Typed Decision

Most SEO advice arrives as a sentence in a document. A sentence cannot be sorted, filtered, assigned, approved or checked against a policy, and it carries its evidence only as far as its author remembered to write it down. A typed decision is the same advice as an object with the same fields every time, so all of those things can be done with it.

The word “typed” has an ordinary meaning in data work. JSON Schema describes itself as a declarative language for defining structure and constraints for JSON data, with which the structure, constraints and data types of data are described, and a validator checks that data conforms^[1]. A typed decision is typed in the same plain sense: a fixed shape that every instance has to fit. QueryRev is not described here as built on that standard, which is named only to explain the idea.

Evidence has to be traceable. The W3C’s PROV documents describe provenance as information about the entities, activities and people involved in producing a piece of data, which can be used to form assessments about its quality, reliability or trustworthiness^[2]. The evidence of a decision works the same way: each line leads back to the rows it came from, so a person can judge it before approving.

The fixed set is nine types: Update, Expand, Create, Merge, Protect, Internal Link, Technical, Conversion and Earn Citation, in five named families. A fixed set lets a queue be filtered by kind and lets a policy say which kinds need whose approval. Each decision carries search and business evidence, technical state and history; a value range with its kind, never a single guess; effort, risk and time to impact; and its place in one prioritised queue. Confidence is shown as product confidence only once it has been calibrated against verified outcomes.

A fixed set of states is also how QueryRev’s own study keeps its data. In the study, 1,339 of the 6,294 sites whose robots.txt could be read, 21.3%, name at least one AI crawler^[3]. For each crawler the study records the rule as one of a few fixed states, such as blocked or partly blocked, and not as a sentence, so that the counts add up and can be read again next month.

By completeness. Count the share of decisions with every field filled: type, target, evidence, value range with its kind, effort, risk and priority. A decision with no evidence is incomplete, whatever its text says.

By traceability. Take a sample of decisions and follow each line of evidence back to its row. Every claim should lead to a figure, and every figure to its source. A line that cannot be followed back is a sentence, not evidence.

By outcome, once the change is made. Set the verified result beside the value range the decision gave, by type. Whether results land inside their ranges is the check on the estimates, and it is how confidence can be calibrated over time.

Illustrative: an invented site, to show the method.

A software company has two setup guides that compete for one query and trade places between positions 5 and 11. An SEO writes in a ticket: “merge the two setup guides, should help.” That sentence cannot be ranked against anything else.

Stored as a typed decision, the same advice has a type, Merge, and a target. Its evidence is a set of known figures that lead to the rows they come from. Its value is a modeled range of $1.1k to $1.7k a month. It has an effort, a risk and a place in the queue. A person can approve, edit, assign, snooze or reject it, and nothing changes on the site while it waits.