Three bets,
worked through

Pinterest said in 2018 that Related Pins from search used the incoming query and the clicked Pin. That does not establish current refinement enforcement. In one desktop walkthrough on 2026-10-12, the pin page cleared the visible search box and related results drifted from the query. Back restored the query and Products filter.

Carry stated refinements across surfaces and visits, visibly and editably.

First test keeping the query and refinements visible and editable on the pin page. Check Back behaviour across paths before proposing a repair; it worked in this walkthrough.

The central run assumes 30% baseline carry-over and a 75% chance that the feature takes its beneficial branch. That second input is a proxy for successful, still-useful persistence, not a measured preference rate. Under the mechanism method, decisions reached on Pinterest rise 6.8% [3.5, 13.3]. Under the research-priors method they move about zero [-0.71%, +0.85%]. The methods encode different assumptions about how persistence helps. These sources provide no isolated causal estimate that resolves the disagreement.

Holding the beneficial-branch input at 75%, the native model changes sign between 70% and 80% baseline carry-over. A version that always carries choices changes sign between 80% and 90%. The native treatment sets carry-over to a sampled value between 80% and 97%. At high baseline carry-over, it can therefore reduce carry-over. Its crossing includes that regression. The always-carry version removes it. Holding baseline carry-over at 30%, the sign changes between 25% and 50% on the beneficial-branch input. All three are unresolved within numerical noise. These are sensitivity results, not cutoffs for dropping the bet.

A stale choice carried into a new task makes results worse. People need to see and clear it.

Pinterest said query-aware Related Pins from search improved engagement 20% over the previous system (2018 talk abstract; metric not defined). Baymard found 27% of shop sites broke Back around filters (2020) and 33% of desktop and 42% of mobile sites cleared the search box (2021): how common the problem is on sites, not the effect of fixing it. Pinterest's skin-tone ranges, Google's shopping preferences and eBay's saved searches show editable persistence is an existing pattern. None of the sources reviewed tests the whole results, pin, related, next-visit chain.

Uptake of a visible control, or long-run return visits.

In the observed desktop interface, Products was inside the filter panel. Pinterest said its 2026 Closeup systems gate shopping candidates and modules. Sponsored tiles appeared on 25 of 26 searches in the one-account walkthrough of 2026-10-12.

Test improved shopping eligibility against the current policy, holding ad load fixed. Treat an explicit switch as a separate hypothesis about user control and uptake.

For the switch as first specified (match intent and show fewer ads), at 15% assumed uptake, decisions rise 0.6% [-0.26, +1.42] and fixed-yield revenue falls 0.35% [-0.70, +0.02]. Matching alone: decisions +0.22% [-0.79, +1.31], revenue +0.19% [+0.11, +0.45]. Fewer ads alone: decisions +0.11% [-0.31, +0.69], revenue -0.53% [-0.97, -0.31]. Components interact, so these are not additive shares.

The switch bundle's median fixed-yield revenue crosses our -0.5% guard between 15% and 30% assumed uptake; individual worlds cross at different settings. In the bundle, no tested uptake meets both the +1% decisions bar and the guard on median; four of 24 worlds do at 15% uptake, before noise screening. Matching alone meets both on median at 100% assumed uptake, a sensitivity setting, not an adoption estimate. Matching alone never breaches the guard in any tested world.

In this model, the revenue cost comes from cutting ads. Test matching at today's ad load first. If ad reduction is also under consideration, cross the two changes in a factorial experiment. Automatic gating and an explicit switch need separate tests.

Pinterest said its best online group cut shopping-generator triggers by about 40%, with sitewide Success Curation volume +0.26%; a separate aggressive group cut triggers by up to 85%. Neither figure measures ad impressions (2026 paper). Pinterest said global module-save volume rose 11.2% (95% CI 6.4% to 16.0%) and UCAN sitewide saves rose 1.0%, with module coverage about 15% (2026 workshop paper). Google's 2015 paper describes cutting mobile search ad load 50% over a sequence of launches and measuring long-run user learning. eBay and Google's AI Overviews ads condition commercial slots on quality or intent.

Whether people would use an explicit switch (uptake is assumed).

Pinterest said it changes link placement based on destination quality and restricts spam distribution. Those sources report no fulfilment rate or causal fulfilment lift.

A declared audit collected on 2026-10-18 (one logged-in account, desktop web, 30 declared queries, 360 organic linked pins; strict rule: no further detour, and the exact item shown can be bought now): each group contained 120 sampled pins. Instructions delivered on 82 of 119 judged destinations (0.689), with one unknown. Default shopping delivered on 56 of 89 judged destinations (0.629), with 31 unknown. Products-filter shopping delivered on 69 of 78 judged destinations (0.885), with 42 unknown. If every unknown failed or delivered, the respective bounds are 0.683 to 0.692, 0.467 to 0.725 and 0.575 to 0.925. These bounds are conditional on the destinations as recorded. Some product-identifying URL parameters were not restored, leaving the exact destination or variant unresolved. Per-query shares among judged destinations ranged from 0.125 to 1.0 for instructions, 0 to 1.0 for default shopping, and 0.5 to 1.0 with Products. Many unknowns are product links whose exact variant could not be confirmed. Common failures: idea listicles with no steps, video-only tutorials and generic pages for instructions; idea pages instead of products in default shopping search; one-click-short landing grids and out-of-stock items under the Products filter. Some failures reflect our strict rubric. It is an audit proxy: the engine samples after a useful step, including sponsored clicks; the audit samples organic linked pins equally by query. Two checks since: a blind second judge (an AI, not a person) agreed on 90% of instruction pins both could judge (kappa 0.75); and a repeat run on the same account a day later moved every share by four points or less (instructions 0.65, default shopping 0.59, Products filter 0.88), though about two thirds of the pins were the same.

Among judged destinations, the instruction share is near the engine's middle input of 0.67, and default shopping is below its middle input of 0.75. This comparison motivates checking the shopping assumption. It does not measure or reject that parameter: the audit and engine sample different events, and many shopping destinations remain unresolved.

Screen or demote results whose destination doesn't deliver, versus improving relevance ranking.

For the native shopping screen D01, at assumed baseline delivery 0.65, decisions rise 2.6% [1.2, 4.7] with no useful-supply loss. They fall 5.9% [-7.4, -2.7] at 80% useful-supply retention. With no supply loss, its paired gap against the exposure-matched relevance intervention is -2.1 points [-5.3, +0.5]. These are global effects from shopping-targeted treatments; the comparison does not establish which improvement is cheaper or more achievable.

For the native shopping screen, relevance wins on median across the tested baseline-delivery values, even with no useful-supply loss. Instruction results differ: with no supply loss, the native instruction fix wins on median at baseline delivery 0.35; its ranking crossing between 0.35 and 0.50 is unresolved. Shopping supply-ranking crossings appear between 95% and 100% retention in individual worlds; instruction crossings also appear at lower retention. None of these clears the noise rule.

The audit suggests testing whether destination type matches the task (a steps page for a how-to query, a product page for a product query), alongside destination screening. The audit does not establish which intervention would create more value.

Pinterest's link-quality and spam enforcement (no outcome published). Google said its March 2024 update cut low-quality, unoriginal content in results by 45%. Airbnb researchers reported 3.7% fewer support-needed bookings at a selected ranking weight, without negatively impacting overall booking conversions. Baymard says 30% of users are likely to leave when the item they want can't be bought.

Long-run trust (the engine has no trust state yet), or how publishers respond.