Methodology
How to read the Behavioral Responsivity Framework
One page that defines what we claim, how we classify our evidence, and the vocabulary our articles reuse. If a claim in any article points here, this is what it points to.
Google is the evidence, not the subject
This framework is not an attempt to reverse-engineer Google. It is an account of the behavioral constraints any retrieval system faces when it tries to infer utility from how people behave. Google, Microsoft, and other search engines are valuable case studies only because they operate at enough scale to expose those constraints publicly — through patents, testimony, and engineering disclosures.
The behavioral principles come first. The implementations follow from them. A patent can be retracted and a system renamed without touching the underlying claim: that satisfaction is latent, and behavior is the least-bad proxy a ranking system has for it.
How we classify evidence
Every claim we publish is tagged by the kind of evidence behind it. The point is to make the boundary visible — so you can always see where established research ends and our own synthesis begins. Very few frameworks in this field draw that line; it is the one we care about most.
Peer-reviewed behavioral science and information-retrieval findings — Joachims, Nisbett & Wilson, Scharkow, Pirolli & Card, Chapelle & Zhang, and equivalents.
Engineering intent documented in granted patents. Evidence of how a system was designed to work — not proof of current deployment.
Real-world implementation: court testimony, engineering disclosures, documented API behavior (e.g. US v. Google, 2023; the Content Warehouse API leak, 2024).
Conclusions derived from the evidence above but not directly documented by any external source — the framework's own predictions, labeled so you can tell them apart.
The Behavioral Signal Hierarchy
Implicit signals sort into four tiers, ordered by how hard each is to fake. This ordering reflects inferential reliability and manipulation resistance — not confirmed ranking weight, which is unknown and varies by query. Articles reference these tiers rather than re-defining them.
Tier 1 — Click
visible · manipulation cost: Low
Which result the user selects. The most visible signal, and the easiest to fake.
Tier 2 — Engagement
attention · manipulation cost: Medium
What happens after the click — dwell, scroll, session-level attention. Harder to fake at scale.
Tier 3 — Reformulation
verdict · manipulation cost: High
Whether the result resolved the need. Requires a full, failed session to manufacture.
Tier 4 — Longitudinal
durability · manipulation cost: Very high
Whether utility persisted after the task — return visits, branded demand. Must produce real recurring value.
↓ Top to bottom, manipulation cost rises — and with it, how far the signal can be trusted.
The full development of the hierarchy — with evidence for each tier — lives in The Behavioral Signal Hierarchy.
The objections we welcome
Sophisticated readers push back — and they should. We answer in a consistent register: concede the literal point, then sharpen to the actual claim.
“Patents don’t prove implementation.”
Correct — and we don’t use them as proof of deployment. We use patents as evidence of engineering intent and architectural possibility. The behavioral science doesn’t depend on any patent; the patents illustrate it.
“Correlation doesn’t imply causation.”
Agreed. We don’t claim behavioral signals cause rankings. We claim retrieval systems face a measurement problem — satisfaction is latent — and behavioral signals are the least-bad available proxies. Whether a given system uses a given signal causally is a separate question from whether that signal carries information about utility.
That is the whole apparatus: a clear claim, four classes of evidence, and a vocabulary that compounds across the cluster. The articles apply it.
Read the articles →