Four cards, two operators, one structural question running through both: what does measurement look like when the signal you relied on disappears? Kevin Indig named the measurement vacuum in marketing. Simon Willison named one place in creative software where human judgment is not optional.
Attribution was always the wrong tool for organic, and AI search makes the gap catastrophic
Kevin Indig's September 28 essay "The Collapse of Attribution" opens with a claim most growth teams treat as settled: attribution works. Indig's argument is that it never really did, not for organic channels.
Attribution was designed for ad spend allocation. Its apparent usefulness for organic channels was always an illusion driven by click data. carries the structural diagnosis. Attribution was built to allocate advertising budgets. It assumes a closed-loop, intent-driven path where a user sees an ad, clicks it, and converts. Organic discovery is not that. A reader finds a piece of content, returns three weeks later from a direct URL, and converts without a referral event that any dashboard can capture. Clicks gave the appearance of a causal chain for organic. Indig's collaborator George Bonaci, VP Growth at Ramp, names the consequence: attribution has become "a crutch replacing critical thinking." The technology changes removing clicks are not breaking measurement. They are exposing a measurement gap that was always structural.
AI search can underattribute marketing channel contribution by 10x because the discovery path no longer generates a referral signal. extends the argument to AI search specifically. When an AI system cites a brand in a generated answer, no referral signal reaches the brand's analytics. The user may act on the recommendation by typing a URL directly or searching by brand name. Neither generates a trace back to the AI citation. Indig cites research putting the underattribution factor at approximately 10x. A brand well-positioned in AI search may see its influence on decisions grow while its attribution dashboards record zero change.
The third card, Three metrics with non-overlapping blind spots, triangulated together, produce a more reliable channel-contribution estimate than any attribution model., is the replacement framework. Indig proposes triangulating three metric types with non-overlapping blind spots: exposure metrics (share of voice, brand mention frequency), behavioral signals (branded search volume, direct traffic), and causal experiments (geo holdouts, randomized control groups). No single leg is accurate. Triangulating them produces a confidence range. Where all three point the same direction, confidence is high.
- Kevin Indig, Attribution was designed for ad spend allocation. Its apparent usefulness for organic channels was always an illusion driven by click data.. Attribution was designed for paid channels; organic use was always an illusion driven by click data that is now gone.
- Kevin Indig, AI search can underattribute marketing channel contribution by 10x because the discovery path no longer generates a referral signal.. AI search removes the referral signal entirely, underattributing channel contribution by approximately 10x.
- Kevin Indig, Three metrics with non-overlapping blind spots, triangulated together, produce a more reliable channel-contribution estimate than any attribution model.. Combining three metrics with different blind spots replaces attribution when referral signals collapse.
The one game loop a coding agent cannot design
Simon Willison's September 27 year-in-review of 2026 includes one observation about game development that sits outside his usual engineering focus. The observation is precise enough to stand as its own card.
Agents can build functional game interfaces but cannot independently engineer engaging gameplay loops, because fun is a subjective judgment that no specification can substitute for. states the boundary: coding agents can build functional game interfaces, but they cannot independently engineer engaging gameplay loops. The reason is structural. Agents verify output against stated criteria. Gameplay quality is not verifiable against any specification because fun is a subjective, iterative judgment discovered only through human play. An agent produces code that runs. A human player discovers whether running the code is worth repeating. The agent cannot initiate that discovery cycle; it can only execute changes once a human has named what needs changing. Willison's observation connects to his broader argument about direction as the residual human skill: once agents handle the mechanical work, the work that remains is judgment about experience that specifications cannot capture.
- Simon Willison, Agents can build functional game interfaces but cannot independently engineer engaging gameplay loops, because fun is a subjective judgment that no specification can substitute for.. Agents build game interfaces; humans must still engineer the gameplay loop because fun cannot be specified or verified without play.