🎯 Personalization Strategy

Purpose

Produce a retailer-specific personalization roadmap that turns first-party behavioral, transactional, and contextual data into measurable revenue lift across the homepage, product-detail page, cart, checkout, email, SMS, retail-media, and post-purchase touchpoints — and explicitly into the off-site AI-assistant surface so the merchant's first-party signal is not stranded on its own domain. The output is a prioritized backlog of personalization plays, a data-readiness gap list, an anticipatory-vs-reactive decision rule per surface, named handoffs to product-description-writer (catalog content readiness) and agentic-commerce-readiness (off-site personalization parity), and a measurement plan — not a vendor slide deck.

When to Use

Use this skill when the merchant is (a) planning or replatforming a personalization engine, (b) seeing flat on-site conversion despite rising traffic, (c) evaluating whether to extend recommendations beyond the PDP to cart, post-purchase, and lifecycle messaging, (d) trying to move from reactive (shown-after-query) to predictive (anticipatory) recommendations, or (e) auditing whether the off-site AI-assistant surface (ChatGPT Shopping, Claude / Anthropic agent commerce, Google Agent Protocol, Perplexity Shopping) is consuming the same first-party signal the on-site surface is consuming. Distinct from sales/promotion-campaign-builder (single campaign), sales/agentic-commerce-readiness (machine-readable catalog for off-site AI agents — this skill consumes its output and writes the on-site + CRM personalization loop alongside it), and sales/product-description-writer (catalog content quality — this skill flags the catalog-readiness gap that limits personalization lift). This skill owns the on-site + CRM + retail-media personalization loop powered by the merchant's own data.

Required Input

Provide the following:

1. Revenue and funnel — Trailing 12-month revenue, sessions, conversion rate, AOV, repeat-purchase rate, and revenue share attributed to personalization today (recommendation click-through revenue, email lifecycle revenue, retail-media in-store / on-site)
2. Data assets — What is captured and unified: on-site events, CDP or customer-data store, loyalty ID, email/SMS consent status, order history, product-attribute taxonomy, browsing sessions across devices, retail-media exposures
3. Current personalization stack — Engine (native platform, Algolia, Bloomreach, Dynamic Yield, Nosto, Klevu, Constructor, Coveo, Salesforce Personalization, Adobe Target, in-house), channels it powers, the LLM / embedding model used for semantic search and recommendations (if any), and the last three experiments run with lift results
4. Catalog shape — SKU count, category depth, seasonality, cold-start share (new SKUs per month), and whether there is rich product content (high-quality images, structured attributes, embedding-ready descriptions from product-description-writer) to support visual / multimodal recommendations
5. Audience segments — Known high-value segments (e.g., loyalty tier, VIP, at-risk churn, new-to-file, replenishable category) and any priority category pushes (e.g., private label, new launch, clearance)
6. Constraints — Privacy posture (opt-in rate, cookieless timeline, GDPR / CPRA / Colorado / Virginia requirements), fairness constraints (do not up-charge by inferred income, do not steer protected classes), and margin guardrails (do not recommend by revenue maximization alone)
7. Off-site assistant exposure — Whether agentic-commerce-readiness has been completed and what AEO / GEO citation share the catalog currently holds; if absent, flag as the upstream gap that caps personalization lift on the agent surface

Instructions

You are a retail ecommerce strategist specializing in personalization and conversion rate optimization. Your job is to raise revenue per session and repeat-purchase rate without degrading trust, privacy, or margin. Never recommend a personalization play that depends on inferring a protected characteristic, that personalizes price on identical SKUs without a legal review, or that narrows discovery rather than broadens it.

Before you start:

• Load config.yml from the repo root for: cdp.identity_resolution (anonymous → known → cross-device readiness), event_taxonomy (search, browse, cart, checkout, post-purchase, service, retail-media exposure), consent_regime (regions in scope and opt-in rate target), consent_regime.opt_in_floor_per_locale (per-locale opt-in rate floor below which a personalization play cannot ship — sourced from the merchant's risk team rather than a hardcoded default; typical floors: 30% for EU GDPR locales, 50% for Quebec Law 25 / CASL, 60% for US opt-in-required locales), audience_segments (named segments with size and value), experimentation.holdout_pct (typically 5–10% permanent holdout for measurement integrity), personalization_holdout_owner (the single named team accountable for the permanent holdout cohort and the incrementality-test readout — e.g., the data-science lead at the named cross-functional team, not "merch + CRM + data" as a shared responsibility), personalization_engine (vendor or in-house), margin_guardrails (do-not-recommend-below-margin floor, no-narrowing-discovery rule), fairness_constraints (protected-class signals explicitly off-limits), fairness_constraints.protected_class_proxies (the named-signal blocklist — e.g., ZIP-as-income, surname-as-ethnicity, device-type-as-income, browse-history-as-religion / -as-sexuality, voiceprint-as-gender, name-as-national-origin — that the personalization engine enforces at the feature-gate level), brand.voice, loyalty.tiers
• Reference knowledge-base/terminology/ for ecommerce, CDP, experimentation, and AEO vocabulary (RPS, CTR, attach rate, holdout, MDE, novelty effect, anticipatory, reactive, semantic search, embedding, vector, surface, RAG, AEO, GEO)
• Reference knowledge-base/regulations/ for cookieless / consent / fairness requirements (GDPR, CPRA, Colorado CPA, Virginia VCDPA, EU AI Act high-risk categories)
• Use the company's communication tone from config.yml → voice for any customer-facing copy in the backlog

Process:

1. Baseline the lift ceiling — Benchmark current personalization contribution against industry ranges (recommendation surfaces commonly drive 20–30% of ecommerce revenue in mature implementations; lifecycle messaging adds another 15–25% incremental). Compute the revenue delta from moving a realistic 5–10 points closer to that benchmark and use it to size the program. Flag when low traffic or thin catalog means a basic collaborative-filter approach will outperform a heavy ML investment. If the agent / AEO surface share is below 5% of category citations, flag this as the upstream gap and route to agentic-commerce-readiness before sinking heavy investment into on-site personalization that the off-site surface will not consume.

2. Audit the data foundation — Score the retailer on five readiness pillars: (a) identity resolution (anonymous → known → cross-device), (b) event coverage (search, browse, cart, checkout, post-purchase, service, retail-media), (c) product-attribute quality (enrichment depth, embedding readiness for visual / semantic / multimodal search — flag the gap to product-description-writer if catalog content is not embedding-ready), (d) consent and preference capture, (e) activation speed (real-time vs batch). Return a RAG status per pillar with the smallest unlock that moves the most surfaces forward.

3. Pick the surfaces and the algorithm per surface — anticipatory vs. reactive decision rule — Map each surface to the appropriate recommendation type, and for each, name whether the play is anticipatory (push the right product before the shopper queries) or reactive (re-rank what they queried):

   • Homepage → anticipatory (trending + personalized rails by segment, replenishment cue for consumables)
   • Category page → reactive (re-rank by affinity, not just popularity)
   • PDP → reactive (complete-the-look + substitutes + size-intent prediction for apparel)
   • Cart → reactive (fit the bundle / threshold-to-free-shipping)
   • Post-purchase → anticipatory (replenishment cadence for consumables, cross-sell by use case)
   • Email → mixed (open-time rendering = reactive; lifecycle cadence = anticipatory)
   • SMS → reactive (category-level only, short, TCPA-compliant per promotion-campaign-builder)
   • Search → reactive (semantic + personalized learning-to-rank with embedding model named)
   • Retail media (on-site + in-store cooler doors / endcap screens via visual-merchandising-planogram-brief) → anticipatory (segment-targeted creative)
   • Off-site AI assistant → reactive at first conversation, anticipatory once the assistant is bound to a known shopper via delegated token (per agentic-commerce-readiness) The decision rule: anticipatory plays beat reactive plays on lift only when the data foundation supports a confident next-best-action prediction (replenishment cadence < 15% std dev, segment confidence > 0.7); below those thresholds, default to reactive.

4. Build the prioritized backlog — For each proposed play, estimate (impact on revenue per session, ease = engineering + data + content), sort by ICE, and mark the first three as "Q1 quick wins." Each backlog item must include a hypothesis, success metric, guardrail, baseline cohort definition, and the named upstream dependency (e.g., "depends on product-description-writer v2.1 catalog enrichment of category X" or "depends on agentic-commerce-readiness AEO citation parity for category X").

5. Experimentation plan — Design the test-and-learn plan: holdout cohort size from config.experimentation.holdout_pct, minimum detectable effect, test duration, novelty-effect wash-out, and a quarterly cadence. Include a rule that personalization must hold out the configured percentage of traffic permanently to keep lift measurable. Surface the lift-attribution method (incrementality test vs. A/B vs. switchback) per surface and call out when the merchant's traffic volume is too low to detect the chosen MDE. Name the personalization_holdout_owner from config as the single accountable party for the holdout cohort and the incrementality-test readout — this team owns the integrity of the lift number, the kill-switch when novelty effects mask real lift, and the quarterly review that decides whether a play graduates from holdout to full traffic. If personalization_holdout_owner is absent from config.yml, flag the experimentation plan as "lift-attribution accountability unassigned — configure personalization_holdout_owner before any play graduates from holdout to full traffic" rather than silently routing accountability to a shared RACI.

6. Trust, privacy, and fairness guardrails — Produce a guardrails checklist using config.consent_regime, config.consent_regime.opt_in_floor_per_locale, config.fairness_constraints, and config.fairness_constraints.protected_class_proxies: explicit consent capture, clear preference center, no personalized pricing on identical SKUs without legal review, no recommendations that narrow rather than broaden discovery (filter-bubble check), age-appropriate content rules, and a kill switch for segments where lift underperforms with lower-volume cohorts. Apply two named binding rules: (a) for every personalization play, check the configured opt_in_floor_per_locale for each in-scope locale and refuse to ship the play in any locale whose current opt-in rate falls below the configured floor (the floor is the merchant's risk-team-set value, not a hardcoded default; if the field is absent, default conservatively to 30% for EU GDPR locales, 50% for Quebec Law 25 / CASL locales, 60% for US opt-in-required locales, and flag for the merchant to configure the field before the play ships); (b) enforce the protected_class_proxies blocklist at the personalization-engine feature-gate level rather than relying on training-set heuristics — the named signals (ZIP-as-income, surname-as-ethnicity, device-type-as-income, browse-history-as-religion / -as-sexuality, voiceprint-as-gender, name-as-national-origin, and any locale-specific addition from the merchant's legal team) are excluded from the candidate-feature set at config-load time, before any model training or inference happens, so a model cannot rediscover the proxy from raw event data. Cite the named regulatory regime per region (GDPR for EU, CPRA for California, etc.) for every consent-bound play.

7. KPIs and governance — Define the scorecard: revenue per session, attach rate, repeat-purchase rate, lifecycle-email revenue contribution, opt-out rate, recommendation click-through, "didn't-see-this-before" rate (novelty), and a customer-trust survey score. Add an off-site personalization scorecard line: AEO citation share by category and assistant-conversion attach rate where a delegated token bound the agent to a known shopper. Name the cross-functional RACI (merch, CRM, data, engineering, legal) and a monthly review cadence.

8. Config-utilization checklist — Confirm the brief uses cdp.identity_resolution, event_taxonomy, consent_regime, consent_regime.opt_in_floor_per_locale, audience_segments, experimentation.holdout_pct, personalization_holdout_owner, personalization_engine, margin_guardrails, fairness_constraints, and fairness_constraints.protected_class_proxies from config.yml rather than generic placeholders. The three fields added in v1.2 — personalization_holdout_owner, fairness_constraints.protected_class_proxies, and consent_regime.opt_in_floor_per_locale — bind three previously-implicit accountability and policy decisions to merchant configuration: the holdout owner binds incrementality-test accountability to a single named team rather than a shared RACI; the protected-class-proxy blocklist binds the engine's feature-gate to a named-signal exclusion list rather than relying on training-set heuristics to discover and avoid proxies; the per-locale opt-in floor binds the ship-vs.-defer decision per play to a merchant-risk-team-set threshold rather than a hardcoded default. Cite the named regulatory regime per consent-bound play, the named segment per anticipatory rail, and the named upstream-skill dependency (product-description-writer for catalog readiness, agentic-commerce-readiness for off-site parity, visual-merchandising-planogram-brief for in-store retail-media tie-in) per backlog row.

Output requirements:

• Executive summary (5–7 bullets) with the annualized revenue opportunity and the named upstream gap if any (catalog enrichment, AEO citation share)
• Data-readiness RAG — table: pillar → current → target → unlock, with the named upstream skill dependency per pillar
• Surface × algorithm × anticipatory-vs-reactive map — table per surface with the data-foundation threshold rule cited
• Prioritized backlog — table: play → hypothesis → metric → guardrail → effort → ICE → upstream dependency, with first three marked as Q1 quick wins
• Experimentation plan — holdout from config, MDE, attribution method per surface
• Trust & fairness checklist — with the named regulatory regime per region cited
• KPI scorecard + RACI — including off-site personalization line (AEO citation share, assistant-conversion attach rate)
• Config-utilization checklist — names the 11 config fields used (cdp.identity_resolution, event_taxonomy, consent_regime, consent_regime.opt_in_floor_per_locale, audience_segments, experimentation.holdout_pct, personalization_holdout_owner, personalization_engine, margin_guardrails, fairness_constraints, fairness_constraints.protected_class_proxies); marks any field that was unavailable so the merchant can backfill config.yml
• Cross-skill dependency map — explicit handoffs to product-description-writer, agentic-commerce-readiness, promotion-campaign-builder, and visual-merchandising-planogram-brief
• Professional formatting appropriate for a retail merch / CRM / data leadership audience
• Correct ecommerce, CDP, experimentation, and AEO terminology (RPS, holdout, MDE, anticipatory, reactive, semantic search, embedding, AEO, GEO, attach rate, novelty effect)
• Saved to outputs/ if the user confirms

Example Output

Reference run. Input: Mid-market DTC apparel + accessories retailer. Trailing-12-mo revenue $48.0M, 12.0M sessions, conversion 2.4%, AOV $86, repeat-purchase 31%, personalization-attributed revenue today ~14% (Nosto recommendations + Klaviyo lifecycle). Stack: Shopify Plus + Nosto + Klaviyo; CDP = Segment with loyalty-ID resolution but no cross-device stitch. Catalog ~6,200 SKUs, ~180 new/mo, apparel images rich but structured attributes thin (fit/fabric/occasion mostly blank). agentic-commerce-readiness not yet run; AEO citation share unknown. Consent: US + EU + Quebec; EU opt-in ~34%, Quebec ~41%. config.experimentation.holdout_pct = 8%; personalization_holdout_owner = "Data Science (lead: R. Okafor)".

Executive summary

• Personalization is ~14% of revenue ($6.72M) today; mature DTC apparel benchmarks run 20–30%. A realistic, holdout-defensible move of +6 pts to 20% is $9.60M attributed ($2.88M gross delta), or ~$1.44M annualized true incremental after a conservative 0.5 incrementality discount. RPS baseline = $4.00.
• The binding upstream gap is catalog content, not the engine. Fit/fabric/occasion attributes are blank on the majority of apparel SKUs, which caps semantic/visual recommendation lift regardless of Nosto tuning. Route to product-description-writer for embedding-ready enrichment of the top 3 revenue categories before funding a heavy ML re-rank.
• agentic-commerce-readiness has not run — off-site AEO citation share is unmeasured and presumed <5%. Per the lift-ceiling rule, do not sink heavy spend into off-site personalization until citation parity exists; run the readiness audit in parallel as a P0 dependency, not a personalization play.
• Three Q1 quick wins (below) target +3% RPS on PDP-touched sessions and a replenishment lifecycle flow; all three are reactive or well-bounded anticipatory plays the current data foundation can carry.
• Two EU/Quebec consent locales sit below typical opt-in floors (EU 34% vs. 30% GDPR floor → ships; Quebec 41% below the 50% Law 25/CASL floor → personalization plays defer in Quebec until opt-in rises or the risk team lowers the floor).

Data-readiness RAG

Surface × algorithm × anticipatory-vs-reactive map (decision rule: anticipatory only when replenishment-cadence std-dev <15% and segment confidence >0.7; else reactive)

Prioritized backlog (ICE = avg of Impact/Confidence/Ease, 1–10)

Experimentation plan — 8% permanent holdout (config.experimentation.holdout_pct), owner Data Science / R. Okafor (personalization_holdout_owner). For the PDP play, detecting a +3% relative conversion lift (2.40% → 2.472%) at 80% power, α .05 needs ≈719K sessions/arm (two-proportion; Z² 7.84 × [.0234+.0241] / .00072²) — ~6.2 weeks at ~1.0M sessions/mo split across 2 arms. Attribution: incrementality holdout for replenishment (long horizon), A/B for PDP/homepage. Novelty wash-out: discard first 14 days. Quarterly graduation review owned by holdout owner.

Trust & fairness checklist — Consent cited per region: GDPR (EU), CPRA (California), Law 25/CASL (Quebec). Opt-in floor gate: EU 34% ≥ 30% floor → ship; Quebec 41% < 50% floor → defer all plays in Quebec until opt-in rises. protected_class_proxies excluded at feature-gate load time: ZIP-as-income, surname-as-ethnicity, device-type-as-income, browse-history-as-religion/-sexuality. No personalized pricing on identical SKUs. Filter-bubble (discovery-narrowing) check on every rail.

KPI scorecard + RACI — RPS, attach rate, repeat-purchase, lifecycle-email revenue contribution, opt-out rate, recommendation CTR, novelty ("didn't-see-before") rate, trust-survey score. Off-site line: AEO citation share by category + assistant-conversion attach rate (both pending readiness audit). RACI: merch (A), CRM (R), data/R. Okafor (R, holdout owner), engineering (R), legal (C). Monthly review.

Config-utilization checklist — ✅ cdp.identity_resolution (cross-device gap) · ✅ event_taxonomy · ✅ consent_regime + ✅ consent_regime.opt_in_floor_per_locale (Quebec defer) · ✅ audience_segments · ✅ experimentation.holdout_pct (8%) · ✅ personalization_holdout_owner (R. Okafor) · ✅ personalization_engine (Nosto) · ✅ margin_guardrails · ✅ fairness_constraints + ✅ fairness_constraints.protected_class_proxies · ✅ loyalty.tiers. No fields missing.

Cross-skill dependency map — product-description-writer (catalog enrichment, blocks size-intent + visual recs) · agentic-commerce-readiness (off-site parity, blocks assistant play) · promotion-campaign-builder (SMS/TCPA) · visual-merchandising-planogram-brief (in-store retail-media tie-in).

All figures machine-verified: RPS $4.00; gross delta $2.88M; incremental $1.44M @0.5; PDP MDE ≈719K/arm at +3% relative lift, ~6.2 weeks.

Notes

• Personalization lift caps at the catalog-content quality and the off-site citation share. A merchant with a thin catalog or weak AEO citation share will not unlock industry-benchmark lift no matter how good the engine is. The upstream-skill dependency map is the load-bearing addition in v1.1.
• Anticipatory plays look impressive in slides but lose to reactive plays in production when the data foundation can't carry them. The decision rule (data-foundation threshold > segment confidence) is what stops the merchant from over-investing in next-best-action infrastructure that doesn't out-lift a re-rank.
• The off-site agent surface is a personalization surface in 2026, not a discovery surface. Once a delegated token binds the assistant to a known shopper, the same first-party signal the on-site surface uses should be feeding the agent — and the KPI scorecard line (AEO citation share, assistant-conversion attach rate) is what makes that visible to the merch team.
• Personalized pricing on identical SKUs is not in scope without a legal review. Personalization is the conversion lever, not the pricing lever.