The problem this addresses

Policy audiences increasingly get their first answer from a language model, not a search results page. When an AI answers a question about tax reform or education outcomes, the institutions it names — and the ones it leaves out — quietly decide who is treated as the authority. Traditional web analytics stop at the click and cannot see any of this. Meridian measures it directly.

How the measurement works

The starting point is a set of 168 policy questions, 12for each of the OECD's fourteen substantive directorates, from taxation and education to AI governance and development co-operation. The directorate structure is used deliberately as the sampling frame, so the measurement reflects the OECD's actual portfolio rather than an arbitrary list of topics. Each question is put to a language model exactly as a policymaker or journalist might ask it, with no mention of any institution. The answer then goes through a second, controlled analysis pass (run at temperature 0 for consistency) that extracts, in structured form, which of the four tracked institutions appear, in what order they are first mentioned, and whether each mention is backed by a specific named output — “PISA”, say, or “the BEPS framework” — rather than a passing reference. Every answer becomes a small record of who was cited, how prominently, and how concretely.

How those records become a score

From this raw evidence, each institution receives an OECD AI Visibility Score built from four components, each capturing a different dimension of visibility. Presence Rate asks how often the institution appears at all. Position Weight rewards being named first, since the first institution cited anchors the answer. Citation Depth distinguishes a specific, attributable reference from a generic name-drop, because a cited PISA figure carries more communicative weight than the bare word “OECD”. Share of Voice measures presence relative to all four tracked bodies combined. The weights reflect a deliberate judgement: appearing at all matters most, prominence and concreteness next, and competitive share provides context. Every weight is documented in the code and adjustable; nothing in the score is hidden.

What the data showed

Measured this way, the OECD is the most visible institution overall: it holds 51% of all institutional mentions (27 of 53) and a composite score of 22.3 — 2.0× UN's 11.4, the next most-visible body. But the visibility is highly uneven, and the pattern is the finding.

The OECD dominates where it owns a single, well-branded product: taxation (63.1), financial & enterprise (51.9), education (35.0) and AI governance (35.0) — BEPS and the global minimum tax, corporate governance standards, the AI Principles, PISA. In three directorates, however, the OECD scores exactly zero: employment & social, health and entrepreneurship & regions. The answers there speak in generic terms or cite national bodies, and never name the OECD — even though the OECD publishes actively in all three, through the Employment Outlook, Health at a Glance and the SME and Entrepreneurship Outlook. An independent regex cross-check over all 168 answers confirms this is genuine, not a measurement artifact: it found no institution mention that the extraction step had missed.

Why the pattern looks like this

The data points to a clear interpretation: AI visibility tracks brand concentration, not institutional effort. Where the OECD's work is consolidated under one recognizable name, the model reaches for it. Where the contribution is spread across many publications without a dominant brand, or another institution holds the stronger topical association, the work is present in the world but absent from the answer. That is the core communications insight this tool surfaces: in AI-mediated channels the OECD's risk is not the quality of its work but the discoverability of it — and that discoverability is uneven in a way traditional metrics would never reveal.

How the OECD could act on this

Each finding points to a concrete response, and the zero-visibility directorates are the priority. The second half of this tool addresses exactly that: the GEO audit checks whether a given OECD page is structured to be read and cited by AI systems — quotable passages, structured data, an llms.txt index, and whether AI crawlers are being served or blocked. Testing surfaced a tangible example: the main OECD.org site returned an automated-access error (HTTP 403) to a standard request, itself a discoverability risk worth investigating. Beyond page-level fixes, the measurement would be strengthened for production use by running across several models rather than one, repeating runs on a schedule so the score becomes a monitored trend rather than a snapshot, and testing question phrasing systematically to separate genuine invisibility from artifacts of wording.

How this maps to the role

This project is a working instance of the responsibilities of the Junior AI & Communications Intelligence Officer role (COM/CISC). Analysing OECD visibility in AI-mediated environments — LLM referrals, generative search, bot-driven traffic — is the visibility score itself. Developing metrics and proxies for OECD presence in AI systems is the four-component OAVS. Exploratory work on Generative Engine Optimisation and AI discoverability is the GEO audit module, and monitoring LLM bot activity and content access is its crawler and robots.txt analysis. Combining traditional communications metrics with AI-era indicators into new methodologies and reporting frameworks is what the per-directorate view does; the dashboard is a communications performance reporting tool by construction. And the scheduled, automated collection — each run appending a comparable snapshot — is the continuous improvement of analytical processes the role calls for. The tool is deliberately transparent end to end, because measurement that informs communications strategy has to be defensible to the teams who act on it.