Methodology · SARI v0.2
How the SARI is scored, and how to cite it
The Scaletific Agent-Readability Index scores how legible a website is to AI agents, out of 100. This page is the single reference for the rubric, the sample, the sites we leave out of the averages, the version history and the citation. Every headline figure on this site comes from the numbers below.
The headline figures
One set per wave. Scored sites are those where we sampled at least one page (high and medium confidence). Every mean, median, range and cohort figure uses the scored set. Site-level adoption counts (llms.txt, sitemap, MCP, AI bot directives) do not depend on page sampling, so they cover every audited site.
| Wave | Audited | Scored | Excluded (low confidence) | Mean | Median | Range |
|---|---|---|---|---|---|---|
| Wave 1: publishers | 103 | 93 | 10 | 49.9 | 57 | 15 to 81 |
| Wave 2: e-commerce | 262 | 169 | 93 | 30.9 | 28 | 10 to 65 |
Wave 1 adoption, over all 103 audited publishers: 6 have an llms.txt file, 5 have an MCP well-known endpoint, 100 have a sitemap, 75 carry AI bot directives in robots.txt, and 8 of those 75 treat training and citation crawlers differently.
Why you may have seen other numbers
Before 2026-11-06, a few pages showed a Wave 1 mean of 45.8 and a median of 54. Those figures included the 10 low-confidence sites, which scored near zero because no page could be sampled. That says more about whether we could reach their articles than about how legible they are, so they are not in the headline. The scored-set figures (49.9 mean, 57 median, n=93) are the ones to quote. No site was re-audited and no score changed.
The rubric
100 points across five categories. Three are universal and two are forked per vertical. Every signal is binary or near-binary and detectable in code, so two auditors running the same script on the same site get the same score. No human judgement enters the score.
| Category | Points | What it measures |
|---|---|---|
| Discovery | 25 | llms.txt, sitemap, explicit AI bot directives in robots.txt, MCP well-known endpoint |
| Page Structure | 30 | Structured data for the page type (Article for publishers, Product for e-commerce) |
| Identity & Attribution | 20 | Author and publisher identity, or brand and reviews for e-commerce |
| Content Addressability | 15 | Stable URLs, heading anchors, a clear main content block |
| AI Bot Policy Clarity | 10 | Separate treatment of training crawlers and citation crawlers |
Is llms.txt worth 10 points? The weight test
In v0.2, llms.txt is the largest single signal: 10 of the 25 Discovery points. Two large studies have since measured it. SE Ranking (15 October 2026, nearly 300,000 domains) found no correlation between having an llms.txt file and how often AI assistants cite a domain, and its model grew more accurate with the file left out. Ahrefs (15 June 2026, 137,210 domains) found that 97% of valid llms.txt files received no requests at all in May 2026, and AI search crawlers made about 1% of the requests the rest received.
So we tested our own weights. Under a v0.3 proposal, llms.txt drops to 3 points, a sitemap rises to 7 (search engines document sitemaps as a discovery input), and AI Bot Policy rises to 15, with 7 for treating training and citation crawlers differently, because an assistant can cite only the pages its search crawler is allowed to read. We rescored every scored site from its saved findings, with no re-crawl. The rankings barely move: rank correlation 0.99 for publishers (8 of the top 10 unchanged) and 0.996 for e-commerce (9 of 10). Medians move from 57 to 60.3 and from 28 to 30, because so few sites have llms.txt (6 scored publishers, 2 scored stores).
Waves 1 and 2 therefore stay on v0.2 exactly as published, and v0.3 applies from Wave 3. The test script and its output are in the research repository (sari_v03_sensitivity.py). Sources: SE Ranking, Ahrefs.
Sample and exclusions
Wave 1 audited 103 publisher sites on 8 May 2026: 87 high confidence (three articles sampled), 6 medium confidence (one or two articles), and 10 low confidence (no article could be sampled). 270 articles were sampled in total. Wave 2 audited 262 e-commerce stores: 160 high confidence, 9 medium and 93 low, scored on product detail pages instead of articles.
Low-confidence sites stay in the published dataset, flagged, so nothing is hidden. They are left out of the mean, median, range and cohort medians. The 10 Wave 1 exclusions are: daringfireball.net, defector.com, espn.com, gizmodo.com, hbr.org, notebookcheck.net, searchengineland.com, theringer.com, time.com, webmd.com.
Version and changelog
Current rubric version: v0.2. Newest first.
| Date | Version | Change |
|---|---|---|
| v0.3 proposed (Wave 3 onwards) | llms.txt tested against two large studies (SE Ranking, 15 October 2026, nearly 300,000 domains: no correlation with AI citations; Ahrefs, 15 June 2026, 137,210 domains: 97% of llms.txt files received no requests). Proposed for Wave 3: llms.txt 10 to 3 points, sitemap 5 to 7, AI Bot Policy 10 to 15 (differentiated training and citation treatment 4 to 7). Rescoring both published waves under these weights keeps the rankings almost unchanged (rank correlation 0.99 and 0.996), so Waves 1 and 2 stay on v0.2 as published. | |
| v0.2 (document revision 2) | Published this methodology page. Reconciled every headline figure to one scored set per wave (Wave 1: 93 scored of 103 audited, mean 49.9, median 57.0). Earlier pages that showed mean 45.8 and median 54.0 had included the 10 low-confidence sites; they now use the scored set. No site was re-audited and no score changed. | |
| v0.2 | Wave 1 (publishers) published: 103 sites audited on 8 May 2026, dataset and CSV released under CC-BY-4.0. | |
| v0.2 | Category 5 reweighted to reward differentiated treatment of training and citation bots. Bot user-agent list updated to match operator documentation. | |
| v0.1 | Initial rubric. |
Cite this
Released under CC-BY-4.0. Reuse with attribution. Quote the scored-set figures and say the sample, for example “mean SARI 49.9 across 93 scored publishers”.
The dataset
Nouriel, M. (2026). SARI Publisher Audit Dataset (Wave 1) [Data set]. Automation Switch. https://automationswitch.com/research/agent-legibility-audit/dataset
This methodology
Nouriel, M. (2026). Scaletific Agent-Readability Index (SARI) methodology, v0.2 [Methodology]. Automation Switch. https://automationswitch.com/research/agent-legibility-audit/methodology
Browse the full dataset, read the publishers deep dive or the e-commerce deep dive, or go back to the research hub. This page lives at /research/agent-legibility-audit/methodology and will keep that address.