Skip to content

Methodology v1

This page describes in full how we measure. The methodology is frozen by edition: whatever is published in an edition was measured with the version documented here at that time, and changes between versions are recorded in the changelog at the bottom. Version v1 · in force since August 2026. This is the English translation of the Spanish original, which prevails in case of doubt.

What we measure

Two different things, with two different instruments:

  1. Technical readiness («agent-readiness»): whether a brand's website can be read and understood by AI assistants and their crawlers. Instrument: the technical scanner (below).
  2. Actual visibility («share of voice»): how often and in what position the assistants mention each brand when answering real consumer questions. Instrument: answer sampling (below).

The census

Each sector has a public census of brands and domains. Inclusion criterion: operating in Spain in that sector, with an active website of its own. The census grows continuously and includes any brand the assistants themselves mention that is not yet listed. No brand can pay to get in, get out or change sector. Initial sectors: invoicing software, car insurance and direct-to-consumer mattresses.

Technical scanner: the 0–100 score

Our own crawler audits every website in the census without executing JavaScript and without bypassing any protection. The weights of the score:

DimensionWhat we checkWeight
AI crawler access Whether robots.txt allows access to the home page for GPTBot (8), PerplexityBot (8), OAI-SearchBot (7), ClaudeBot (7) and Google-Extended (5). With no robots.txt, the protocol convention is that access is allowed, and it is scored that way.35
Content readable without JavaScript Characters of visible text the home page serves to a client that does not run JS: ≥ 2,000 → 25 · ≥ 800 → 15 · ≥ 300 → 8 · fewer → 0.25
Structured data JSON-LD (schema.org) present on the home page (10) and with types that are meaningful to an agent (Product, Offer, Organization, LocalBusiness, SoftwareApplication, FAQPage, Service, etc.) (10). Purely decorative types (WebSite, BreadcrumbList…) do not score the second part.20
llms.txtA real /llms.txt file (a 200 that returns HTML does not count).10
SitemapDeclared in robots.txt or present at /sitemap.xml.7
FeedRSS/Atom linked from the home page.3

What the census represents, and what it does not. The census brings together the brands with a recognisable presence in each sector, not a random sample of Spanish e-commerce. Our figures describe the census brands in those sectors and should not be read as «X% of Spanish websites»: sector leaders have technical teams that the average online shop does not, so their results will be better than those of the whole. When we publish a figure about the whole country it will be with a census designed for that, and we will say so.

Reliability of a single pass. Some websites respond intermittently: the same address rejects the scan once and accepts it five minutes later. Every check that fails because of a network cut or a temporary rejection is retried before being counted as negative; a 404, which is a definitive answer, is not retried. And since those rejections come and go, the score we publish is that of the most recent valid scan: if the website responded to us in any pass, that is the measurement that counts. Only websites that have never let us read them appear as «not auditable». We also discard responses that are not the brand's website (anti-bot verification screens and domains parked at the registrar), because scoring them would mean publishing another page's score believing it to be theirs.

Websites that reject the scan. Some websites answer our crawler with a 401, 403 or 429: their firewall rejects clients it does not recognise. Those websites get no score, because we do not score blindly what we have not been able to read, and they appear in the census saying exactly that. We publish the fact because it is informative in itself: a site that rejects clients it cannot identify usually rejects AI crawlers too, even if its robots.txt says otherwise.

Known limitations of v1, published on purpose: (a) the scanner identifies itself as an ordinary browser and does not simulate each AI crawler's user agent, so it measures what is declared in robots.txt plus what is served to a client without JS; (b) it only audits the home page, not the rest of the site; (c) it does not yet assess whether the checkout process is accessible to agents. All three are candidates for v2.

Readiness: Base and Authority

In addition to the census technical score, since 9 September 2026 we measure a website's readiness with two scores out of 100: the Base (what a website must have; finite) and the Authority (reviews, content, links, citations; logarithmic). The full rubric is in the readiness methodology (in Spanish). The public pre-assessment of the Base is available at https://app.nombradia.com/api/preestudio?web=yourdomain.com.

Share-of-voice sampling

The v1 design:

Which assistants we measure, and why those

A visibility index is only as good as its selection of assistants, so we apply two criteria and make them public:

  1. People in Spain use it. We measure the channel through which a Spanish consumer actually asks for recommendations. An assistant with marginal use in Spain adds no signal to the index: it adds noise, and multiplies the cost of each edition.
  2. It is an independent observation. Two assistants that use the same base model underneath are not two opinions: they are the same one counted twice. Including them would artificially inflate that model's weight on the result without it showing in the table. That is why we do not measure products that are a wrapper around a model we already measure, and why we record the exact model that answered each sample.

The second criterion has an uncomfortable consequence that we accept: if it is ever shown that two models we measure share an origin, we will say so and adjust the selection in the next edition, with the change documented in the changelog below.

Niches and difficulty of each front v1 · 4-10-2026

This is what each customer sees in their dashboard, not public data: it feeds no ranking, no seal and no published figure.

Three percentages are calculated from that:

The difficulty comes from these rules, applied in order (the first one that applies wins):

  1. Fewer than 20 answers in the front: not enough data yet.
  2. Fewer than 30% of answers name any business: here the AI explains, it does not recommend anyone.
  3. T greater than 0 and T equal to or greater than L: you are ahead; the goal is to stay there.
  4. H below 15% and the AI names someone (big brands, or there is a bar): very high (it only names big brands). If the only businesses that appear are rivals the customer has set aside or names from a single answer, it is not «very high»: the next rule applies.
  5. L equal to 0: open ground (nobody within reach appears yet).
  6. H below 40% or L − T greater than 50 points: high.
  7. L − T between 20 and 50 points: medium.
  8. L − T below 20 points: low.

Realistic target: to appear as often as the best rival within reach (L). If a big brand the customer is not watching appears more often than them, they are always told, even if they are ahead of the rivals their size. Indicative time frame: low, 1–2 months; medium and open ground, 2–4; high, 4–8; very high, more than 8. It is a hypothesis of this version, which we will review in January 2027 with our own data: assistants that search the web reflect published and indexed changes within weeks, and those that answer from memory, in their next version. It is not a promise, and we do not charge for results.

Setting a rival aside («not my competitor») only changes the customer's targets, never the measurement: the rival stays in the full table and in the verbatim answers, and the date it was set aside is recorded.

Raw data and reproducibility

Every sampled answer is archived in full (question, complete answer, model, date, sources) and every technical scan keeps the original HTTP responses. With each edition we publish the aggregates and the question universe; the complete raw archive is kept permanently and supports any verification or replication.

Changelog