# AI in the UGC loop, part 4, personalisation: the right clip for the right shopper Nine product pages out of ten sort UGC newest first, a sort built for the brand's convenience. The fix is a four-rung ladder up to 1:1 persona matching. By Rohin Aggarwal · 2026-05-27 A shopper in Munich, a shopper in Manila and a shopper in Manchester open the same PDP. The hero clip should be three different videos. Most stores serve them one. The lift from getting it right is not subtle, and the cost of implementing it is smaller than that lift makes you expect. **AI in the UGC loop · part 4 of 4** - Nine PDPs out of ten sort UGC "newest first", a sort built for the brand’s convenience, not the shopper. - Recency has only weak correlation with conversion. Affinity, session context and persona fit matter far more. - Personalisation is a four-rung ladder: recency, rule-based, segment-based, persona 1:1. Most brands sit on rung one. - Track per-persona conversion uplift versus the newest-first baseline, as a clean A/B, held for a full retail cycle. Open any ten product pages today and look at the UGC section. Nine of them will be sorted "newest first". The tenth will be "highest engagement". Both are sorting strategies built for the brand’s convenience, not the shopper’s experience. This is the last and most expensive blind spot in the UGC stack. You sourced well in [part 1](/blog/ai-ugc-loop-ingestion), tagged thoroughly in [part 2](/blog/ai-ugc-loop-tagging), moderated cleanly in [part 3](/blog/ai-ugc-loop-moderation), and then you show every shopper the same six clips in the same order. The shopper hunting for a wedding-guest dress gets the same lookbook as the one buying gym leggings. And the merchandising team wonders why PDP conversion has plateaued. ## Why "newest first" loses Newest-first is not bad so much as blind. It ignores three things that move conversion more than recency does. - Affinity, what this shopper responds to, from browsing history and past purchases. A shopper who keeps clicking outdoor lifestyle clips wants more of those, not whatever was uploaded yesterday. - Context: what this session is about, from the search query, landing page, time of day and device. A shopper arriving from "wedding guest dress" wants formal-occasion content. - Persona fit: body type, age, aesthetic, geography. "Looks like me" content is a strong conversion signal in apparel and beauty in particular. Newest-first uses none of these. It optimises for one signal, recency, which correlates only weakly with conversion. The opportunity cost is real, and AI is what finally makes per-shopper matching operational instead of a merchandiser hand-curating every page. ## The personalisation maturity ladder Most brands are stuck on rung one. The conversion sits on the three rungs above it. ### The four rungs _Lift figures are composite ranges from public UGC-personalisation benchmarks, expressed against a newest-first baseline. See the note on numbers, these are not Idukki-measured customer results._ **Rung 1–2: Recency → rule-based** Where most brands live, and the first step off it. +5–10% — Rule-based vs newest-first - ✓ No tooling needed for recency - ✓ Rules are easy to reason about - ✗ Recency ignores the shopper entirely - ✗ Rule maintenance becomes a part-time job - ✗ Rules do not compose well **Rung 3: Segment-based** The system learns clusters of similar shoppers and sorts UGC per segment. +10–20% — Segment vs newest-first - ✓ Scales without per-page curation - ✓ Three to ten segments covers most traffic - ✗ A segment is still an average - ✗ Misses within-segment variation **Rung 4: Persona 1:1** A persona profile per shopper; UGC matched at the asset level. +20–45% — Persona 1:1 vs newest-first - ✓ Two shoppers in one segment can see different galleries - ✓ Matches on the dimensions that actually drive conversion - ✗ Needs good tagging and persona coverage to work - ✗ Requires A/B discipline to prove - **+5–10%** — Rule-based vs newest-first (First step off the recency sort) - **+10–20%** — Segment-based vs newest-first (Three to ten segments covers most traffic) - **+20–45%** — Persona 1:1 vs newest-first (Per-shopper persona profile, per-asset matching) - **60%+** — Persona coverage needed before A/B is meaningful (Below this, you mostly test the unknown-shopper fallback) _Personalisation lift vs newest-first baseline (consolidated industry benchmarks)._ ## What this looks like on a product page Three shoppers land on the same midi dress at the same time. Shopper A is a new visitor from Pinterest on mobile who has browsed wedding-guest content all morning. Shopper B is a repeat buyer on desktop who has bought two casual dresses before. Shopper C clicked through from a "new arrivals" email and skews younger and trend-led. On a rung-one site all three see "newest first". On a rung-four site, A sees the dress styled as a wedding-guest look on a model with similar colouring; B sees everyday styling with reviews about fit and washability; C sees younger models, trend-led pairings, creator audio. Same product, three galleries, three conversion rates. The lift does not come from more content. It comes from matching what is there to who is there. > The PDP is the last 30 seconds of a decision. Personalisation is making those 30 seconds about the shopper in front of you, not the asset you happened to upload most recently. ## The one number to track The headline metric is per-persona PDP conversion uplift against the newest-first baseline, read weekly per persona. Set it up as a clean A/B from day one: control sees the recency sort, treatment sees the personalised sort. Hold it for a full retail cycle, four weeks minimum and eight ideal, so the result spans both new and returning shopper mixes. **60%+** — Persona coverage you need before the test is meaningful (Below this, personalisation is mostly serving the unknown-shopper fallback, so you are not actually testing personalisation.) > What I like best about Idukki is how easy it is to launch experiments and personalization campaigns without needing heavy dev support. The UI is intuitive, setup is quick, and everything is geared toward helping you move fast and see results. On top of that, their team is incredibly responsive and proactive, they don't just support you, they collaborate with you to grow your CRO strategy. > — Tebogo M., CRO Specialist, verbatim, G2 review, April 11 2025 ## Three things to do this quarter 1. Run the baseline A/B, just two cells, newest-first versus a basic affinity sort. The lift number makes the case for going further. 2. Define your top five personas. Not fifty. Five, covering 70%+ of traffic. One page each: what they look like, what they shop for, what UGC resonates. 3. Audit asset utilisation. Pull last quarter’s UGC and find what percentage was served at least 100 times. Below 50% and your sort is wasting most of your library. **Where Idukki fits:** This is what AI Personas is for: per-shopper persona profiles, per-asset matching, guardrails the merchandiser controls, and A/B reporting baked in. Combine it with AI Player to make the gallery itself adaptive: the right clip, in the right format, at the right moment in the session. The pipeline you built across the previous three posts becomes a real merchandising surface, not a static gallery. **The series, in four lines:** Ingestion: stop chasing, build inbound capture, track net rights-cleared assets per week. Tagging: let AI type, let humans edit, track P75 time-to-tag in hours. Moderation: a three-tier queue with SLAs: track latency, escape rate, overturn rate. Personalisation: climb the ladder from recency to 1:1, track per-persona uplift. If you only do one, do tagging; everything else compounds on it. That is the series. The product view of this stage lives on the [AI shopper](/ai-shopper) and [AI Player](/ai-player) pages. Start at [part 1](/blog/ai-ugc-loop-ingestion) if you came in here first. [Get the full series. AI in the UGC loop — All four parts plus the pipeline self-audit worksheet, in one file.](/downloads/ai-in-the-ugc-loop-series) ### Sources + note on numbers - [Nosto, Ecommerce Personalization research](https://www.nosto.com/resources/) — Personalisation conversion-lift benchmarks across retail. - [Bazaarvoice, Shopper Experience Index](https://www.bazaarvoice.com/resources/) — UGC-on-PDP conversion behaviour. - [Baymard Institute, product page UX research](https://baymard.com/research) — Shopper behaviour on PDP content surfaces. - Note on numbers — The ladder lift ranges are composite figures consolidated from the public personalisation benchmarks above, expressed against a newest-first baseline. They are representative, not Idukki-measured customer results. --- Canonical: https://idukki.io/blog/ai-ugc-loop-personalization Tags: ugc, personalization, ai-personas, ai-in-ugc-loop