Agent discoverability of Romanian brands and agencies: an ADO Score measurement of 130 domains, September 2026

Le résumé fourni par la source

Raw data from an original study measuring the discoverability of 130 Romanian domains for autonomous AI agents, using the ADO Score (Agent Discovery Optimization) framework proposed by Exista.io (Marco, 2026, doi:10.5281/zenodo.18728629) with a published, machine-verifiable operationalisation by Websem. The sampling frame is the 43 brands named by AI engines in the series' episodes 1-3 (luxury jewelry, books, electronics) plus the 87 marketing agencies and sites cited by the engines in episode 4. Each domain was probed on 8 September 2026 with a standard-library Python script (no JavaScript) for twelve signals: A2A Agent Cards on the current and legacy paths, MCP discovery (RFC 9728 oauth-protected-resource, mcp.json), ai-plugin.json, robots.txt and AI-crawler blocking, llms.txt, homepage JSON-LD Organization markup, server-side text, sitemap freshness and a Wikidata entity whose official website equals the domain. Results: no domain serves an Agent Card; mean ADO Score 17.2 (median 19.7, max 35); the two agent-specific dimensions (Agent Card 30 and interoperability 15) average 0.27 points; the 8 domains with MCP discovery obtained it from Shopify (3) or a WordPress plugin (4), with one self-built server; Spearman correlation between layer-1 AI visibility and ADO Score across the 41 brands is 0.31.Method. On 8 September 2026 a standard-library Python probe (published) requested, for each of 130 domains, the homepage without JavaScript, robots.txt, llms.txt and llms-full.txt, sitemap.xml, the A2A Agent Card paths (/.well-known/agent-card.json and /.well-known/agent.json), the MCP discovery paths (/.well-known/oauth-protected-resource, /.well-known/mcp.json) and /.well-known/ai-plugin.json, and queried Wikidata for an entity whose official website (P856) equals the domain. A file counts as served only if it parses as JSON and carries the protocol's mandatory fields; an HTML or generic-JSON answer at a well-known path does not count. The ADO Score keeps the Exista.io weights (30/25/20/15/10) with one binary, machine-verifiable criterion per point, documented in scor-ado.md. Domains were not chosen by the authors: they are the brands the AI engines named in episodes 1-3 and the domains the engines cited in episode 4. The author's domain (websem.ro) is reported but excluded from the ranking and the means.Key findings. No domain out of 126 with an HTTP response serves an A2A Agent Card, on either the current or the legacy path. The 30-point dimension is empty across the board. The mean ADO Score is 17.2/100 (median 19.7, maximum 35.0, IONA). 55 domains score 21-40, 61 score 1-20, 9 score exactly 0; none exceeds 40. The two agent-specific dimensions (Agent Card, interoperability) average 0.27 of 45 possible points; trust signals (8.2/25) and freshness (3.8/10) supply almost the whole score. Interoperability comes from the platform: of the 8 domains with MCP discovery, 3 are Shopify stores, 4 run a WordPress plugin and 1 built its own server. One domain (snsys.ro) publishes a hand-written mcp.json. Agencies selling AI visibility score slightly higher than the brands (17.9 vs 15.9), entirely through llms.txt (66% vs 29%) and JSON-LD (80% vs 59%); brands lead on Wikidata (29% vs 4%). Layer 1 transfers weakly to layer 2: Spearman rho = 0.31 between AI-answer visibility (episodes 1-3) and ADO Score across 41 brands; eMAG, at 94% AI visibility, scores 22. Ten of 130 domains cannot be read by a non-browser client (six blocked, four without response), seven of them electronics retailers; libris.ro returns 200 with a generic 'Forbidden!' JSON on every .well-known path, a real false positive. Contents. Three CSV files (one row per domain with the ADO Score and its five dimensions plus every raw signal; adoption rate per signal and group; score statistics per dimension and group), four charts as SVG, the scoring specification, the probe script (stdlib Python) and a data dictionary.Denominators. Signal adoption rates use 126 as denominator (domains with an HTTP response; 130 in the frame). Score statistics use 125 (the 126 minus websem.ro, the author's domain, reported separately). Per-group rates use 41 brands and 85 agencies for signals, 41 and 84 for scores. Domains that answered 403/503/406/202 are counted as responding and remain in the ranking with the signals that could be measured; the four domains with no HTTP response are listed but not scored.Limits. A single pass on a single day (8 September 2026); no volatility measurement. The scoring criteria are Websem's operationalisation of the Exista.io dimensions and weights, not the Exista.io instrument. Binary signals: presence, not quality, is scored. Google Knowledge Graph, certifications and cross-source address/phone consistency are in the framework but could not be probed. Wikidata credit requires an entity with the official-website property equal to the domain, found through the Wikidata search API. The 130-domain frame is not representative of the Romanian economy; one domain (amazon.de) is not Romanian. Ten domains could not be read (six blocked, four without response); their scores are incomplete by construction. Layer-1 visibility values come from three episodes with different denominators and are compared as plain percentages. The author's domain is in the frame (reported, excluded from ranking and means); the author sells services on the measured subject.Responsibility note. All brands named in this dataset were named by the AI engines, not by the authors. The material is market research and does not evaluate the quality, services or commercial performance of any brand. Trademarks belong to their rightful holders and are used for identification within a factual analysis. Corrections: office@websem.roCanonical version of the analysis: https://websem.ro/resurse/aeo/studiu-ado-agenti-ai-romaniaThis is episode 5 of an ongoing series measuring how Romanian markets appear in AI answers. Episode 1 covered luxury jewelry (DOI 10.5281/zenodo.21724399). Episode 2 covered the book market (DOI 10.5281/zenodo.21736211). Episode 3 covered the electronics and IT market (DOI 10.5281/zenodo.21821185). Episode 4 covered llms.txt adoption among Romanian marketing domains (DOI 10.5281/zenodo.21861572).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.