do not read a page the way a browser does. We crawled 118 recently launched sites and scored each on nine checks that determine whether an AI assistant can describe and cite them.
| Score band | Sites | Share | Distribution |
|---|---|---|---|
| 0-39 | 19 | 16.1% | |
| 40-59 | 22 | 18.6% | |
| 60-79 | 41 | 34.7% | |
| 80-89 | 22 | 18.6% | |
| 90-100 | 14 | 11.9% |
Only 1 of 118 (1%) scored a clean 100. The median launch leaves roughly one third of the available signal on the table.
These are the checks that failed most often, as a share of all 118 audited sites:
| Failed check | Sites | Share | Rate |
|---|---|---|---|
| No structured data | 69 | 58.5% | |
| No llms.txt | 59 | 50.0% | |
| No canonical tag | 38 | 32.2% | |
| Homepage copy under 800 words | 38 | 32.2% | |
| No working sitemap.xml | 34 | 28.8% | |
| AI crawlers not explicitly allowed in robots.txt | 31 | 26.3% | |
| Missing H1 | 31 | 26.3% |
Organization, Product or FAQPage markup an assistant has to infer what you sell from prose. Inference is where it guesses, and guesses get attributed to a competitor.llms.txt. It is a plain-text file that tells assistants what your site covers and what may be quoted. It takes fifteen minutes to write and almost nobody does it./robots.txt, /llms.txt, /sitemap.xml, then parse the homepage for JSON-LD, canonical, H1 and body text. No headless browser, no API key. The scoring is nine equally weighted checks.Sample: domains submitted as Show HN posts in the 45 days before 2026-11-14, one crawl each, English-language sites only, social and code-hosting URLs excluded. Failures caused by a temporary outage were retried twice and then dropped. Scores describe technical readiness for AI retrieval, not traffic, revenue or ranking. A site can score 100 and still sell nothing — this benchmark measures whether an assistant can read you, not whether anyone asks.
Fix the checks this benchmark measures →Ten tools, one-time from $19. Each ships the exact fix for a failed check — structured data, llms.txt, canonical, body copy. No subscription.