This research brief separates what publishers observed, what respondents said, and what we recommend testing. Reviewed November 7, 2026. These are external findings, not measured outcomes from Outcome Driven or its clients. Four sources have a dated release or update within the research window; LiftArchive identifies its version only as November 2026, so its exact first-publication day is unconfirmed.

1. Shopping agents need a measurable completion path

2026 Peak Season Benchmark — Quantum Metric, November 5, 2026. The report combines a 1,500-person U.S./U.K. survey with platform behavior. It reports 48 abandoned retail carts per completed agent purchase, versus about three for retail traffic overall. Assisted sessions converted about 30 times as often as fully agentic sessions. Those ratios are not abandonment percentages or a comparison with human-only conversion.

Limit: platform business/session counts are not fully disclosed, agent classification is proprietary, and intent differs between cohorts. This is observational vendor research.

Test: instrument product access → valid quote → cart → authorized handoff → completed order. Diagnose legitimate blocked requests without removing fraud controls. Measure completion per eligible task, error type, refunds and net margin.

Editorial angle: Your Website Has a Second Customer: The Shopping Agent. The thesis is that delegated shopping introduces a distinct task journey whose failure points need their own measurement.

2. Published experiment wins are a selected sample

A/B Test Win Rate: What Share of 1,824 Real Tests Won? — LiftArchive, November 2026 version. Of 1,824 publicly shared results from 442 sources, 83% were described as wins, 8% as losses and 9% as flat. Its comparison with complete experiment portfolios illustrates publication bias; the organizations and testing methods are not a controlled comparison.

Limit: this is a curated public corpus, not raw experiment validation or a representative base rate. Definitions, statistical methods and designs vary.

Test: record every launched experiment, including flat results, with a predefined primary metric, uncertainty, duration, allocation and profit guardrails. Review incremental contribution margin and decisions changed.

Editorial angle: Why Every CRO Case Study Looks Like a Winner. The thesis is that selective reporting distorts expectations and hides the value of rejecting an ineffective change.

3. Better conversion among visitors can coexist with fewer customers

Conversion Rate Benchmarks for 2026 — First Page Sage, updated November 2, 2026. The AI-versus-search client panel covers 221 sites, 12 industries and over 3.8 billion sessions. Reported stage rates imply 10.3 customers per 1,000 ChatGPT visits versus 2.6 for Google organic. The publisher calculates these figures by multiplying stage rates; they should not be presented as pooled observed customer counts.

Limit: observational client selection, differing intent and conversion definitions, and no demonstration of incremental demand. Some component cohorts had invested in GEO. Large session totals do not remove selection bias.

Test: report identifiable referrals, qualified customers and retained margin alongside total discovery volume; keep unknown attribution visible. Measure absolute qualified demand and use an eligible holdout where feasible.

Editorial angle: AI Traffic Converts Better. That Doesn’t Mean It Creates More Customers. The thesis separates visitor quality, channel scale and causal lift.

4. Permission is part of the buying experience

2026 AI Holiday Shopping Survey — Talkdesk / Centiment, November 4, 2026. The survey includes 1,048 U.S. shoppers and 312 retail leaders at AI-using businesses. It reports willingness to let an agent purchase using payment details of 15% with prior review and 10% without. These are stated preferences, not a checkout A/B test.

Limit: self-report, restricted retail sample and no observed effect of adding review controls.

Test: make recommend, prepare and execute distinct permissions. Evaluate a clear review screen, editable details and human recovery. Measure authorized completion, unwanted actions, corrections and repeat purchase; never test by withdrawing required consent.

Editorial angle: Agentic Commerce Needs a Review Screen Before It Needs More AI. The thesis is that understandable control can make delegation usable; its conversion effect still needs measurement.

5. Staying is not the same as choosing

Financial Life Moments 2026 National Study — CSI / Center for Generational Kinetics, November 4, 2026. In a weighted online survey of 1,000 U.S. account holders, 51% said switching difficulty was their main reason for staying; 40% had considered leaving after feeling unsupported during a life event. The publisher reports a ±3.1-point margin of error.

Limit: stated reasons and intentions do not establish observed retention effects. Weighting does not by itself establish a probability sample. Banking has unusually high switching costs, limiting transfer to other businesses.

Test: compare a consented, context-specific onboarding or renewal sequence with normal communication using a holdout. Measure task success, voluntary repeat behavior, opt-outs and 90-day retention.

Editorial angle: Inertia Isn’t Loyalty. The thesis is that retention should be read alongside active preference and successful customer outcomes.

An experiment credibility checklist

  1. What decision, population and primary metric were specified before launch?
  2. Was assignment randomized? What changed together?
  3. How many eligible people and outcomes were observed, for how long?
  4. What uncertainty and stopping rule support the decision?
  5. Did revenue, margin, returns or customer effort worsen?
  6. Were segments planned beforehand or discovered afterward?
  7. Where are the flat and losing tests in the same portfolio?
  8. What can this evidence justify changing in our business?

Our strongest editorial opportunity remains the CRO credibility checklist: it helps readers judge any vendor case study without pretending that another company’s win rate predicts their next test. The companion agent series applies the same discipline to a new interface: a working API is an operational result; profitable, wanted customer outcomes require further evidence.

Use the Outcome Experiment Canvas to specify one test and its decision rule.

Continue with the companion guides

Published November 7, 2026. Recommendations and scenarios are editorial proposals, not measured portfolio or client outcomes.