Can an AI agent find you?
130 Romanian domains · 12 machine-verifiable signals · ADO Score 0–100 after the Exista.io framework · 0 Agent Cards · mean score 17 · 8 September 2026
Until now, this series has measured one thing: when a person asks ChatGPT or Gemini "where do I get a good laptop", which brands come back in the answer. The person reads, the person chooses. But a second kind of customer is arriving. An AI agent receives an objective ("find a supplier of corporate jewelry for 200 gifts, under a given budget, with an invoice") and solves it on its own. It does not search Google and does not read marketing copy. It requests configuration files at standardised paths, checks structured data, cross-references sources and, if it cannot find what it needs in a format it can process, moves on to the next candidate. The first customer is persuaded; the second is verified.
This is the fifth episode in the series measuring how Romanian markets appear in front of artificial intelligence. The first three measured the human layer: jewelry, books, electronics. The fourth measured a technical artefact, llms.txt. This one measures the next layer, the agents', using the framework Exista.io published in February 2026 under the name AI Visibility Stack: whether an autonomous agent can find, evaluate and select a Romanian company without a human stepping in. The paper proposes a metric for this, the ADO Score, and makes a prediction: most companies will score near zero, not because they are weak, but because the necessary artefacts are not yet part of standard practice. We tested the prediction on Romania.
Two changes of method from the previous episodes, stated up front. We did not query models; we probed websites: every figure below comes from an HTTP request or a Wikidata query, reproducible with the published script. And we did not build a new sample: we took exactly the brands the AI engines recommended to people in episodes 1–3, plus the marketing agencies from episode 4. That is, the companies that have already won layer 1 and the companies that sell AI visibility. If anyone is ready for agents, it should be them.
In short — what we found
-
Zero Agent Cards. None of the 126 domains with an HTTP response serves
/.well-known/agent-card.json, the file the paper calls "the functional equivalent of a website for agents". Nor on the legacy path,/.well-known/agent.json. The dimension with the largest weight in the ADO Score, 30 points, is empty across the board. - The mean ADO Score is 17 out of 100; the maximum is 35. Nobody leaves the paper's "minimal visibility" band (0–20) by more than a little: 55 domains sit between 21 and 40, 61 between 1 and 20, 9 at exactly zero. Above 40 there is no one.
- All the score comes from old SEO. Trust signals (JSON-LD, Wikidata) and sitemap freshness bring, on average, 12 of the 17 points. Interoperability brings 0.3 of 15. Whoever does well does well for reasons unrelated to agents.
-
The interoperability that exists comes from the platform, not from strategy. 8 domains serve MCP authorisation discovery (
/.well-known/oauth-protected-resource): 3 are Shopify stores (the protocol ships with the platform), 4 run a WordPress plugin (/wp-json/mcp/…) and exactly one built its own server. A single domain out of 130, snsys.ro, an IT firm, wrote a manifest for agents by hand. -
The agencies selling AI visibility do slightly better than their clients (17.9 vs 15.9 points), exclusively through
llms.txt(66% vs 29%) and JSON-LD (80% vs 59%). On Agent Cards, zero as well. - Layer 1 transfers weakly to layer 2. Across the 41 brands, the Spearman correlation between visibility in AI answers (ep. 1–3) and the ADO Score is 0.31. The top 10 brands by AI visibility have a mean ADO of 17.2; the rest, 15.4. eMAG, the most visible brand in the series (94% of answers), scores 22.
- Wikidata is the signal that separates brands from agencies: 12 of 41 brands have an entity with an official website declared; 3 of 85 agencies. But of the 4 domains that have both Wikidata and JSON-LD, only 3 carry the same name in both.
- The closed door: 6 domains answer a JavaScript-free request with 403, 503, 406 or an anti-bot challenge, among them 4 of the 13 electronics retailers the engines recommend (Flanco, PC Garage, Vexio, Quickmobile) and Amazon.de, which additionally blocks all six answer crawlers. An agent-crawler gets exactly what the probe got.
-
A real false positive: libris.ro answers 200 with a generic JSON ("Forbidden!") on any
.well-knownpath. A naive agent would read it as an Agent Card,mcp.jsonandai-plugin.jsonat once. The mandatory-field filter removed it; the lesson stays.
Raw data, free. The three CSV files — scores and every signal per domain, adoption rate per signal and group, statistics per dimension — plus the scoring specification and the probe script are published under CC BY 4.0. Download the data
01 · How did we measure whether an AI agent can find a Romanian company?
Methodology
| Parameter | Value |
|---|---|
| Theoretical framework | AI Visibility Stack / Agent Discovery Optimization — Marco, G. (2026), Beyond AEO: The AI Visibility Stack and the Era of Agent Discovery Optimization, Exista.io Working Paper, doi:10.5281/zenodo.18728629 |
| Metric | ADO Score 0–100, five dimensions with the paper's weights: Agent Card 30 · trust signals 25 · knowledge completeness 20 · interoperability 15 · freshness 10 |
| Operationalisation | Websem — one machine-verifiable criterion per point, published in scor-ado.md
|
| Sampling frame | 130 domains: 43 brands named by AI engines in episodes 1–3 (jewelry, books, electronics and IT) + 87 marketing agencies and sites cited by the engines in episode 4 |
| Probing | 8 September 2026, a single pass, from Romania, ~12 HTTP requests per domain, no JavaScript |
| Signals checked |
/.well-known/agent-card.json and /.well-known/agent.json (A2A) · /.well-known/oauth-protected-resource and /.well-known/mcp.json (MCP) · /.well-known/ai-plugin.json · robots.txt · llms.txt and llms-full.txt · JSON-LD on the homepage · visible text without JavaScript · sitemap.xml (lastmod) · Wikidata (P856 = official website) |
| Instrument |
probe_ado.py, standard Python 3, published with the data; declared User-Agent WebsemADOProbe/1.0
|
| Domains with an HTTP response | 126 of 130 (4 without: 2 timeouts, 1 DNS, 1 invalid certificate) |
| Domains in the ranking | 125 (without websem.ro, the author's domain — reported separately, section 07) |
Four design choices deserve an explanation up front, because they decide what the numbers mean.
We measure what an agent can verify, not what a person can read. The probe does not execute JavaScript, does not interpret marketing copy and does not judge design. It requests files at standardised paths, parses raw JSON and HTML, and queries a public knowledge graph. This is exactly how the Exista.io paper describes a procurement agent in its evaluation phase: "persuasive prose, testimonials, and visual design elements are irrelevant to this process".
The weights are the paper's; the criteria are ours. The paper defines the dimensions and weights of the ADO Score but does not publish scoring criteria. We wrote them, transparently, one binary signal per point, and publish them with the data. This is not the Exista.io instrument and claims no equivalence with it; it is a reproducible measurement of the same dimensions. Anyone can run the probe and get the same figures, on the same day.
A 200 response does not mean "served". Many servers answer 200 with an HTML page on any path, and some answer 200 with a generic JSON error. A file counts only if it parses as JSON and contains the protocol's mandatory fields: for an Agent Card, name and url or skills; for MCP discovery, resource or authorization_servers. Without this filter, one domain in the study would have had an Agent Card, mcp.json and ai-plugin.json simultaneously, all three being the same "Forbidden!" message.
The sampling frame is not the market. The 43 brands are exactly the companies the AI engines recommended to people in previous episodes, the winners of layer 1; the 87 agencies are those the engines cite on marketing and AI topics, i.e. those who sell AI visibility. It is not a representative sample of the Romanian economy. It is, however, the most favourable possible sample for the paper's hypothesis that LLM visibility is inherited by agent discoverability: if it does not hold here, it is unlikely to hold elsewhere.
The main limitation, declared from the start: a single pass, on a single day, with binary signals. A site that publishes its Agent Card the next day does not appear here. The score measures the existence of public artefacts, not the quality of the company.
02 · What do ADO, Agent Card, A2A, MCP and llms.txt mean?
Terminology, once
AI Visibility Stack is the three-layer framework proposed by Exista.io. Layer 1, LLM visibility, is what we measured in episodes 1–3: whether a model names you when a person asks. Layer 2, agent discoverability, is the subject of this study: whether an autonomous agent can find, evaluate and select you without a human in the loop. Layer 3, participation, is whether you yourself have a published agent other agents can transact with; it cannot be measured from the outside without initiating transactions, so it stays out of the study.
ADO, Agent Discovery Optimization, is the discipline of layer 2, and the ADO Score its composite metric, 0–100. The paper's interpretation scale: 0–20 minimal visibility, 21–40 emerging, 41–60 moderate, 61–80 strong, 81–100 dominant.
A2A, Agent2Agent, is the open protocol launched by Google in April 2025 and donated to the Linux Foundation, through which agents talk to each other. Its discovery artefact is the Agent Card: a JSON file at /.well-known/agent-card.json (the current path; the old one was /.well-known/agent.json) declaring the entity's name, description, endpoint, protocol version, capabilities, skills and authentication schemes. The paper calls it "the functional equivalent of a website for agents".
MCP, Model Context Protocol, introduced by Anthropic in 2024, is the protocol through which an agent connects to tools and data. Discovering a protected MCP server goes through /.well-known/oauth-protected-resource (RFC 9728), a JSON declaring the resource and its authorisation servers; /.well-known/mcp.json is an unofficial convention with no standard.
llms.txt is the Markdown file at the site root that gives models a clean map of the content; we measured it in episode 4 on 87 marketing domains, and here we re-read it as an agentic accessibility signal.
JSON-LD / schema.org Organization is the structured markup on the homepage that declares by machine who the entity is: name, URL, sameAs (official profiles), address, telephone, legal identifier. Wikidata is the public knowledge graph behind Wikipedia; an entity whose property P856 (official website) equals the domain is the simplest cross-verifiable identity signal an agent can ask for.
Answer crawlers (OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, Claude-User, Claude-SearchBot) bring content into real-time answers; training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot etc.) collect for future models. The taxonomy is the one from episode 4.
03 · Which Romanian companies are the most discoverable for an AI agent?
The ranking: who can be found
| # | Domain | Group | ADO | Trust | Knowledge | Interop. | Freshness | Agent Card |
|---|---|---|---|---|---|---|---|---|
| 1 | IONA (iona.ro) | brand · jewelry | 35.0 | 13 | 10.0 | 4 | 8 | 0 |
| 2 | TEILOR (teilor.ro) | brand · jewelry | 32.4 | 21 | 11.4 | 0 | 0 | 0 |
| 3 | Litera (litera.ro) | brand · books | 32.4 | 18 | 9.4 | 0 | 5 | 0 |
| 4 | dafe.ro | agency | 31.0 | 13 | 10.0 | 0 | 8 | 0 |
| 5 | difrnt.ro | agency | 31.0 | 13 | 10.0 | 0 | 8 | 0 |
| 6 | privatebrands.ro | agency | 30.8 | 15 | 7.8 | 0 | 8 | 0 |
| 7 | digitaliomarketing.ro | agency | 28.0 | 13 | 10.0 | 0 | 5 | 0 |
| 8 | goai.ro | agency | 28.0 | 13 | 10.0 | 0 | 5 | 0 |
| 9 | seocherry.ro | agency | 28.0 | 13 | 10.0 | 0 | 5 | 0 |
| 10 | creativdigital.ro | agency | 27.9 | 11 | 8.9 | 0 | 8 | 0 |
The rest of the ranking, with all 41 signal columns, is in the main CSV.
Three things are visible in the table before any interpretation. First: the Agent Card column is zero from top to bottom, across all 125 domains, and the interoperability column is non-zero for exactly 9. Second: the scores are compressed into a narrow band. Between rank 1 and rank 20 there are 9 points; between rank 20 and rank 100, another 15. On the paper's scale, 6 domains enter "emerging visibility" with more than 30 points and none reaches "moderate". Third: the top three are brands, not agencies, but for different reasons. IONA is first because it collects every small signal available to a well-maintained online store: JSON-LD with a legal identifier, llms.txt, a fresh sitemap and, alone in the top 3, MCP discovery, courtesy of a WordPress plugin. TEILOR and Litera are there because they have what almost nobody has: a Wikidata entity with a declared official website and an identical name between Wikidata and JSON-LD, the consistency signal the paper places at the base of machine trust.
The mean score over the 125 domains is 17.2, the median 19.7. The nine domains at zero (Amazon.de, Cel.ro, KULTHO, Librăria Română, Vexio and four agencies) carry no signal the probe could credit: no organisation JSON-LD, no Wikidata, no dated sitemap, no files for agents. It does not mean the companies do not exist or do not work; it means that, for a visitor who is not human, there is no machine-readable evidence of who they are.
04 · Where are the points lost: which ADO Score dimensions are empty?
The anatomy of the score
| Dimension | Weight | Mean (125 domains) | Domains with points |
|---|---|---|---|
| Agent Card | 30 | 0.00 | 0 |
| Trust signals | 25 | 8.19 | 102 |
| Knowledge completeness | 20 | 4.99 | 101 |
| Interoperability | 15 | 0.27 | 9 |
| Freshness | 10 | 3.76 | 89 |
| ADO Score | 100 | 17.22 | 116 |
Read by dimension, the figure of 17 tells a simple story: 45 of the 100 possible points, the two dimensions that explicitly concern agents, Agent Card and interoperability, produce on average 0.27. Every other point comes from three things a site has had for ten years for SEO reasons: organisation markup, an entity in a knowledge graph and a dated sitemap. The Exista.io paper builds the ADO Score so that layer 1 "feeds" layer 2; in our data, layer 1 is all there is.
Trust signals are the most populated dimension, with 102 domains above zero, but the mean of 8.2 out of 25 shows how thin the layer is. 92 of 126 domains have Organization-type JSON-LD on the homepage (73%), but with on average 5.4 of the 9 attributes checked: name and URL almost always, logo and sameAs often, address and telephone less so, a legal identifier on just 15 domains. Consistency between the declared URL and the actual domain is good (83 of 91). Name consistency with Wikidata can only be checked on 4 domains, because only 4 have both sources, and passes on 3.
Knowledge completeness has a structural ceiling: half the points come from Wikidata, and only 15 domains out of 126 have a Wikidata entity whose official website is the domain. The other 111 can take at most 10 points of 20, however good their JSON-LD. The 15 entities have between 3 and 9 of the 10 attributes checked; eMAG has all 9, Humanitas, Cărturești and TEILOR 7 each.
Freshness is the only dimension where the agencies clearly beat the brands (4.2 vs 2.8 out of 10): 65 domains have a lastmod within the last 30 days in their sitemap, another 20 within 180, and 30 earn extra points for an llms.txt with a Last-Modified from the last 90 days. It is a cheap dimension and explains part of the agencies' advantage in section 07.
05 · Which agent signals actually exist on Romanian websites?
Adoption, signal by signal
| Signal | All (n=126) | Brands (n=41) | Agencies (n=85) |
|---|---|---|---|
A2A Agent Card (/.well-known/agent-card.json) |
0 (0%) | 0 | 0 |
Agent Card, legacy path (/.well-known/agent.json) |
0 (0%) | 0 | 0 |
MCP discovery (oauth-protected-resource) |
8 (6.3%) | 3 (7.3%) | 5 (5.9%) |
mcp.json |
1 (0.8%) | 0 | 1 |
ai-plugin.json (legacy OpenAI) |
2 (1.6%) | 0 | 2 |
| Wikidata entity with official website = domain | 15 (11.9%) | 12 (29.3%) | 3 (3.5%) |
| Romanian Wikipedia article | 10 (7.9%) | 9 (22.0%) | 1 (1.2%) |
| JSON-LD Organization on the homepage | 92 (73.0%) | 24 (58.5%) | 68 (80.0%) |
llms.txt served |
68 (54.0%) | 12 (29.3%) | 56 (65.9%) |
llms.txt spec-compliant (H1 + summary) |
43 (34.1%) | 8 (19.5%) | 35 (41.2%) |
Sitemap with lastmod ≤ 30 days |
65 (51.6%) | 20 (48.8%) | 45 (52.9%) |
| ≥ 200 words without JavaScript | 111 (88.1%) | 34 (82.9%) | 77 (90.6%) |
| No answer crawler blocked | 117 (92.9%) | 38 (92.7%) | 79 (92.9%) |
robots.txt accessible |
119 (94.4%) | 39 (95.1%) | 80 (94.1%) |
| HTTPS | 126 (100%) | 41 | 85 |
The table has a clear shape: the older and more "SEO" a signal is, the more adopted it is; the newer and more "agent" it is, the more absent. HTTPS, robots.txt and server-side content are above 88%. JSON-LD is at 73%. llms.txt, born in 2024, is at 54%, but with a brutal asymmetry between those who sell it (66%) and those it is sold to (29%). Wikidata, which is not new but requires curation work off-site, is at 12%. And the two files defined by the agent protocols, from 2025, are at 6% and 0%.
It is worth underlining that blocking answer crawlers is almost non-existent: 2 domains out of 126, Amazon.de (all six) and one agency (only ChatGPT-User). Blocking training crawlers is somewhat more frequent, 13 domains, with KULTHO and canopy.ro blocking eight each, but it does not affect the answer layer and is not penalised in the score. The Romanian market is not hiding from AI; it simply has not prepared anything for it.
Content without JavaScript deserves a note as well: 6 domains have under 200 words in their raw HTML, four of them under 25, i.e. homepages that exist only after a script runs. Three domains explicitly display the "enable JavaScript" warning. For a person with a browser they are normal sites; for an agent-crawler they are empty pages.
06 · Where does the interoperability that exists come from?
The platform, not the strategy
The 8 domains serving /.well-known/oauth-protected-resource, the file through which an MCP client learns where to authorise, are the only place in the study where agent protocols appear in production. We read the files, not just the response code, and their origin is unambiguous.
| Domain | Group | What the file declares | Origin |
|---|---|---|---|
| moogu.ro | brand · jewelry | authorization_servers: shopify.com/authentication/… |
Shopify |
| istyle.ro | brand · electronics | authorization_servers: cont.istyle.ro, shopify.com/authentication/… |
Shopify |
| gregoire.ro | agency | authorization_servers: account.gregoire.ro, shopify.com/authentication/… |
Shopify |
| iona.ro | brand · jewelry | resource: …/wp-json/mcp/mcp-oauth-server |
WordPress plugin |
| limitless.ro | agency | resource: …/wp-json/mcp/mcp-oauth-server |
WordPress plugin |
| arobsgrup.ro | agency | resource: …/wp-json/mcp/mcp-oauth-server |
WordPress plugin |
| growwwise.com | agency | resource: …/wp-json/mcp/mcp-oauth-server |
WordPress plugin |
| seo365.ro | agency |
resource: seo365.ro/mcp, authorisation on Supabase |
own server |
Three stores run on Shopify, and Shopify has shipped MCP discovery to every store as platform infrastructure; the merchants did nothing and, most likely, do not know the file exists. Four sites run a WordPress plugin exposing an MCP server at /wp-json/mcp/; that is an installation decision, not an architecture one. A single domain out of 130, seo365.ro, built its own endpoint with its own authorisation.
It is the same pattern we found in episode 4 with llms.txt, where 21% of the files were plugin-generated and none of those followed the specification. The agentic interoperability of the Romanian market, such as it is, is a side effect of the platforms the sites run on. That has a practical consequence: if an agent started tomorrow to prefer suppliers with MCP discovery, the advantage would go not to the companies with a strategy, but to those that happened to be on Shopify.
Two cases fall outside the pattern and deserve naming. snsys.ro, an IT services firm in Bucharest, is the only domain with a hand-written mcp.json: name, version, description, vendor, contact with sameAs to Wikidata and LinkedIn, capabilities by topic and geography, resources (llms.txt, llms-full.txt, a company.json), certifications with issuer and registry code. It is not a standard, but it is exactly the kind of document the paper describes under "semantic clarity": precise terminology, explicit client segment, verifiable certifications. The paradox: the same server answers 406 on the homepage to a non-browser client, so the probe could read its manifest but not its site. The second case, difrnt.ro and again snsys.ro, serve ai-plugin.json, the ChatGPT plugin format retired by OpenAI in 2024. It is an artefact of the previous wave, published in good faith for a protocol that no longer exists. It does not score, but it says something about how fast "what you need to have" changes.
07 · Are the agencies selling AI visibility more discoverable than their clients?
Brands vs agencies
| Brands (ep. 1–3) | Agencies (ep. 4) | |
|---|---|---|
| Domains in the ranking | 41 | 84 |
| Mean ADO Score | 15.9 | 17.9 |
| Median ADO Score | 16.3 | 19.9 |
| Maximum | 35.0 (IONA) | 31.0 (dafe.ro, difrnt.ro) |
| At zero | 5 | 4 |
| Above 20 points | 15 (37%) | 40 (48%) |
| Agentic accessibility index (mean) | 72 | 83 |
The agencies hold a small advantage, fully explained by two cheap signals: llms.txt (66% vs 29%) and organisation JSON-LD (80% vs 59%), both delivered by the WordPress themes and plugins they use. On the expensive signals, those requiring a presence off-site, the brands lead clearly: 12 brands have Wikidata, 3 agencies do; 9 brands have a Romanian Wikipedia article, a single agency does. And on the dimensions that concern agents, both groups are at zero, save for the plugins of section 06.
In other words, those who sell AI visibility have done, on their own sites, the easy part of the homework and nothing of the hard part. It is not an accusation; it is a measure of how new the subject is. But for a buyer of AEO services, the question "what is your agency's ADO Score?" becomes, from September 2026, one that can be answered with a number.
By vertical, among the brands, luxury jewelry has the highest mean (19.0 over 10 brands), followed by the book market (15.8 over 21) and electronics (12.8 over the 10 with a response). The gap in electronics is produced largely by section 09: four of the retailers could not be probed at all.
Websem.ro, the author's domain, was in the sampling frame because the engines cited it in episode 4. We probed it with the same script and report it separately, with no place in the ranking: ADO Score 28.0 (trust 13, knowledge 10, interoperability 0, freshness 5, Agent Card 0), accessibility index 100. It would have ranked 7th. No Wikidata, no Agent Card, no MCP: the same gaps as the rest of the market. The study is not written from a position of advantage.
08 · Are the brands AI recommends to people also the ones an agent can find?
Layer 1 and layer 2
The Exista.io paper states that the layers are cumulative: "a company with strong Layer 1 visibility inherits credibility signals that support Layer 2 discoverability", and leaves the empirical check to "future research". We had, from episodes 1–3, exactly the layer-1 visibility of the 41 probed brands: the share of AI answers that named them. We set it against the ADO Score.
The Spearman correlation is 0.31 (n = 41). Positive, weak, far from "inheritance". The top ten brands by AI visibility have a mean ADO of 17.2; the other 31, 15.4. eMAG, named in 94% of the electronics answers and holding the most complete Wikidata entity in the study (9 of 10 attributes), scores 22, because it has no organisation JSON-LD on its homepage and no file for agents. TEILOR, at 60% in jewelry, scores 32.4. Flanco, PC Garage and Vexio, between 50 and 55% AI visibility, score 16, 18 and 0, because they refuse the connection of a client that is not a browser.
The figure must be read carefully in both directions. It does not say the paper is wrong: the mechanism it describes, LLM citation as a credibility signal for agents, may exist without showing in the public artefacts we measure. It does say that, in September 2026, in Romania, being recommended by ChatGPT to people does not come bundled with being findable by an agent. They are two separate subjects, with two separate to-do lists, and nobody has started the second one.
09 · What happens when an agent knocks and nobody opens?
Inaccessible sites and false positives
The probe requested each homepage first with its own declared User-Agent, then, on refusal, with a browser one. Six domains refused both: Flanco, PC Garage and Vexio with 403, Quickmobile with 503, snsys.ro with 406, and Amazon.de with 202 and an anti-bot challenge page. Four more did not answer at all (Altex and Media Galaxy with timeouts, start-seo.ro with no DNS, daredigital.ro with an invalid certificate). In total, ten of 130 domains cannot be read by an ordinary HTTP client, and seven of them are electronics retailers, precisely the market episode 3 found the most concentrated in AI answers: of 13 stores recommended by the engines, only 6 could be read in full.
The consequence for the score is mechanical: a blocked site cannot receive points for JSON-LD, content or freshness, however much of them it has behind the firewall. Flanco and PC Garage stay in the ranking only through Wikidata. The consequence for agents is worse than the score: the paper describes the specified behaviour of an agent that receives no answer, "to proceed to the next candidate". Anti-bot protections are legitimate and have good reasons; but they do not tell a scraper from a buyer with a budget, and in the absence of an Agent Card saying "come in this way", the agent leaves.
The opposite case is just as instructive. libris.ro answers 200 with valid JSON on any .well-known path: {"servlet":"default","message":"Forbidden!","url":"/.well-known/agent-card.json","status":"200"}. Code 200, parsable JSON, a message of denial. An agent that checks only for the file's existence would conclude that Libris has an Agent Card, mcp.json and ai-plugin.json. The probe required the protocol's mandatory fields and rejected all three. It is not a mistake by Libris, but a property of the real web: as agents start looking for files at standard paths, servers configured to answer "something" to any request will produce false identities. Verifying the schema, not the response code, is the first rule of layer 2.
10 · What are the limitations of this study?
Limitations
We list them in full, because a study that does not declare its limitations is not a study.
- A single pass, on a single day. Everything above describes 8 September 2026. A site publishing an Agent Card on 9 September does not appear. There is no volatility section, because there is no second measurement.
-
The operationalisation of the score is ours. The Exista.io paper gives the dimensions and weights, not the criteria. Another set of criteria, with the same weights, would produce other absolute values; the order and, above all, the zeros would remain. The criteria are published in full in
scor-ado.md, precisely so they can be contested point by point. -
Binary signals, no quality. A JSON-LD with name and URL and one with twelve verified attributes both take the 6 base points; the difference shows only in completeness. A 200-byte
llms.txtand a 60 KB one are "served" alike. - What we could not probe: Google Knowledge Graph (keyed API, non-deterministic results), certifications and accreditations (no machine-queryable public registry in Romania), address or telephone consistency across sources (no second machine source outside Wikidata). All three are in the paper; none is in the score.
- Wikidata depends on search. An entity is credited only if it has the "official website" property equal to the domain and if it was found through the Wikidata search API. An entity without a declared official website cannot be machine-linked to the domain and is not credited; that is a design choice, not an oversight, but it may underestimate large brands. The Litera entity was found in the first pass and kept after verification; the method is noted in the raw data.
- The sampling frame is not representative. 43 brands from three verticals and 87 marketing sites. The figures are about them, not about "Romanian companies" in general. One of the brands, Amazon.de, is not Romanian; it was kept because the engines named it in ep. 3 and is reported as such.
- Ten domains without a read. Four did not answer, six blocked the request. They are in the CSV with their status and with what could be measured (Wikidata, robots.txt where it answered), but their scores are, by construction, incomplete.
- Layer-1 visibility comes from different episodes, with different denominators (10, 180 and 87 answers), and is compared here as a plain percentage. For eMAG, present in two episodes, the higher value was used. The correlation in section 08 is indicative.
- The probe's declared User-Agent may be treated differently from a real agent. We retried with a browser User-Agent on every refusal and recorded the result; no domain answered differently.
- Declared conflict of interest: websem.ro is in the frame and is reported, but excluded from the ranking and the means. The author sells services on exactly the subject measured.
11 · What does this mean for a company in Romania?
Implications
Three implications that follow directly from the data, without exceeding it.
Layer 2 is empty, therefore it is cheap. In a market where nobody has an Agent Card, the first to publish a correct one competes with no one. The file is a JSON of a few dozen lines at a standard path; the A2A specification is public; the cost is measured in hours, not months. The same goes for MCP discovery, which three platforms already deliver for free. The difference between 17 and 50 points, i.e. between "minimal" and "moderate" on the paper's scale, is a list of four files and a Wikidata entry.
Machine identity is the real problem, not the files. Of 126 domains, 15 can be machine-linked to an entity in a knowledge graph, and 3 carry a consistent name between the graph and their own site. An agent cross-checking "who is this company" has, for 111 domains, no second source. Wikidata is public, free and editable by anyone, with notability rules that leave room for most of the companies in this study. It is the biggest lever in the signal table and the least used.
Anti-bot protection needs to learn to tell agents from scrapers. Six domains, four of them large electronics retailers, are invisible to any client that is not a browser. Until the protocols have identification mechanisms accepted by firewalls, the only way to remain accessible to an agent is to declare explicitly where it may enter: an Agent Card, llms.txt, a robots.txt with rules for answer crawlers. Otherwise the "let it in or not" decision is made by a WAF configured in 2019.
12 · Frequently asked questions
What is the ADO Score?
A 0–100 metric proposed by Exista.io (Marco, 2026) for a company's discoverability by autonomous AI agents, from five dimensions: Agent Card (30), trust signals (25), knowledge completeness (20), interoperability (15), freshness (10). This study keeps the dimensions and weights and adds published, machine-verifiable scoring criteria.
Is this the official Exista.io score?
No. It is Websem's operationalisation of their framework, with its own transparent criteria. The paper does not publish scoring criteria, so any ADO Score computed by anyone else is, likewise, an operationalisation.
Why does no domain have an Agent Card?
Because the A2A protocol is a year and a half old, and the file is not generated by any popular website platform, unlike llms.txt or MCP discovery. The Exista.io paper predicted exactly this outcome for February 2026; our data confirms it for September, in Romania.
What is "MCP discovery" and why does it matter that it comes from Shopify?
/.well-known/oauth-protected-resource is the file through which an MCP client learns where to authorise in order to talk to a site's server. Shopify publishes it for every store; the merchant decided nothing. It matters because it shows that agentic interoperability, such as it exists, is an effect of the platform, not of strategy.
Why is websem.ro in the study?
Because the AI engines cited it in episode 4, and this study's sampling frame is everything the engines cited. We probed it with the same script and report it with its score (28.0), but it enters neither the ranking nor the means.
How were the 130 domains chosen?
We did not choose them. They are the brands named by the AI engines in episodes 1–3 and the domains cited as sources in episode 4, deduplicated. A single domain appears in two episodes (emag.ro).
Why did Altex and Media Galaxy not answer?
Their servers sent no response within 20 seconds, twice, at different hours. They remain in the data with the Wikidata entity found and without the other signals. It is a limitation of the probe, not a statement about their sites.
Does a low score mean the company is weak?
No. It means the public artefacts an agent could verify do not exist. The Exista.io paper says the same thing explicitly. The largest companies in the study, eMAG or Amazon, are not at the top.
Why does Wikidata weigh so much?
Because it is the only public, free, machine-queryable source through which an agent can cross-verify a company's identity against its own website. The paper names it explicitly as a trust signal. Google Knowledge Graph would be the second, but cannot be probed without a key.
How can any figure in the study be verified?
With probe_ado.py and analiza_ado.py, published in the repository. The first runs the probe on domenii.csv, the second computes the scores from the raw JSON and checks programmatically that the sum of the dimensions equals the score, on every row. The run takes under an hour.
What should a company do to pass 40 points?
The data shows where the unclaimed points are: a complete A2A Agent Card (up to 30), a Wikidata entity with an official website and basic attributes (8 + up to 10), an organisation JSON-LD with address, contact, legal identifier and sameAs to Wikidata (up to 15 cumulated), a dated sitemap. The study does not measure the effect of these interventions; it measures the state of affairs.
Is the data reusable?
Yes. The three CSVs, the scripts and the scoring specification are published under CC BY 4.0, with attribution.
The data, deposited publicly
Responsibility note
This study measures exclusively the existence of public, machine-verifiable artefacts on 130 domains, on a single day. It does not measure the quality of the companies, their products, services or sales. "X is the most discoverable" does not mean "X is the best", and a score of zero means the signals are missing, not that the company is weak. The domains were named or cited by AI engines in previous episodes, not chosen by the authors. The ADO Score is Websem's operationalisation of a framework published by Exista.io and does not represent Exista.io's instrument or assessment. The figures can be re-verified from the raw data and with the published scripts, and any reported error will be corrected publicly, with versioning.
Original Websem study · Dan Cristian Alexandrescu · September 2026
Raw data: CC BY 4.0 — DOI 10.5281/zenodo.22664431
Top comments (1)
Probing the sites instead of querying the models is the right methodological call — engine answers are a moving target and partly circular with a sample built from engine recommendations. The trade-off is you measure readiness, not actual agent traffic, so my main question is: in your own access logs, do you see requests hitting /.well-known/agent-card.json or the MCP discovery paths? If agents already crawl, the study measures who is ready; if they don't, it measures who is early. Both are useful, but they imply different actions.
The actionable number hiding in your data: the Agent Card dimension is worth 30 of 100 points, zero domains have it, and your observed maximum is 35. One static JSON file at a well-known path is the cheapest 30-point move available to anyone in the sample, and it alone would top the distribution you measured. That asymmetry — 30 points for one file versus 15 for the whole interoperability dimension — is worth spelling out for the agencies reading this, because "put one JSON file on your domain" is a sellable engagement in a way "restructure your data layer" never will be.
The platform-versus-strategy split — 8 domains with OAuth-protected-resource discovery, 7 of them inheriting it from Shopify or a WordPress plugin — matches what I see from the agent side: discovery artifacts spread when platforms ship them, not when consultants recommend them. Which raises the uncomfortable question for the series: does ADO adoption in Romania just wait for Shopify and WordPress to ship Agent Card support, and is the honest timeline "when the platforms move" rather than "when agencies educate their clients"?