Choice Without Measurement Is Theatre
2026-06-07 · 4 min read · Janaina Maia
Giving people an opt-out sounds responsible. But if they cannot see the consequences of using it, the control is mostly theatre.
Search Engine Journal reported that Google is starting to let some UK websites opt out of AI search features without losing standard search ranking. The change follows a UK Competition and Markets Authority requirement that publishers should be able to withhold content from AI search features and AI model training. Google is also testing Search Console controls for AI Overviews, AI Mode, and AI Overviews in Discover, and says the opt-out will not be used as a normal ranking signal.
That is a meaningful shift. But the more interesting design problem is what is still missing: the measurement layer that lets people make the decision intelligently.
A control is not enough if the user is blind.
Google’s new AI performance reporting shows impressions — when a publisher’s content appears in AI features. What it does not yet show is the click data that would tell publishers whether those appearances actually send people back to the site. The CMA’s notes call for click-throughs, click-through rate, and data separated from normal organic search. Without that, publishers can see exposure, but not value.
This is the kind of gap product teams create when they treat control as the end of the design work. A toggle may technically give choice, but choice without feedback is weak. It asks the user to make a consequential decision with only half the instrument panel visible.
The enterprise version of this will matter even more.
This is not only a search and publishing story. The same pattern will show up inside enterprise AI products. Teams will ask users to opt in or out of agent access, generated recommendations, automated summaries, synthetic data, broad context, training use, or workflow automation. The interface may offer a switch, a policy, or a consent banner. But the user will still need to understand what happens when they use it.
If a design leader can turn off AI suggestions in a high-risk workflow, what does that change? Does quality improve or slow down? Are fewer errors introduced? Does review burden drop? Are decisions better documented? Are teams losing useful context or avoiding bad automation? If the product cannot answer those questions, the control is performative.
Measurement is part of trust.
I think AI governance is often discussed as if it lives in policy documents, admin settings, and legal language. But governance also lives in feedback loops. People need to see enough evidence to decide whether a control is working for them.
That means product teams should design controls and measurement together. If we let people restrict AI context, show what changed in answer quality and source coverage. If we let admins require human review, show review volume, correction patterns, and escaped errors. If we let teams opt out of model training, show what data is excluded and what functionality depends on it. If we let users suppress AI-generated material, show whether the experience becomes more useful or just emptier.
Design implications.
- Pair every AI control with evidence: users should understand the practical effect of the setting, not just its label.
- Separate visibility from value: impressions, mentions, or appearances are not the same as usefulness, traffic, quality, or trust.
- Show before-and-after impact: help teams compare what changes when AI is enabled, narrowed, reviewed, or turned off.
- Make trade-offs explicit: a good control explains what the user gains, what they lose, and what remains unknown.
- Design for negotiation: in enterprise settings, measurement gives users and organisations leverage to set better boundaries.
My take.
The next stage of AI product maturity is not more toggles. It is controls that are backed by understandable evidence.
Google’s AI search opt-out is a useful step, especially because it suggests regulators are starting to insist that participation in AI systems should not be all-or-nothing. But the missing click data is the real lesson. A choice becomes meaningful when people can evaluate it.
If an AI product gives me a control but hides the impact, it has not designed agency. It has designed plausible deniability.