Sooner or later, in a meeting about which vocabulary to adopt, someone puts a number on the screen. The vocabulary is used on a vast number of web pages, supported by every major tool, downloaded more often than all the alternatives together. The number tends to end the discussion, and it is easy to see why: after months of debate about definitions, here at last is something objective.
The question the number answers, though, is a different one from the question the meeting was asked. The meeting needed to know which vocabulary would let this organisation exchange data with its own partners. The number says how many organisations, somewhere, use the vocabulary for something.
The cobra bounty, again
In the first part of this series I told the story of the cobra bounty in colonial Delhi: a reward for dead cobras that led people to breed them, and a city with more cobras once the reward was withdrawn and the breeders released their stock. Goodhart’s law, in Marilyn Strathern’s phrasing, gives the mechanism: “When a measure becomes a target, it ceases to be a good measure.”
Each of the four approaches our field uses to pursue semantic interoperability rewards something countable in place of the outcome. The mandate counted compliance; bilateral alignment counted connections. The open-ended critical-mass model counts adopters. Whichever vocabulary attracts the most users becomes the default, and everyone else is expected to follow.
An adoption count measures reach. A governed ecosystem needs evidence of convergence among the particular parties who exchange data, and a single global number has no way to provide it.
A count without names
The trouble lies in what an adoption count records. It is one number. It carries no information about who the adopters are, what they use the vocabulary for, or whether any of them matter for your exchange.
Picture a logistics company that must exchange shipment data with five partners. A vocabulary used on millions of websites tells it nothing useful if those five partners describe shipments with something else. For this decision, the large number is empty. The opposite also happens: a modest vocabulary with a few dozen adopters may be exactly the one the five partners run, and the global ranking will place it far down the list.
Influence and merit produce the same number
A count also cannot distinguish why adoption happened. A vocabulary can spread because it models its domain well, or because the body that publishes it has the standing, the marketing or the tooling to make adoption easy. Both produce the same number. This is the difference between authority, the standing of whoever publishes a vocabulary, and authoritativeness, the evidence that the parties relevant to a scope actually build on it. The critical-mass model reads the second from a figure that mostly reflects the first.
Popularity breeds its own cobras
Once the count becomes the target, behaviour adjusts to it. Publishers compete for adopters, and the easiest adopter to win is one who takes a handful of terms and models everything else locally. Adopters, for their part, gain from being seen to use the popular vocabulary, so they declare it, use the parts that are easy, and extend it privately where their real needs begin.
The count rises. Underneath it, divergence rises too: many parties share a thin layer of terms and each has built its own model for the rest. When the incentive that drove adoption shifts, because a search feature changes or a campaign ends, the thin adopters have little holding them to the shared layer, and the ecosystem is left with more local variants than it had before.
schema.org is the clearest case of what this model can achieve, and of where it stops. Launched in June 2011 by Bing, Google and Yahoo, with Yandex joining later that year, it reached a large part of the web because adopting it changed how pages appeared in search results. That achievement is real. It also sets the ceiling of the approach: the adopters are unidentified, the scope is unbounded, and no particular set of parties can point to the deployments as evidence that they, specifically, have converged.
Counting adoption well
Adoption remains one of the most valuable signals a programme can have, once it is bounded twice. The first bound is the party set: the identified participants of your ecosystem, such as the members of a consortium, the agencies of an administration, the business units of an enterprise or the participants of a data space. The second is the representation scope: the questions your models must be able to answer, written as competency questions.
Within those bounds, adoption means live deployments by named parties, for the purpose in hand. The question changes from “how many use this vocabulary anywhere?” to “how many of the parties I exchange data with run this design in production, for this purpose?”. An answer such as “four of the six banks in our consortium use it for clearing” is evidence a modeller can act on. It also sits naturally next to two other signals, which parties import the vocabulary and which map their own terms to it, so that a strong adoption figure resting on a thin structural commitment becomes visible.
Five questions to ask of an adoption figure
The next time a number settles a modelling discussion, these questions will show how much of it applies to you.
- Who are the adopters? If the figure cannot name them, it measures reach.
- Which of them are parties we exchange data with?
- For what purpose do they use it, and does that purpose match the exchange we need?
- How much of the vocabulary do they use, and how much have they extended locally?
- What would the figure look like if we counted only our partners, in production?
Keeping the count, changing what it counts
Popularity carries real information, and a vocabulary with wide reach often deserves serious consideration. The change is to stop letting the global number decide on behalf of a particular ecosystem. Counting adoption among identified parties, for a bounded purpose, alongside the other signals of convergence, is the approach behind the framework we are developing at METALEVELS.
Next, and last, in this series: the approach that has produced the strongest convergence of all, and the price it asks before anyone can join.
If you would like to know what adoption looks like among your own partners, start a conversation with us.