Skip to content
Docs menu

Signal-to-noise

Signal-to-noise — answer-bearing links

In short: Keep the file to the pages that answer questions, 92 links or fewer; cut tag, pagination, cart and login links first.

  • Category: Curation Quality · weight 16 of 100 within the category
  • Curation Quality is worth 25 in generic and commerce, 20 in regulated and technical (points of 100, by scoring profile)
  • Audit ID C1 (as shown on reports) · rubric v2.1

What it checks

The share of the file’s links that lead to answer-bearing pages.

Why it matters

Every link costs an agent attention and context. Tag archives, pagination, cart and login pages crowd out the pages that actually answer questions.

When it applies

Applies when the file has links. A file with no links is Not applicable here.

An audit that doesn't apply, or that we couldn't verify, is removed from your score and its weight shared among the Curation Quality audits that do apply. It is never scored as a zero.

How it's scored

From the number of links: min(1, 60 ÷ links + 0.35). Up to 92 links score 1.0; 120 links score 0.85, 200 score 0.65, and the score never drops below 0.35. Every link the parser reads counts, ## Optional included.

How to fix

  1. Remove links to tag, category and pagination pages, and to cart, login and account pages.
  2. Replace index dumps with the handful of pages people actually ask about.
  3. Ask of each link: would an agent quote this page in an answer?

Frequently asked questions

How many links can an llms.txt have before Signal-to-noise costs points?
Up to 92. Today the score is min(1, 60 ÷ links + 0.35), so 92 links or fewer earn full marks, 120 links score 0.85, 200 score 0.65 and 600 score 0.45, and it never drops below 0.35. The link count is how the engine measures the rubric’s answer-bearing ratio: the longer the list, the more of it tends to be padding.

Sources: LLMSRank — Signal-to-noise audit · llmstxt.org — The /llms.txt file, v2

Do links in the Optional section, bare URLs and repeated links all count toward the total?
Yes. Every list item the parser reads as a link counts: - [Name](url) lines and bare - https://… lines in every section, ## Optional included, since llms.txt v2 made Optional a convention for links an agent may skip rather than something outside the file. A page linked twice counts twice here, and also lowers the unique-URL share in URL hygiene.

Sources: llmstxt.org — The /llms.txt file, v2 · LLMSRank — Signal-to-noise audit

Which links should I cut first to improve my signal-to-noise?
The ones an agent would never quote: tag, category and pagination pages; cart, login and account pages; and auto-generated lists such as every blog post or product variant, which one link to their index page can replace. The proposal asks for “concise, expert-level information” and warns that a sitemap “will include a lot of information that isn’t necessary to understand the site”. Every cut also helps Token budget and Curation, not a sitemap.

Sources: llmstxt.org — The /llms.txt file, v2 · LLMSRank — Signal-to-noise audit

My docs site has thousands of pages. How do I stay within the link limit?
Split it by path. llms.txt v2 lets a file sit at any path, such as /docs/llms.txt, covering the pages under it, and says agents should use the most specific file that applies. LLMSRank grades the file that covers the URL you scan (see Covering llms.txt), so each part of a large site can carry its own curated list of 92 links or fewer.

Sources: llmstxt.org — The /llms.txt file, v2 · LLMSRank — Covering llms.txt audit

How do real llms.txt files do on signal-to-noise?
In our research, 174 of the 187 sites in our corpus (captured 2026-10-10, scored with rubric v2.1) had links to score, and 113 of them (64.9%) earned full marks; the mean score was 0.881. netlify.com, expedia.com, cloudflare.com, attio.com and retool.com are among the files at full marks.

Sources: LLMSRank — Method · LLMSRank — Signal-to-noise audit