Source-linked AI summary
TwinStore: A Synthetic Consumer Platform Using Vision--Language Models
Minhyeok Lee
TL;DR
Retail decisions are hard to evaluate before purchase histories arrive: field experiments are slow and confounded, demand models miss shelf visuals, and synthetic consumers have mainly been validated against stated responses. TwinStore combines a digital store twin with vision–language shoppers that decide from shelves, record receipts, and support retail decision tools. On held-out purchase data, shoppers approach households' natural variation: category-mix divergence is 0.017–0.018 versus real drift of 0.012.
Problem
Retail decisions are hard to evaluate before purchase histories arrive: field experiments are slow and confounded, demand models miss shelf visuals, and synthetic consumers have mainly been validated against stated responses.
Source: §1 · p. 2Method
TwinStore combines a digital store twin with vision–language shoppers that decide from shelves, record receipts, and support retail decision tools.
Source: §1 · p. 2Results
On held-out purchase data, shoppers approach households' natural variation: category-mix divergence is 0.017–0.018 versus real drift of 0.012.
Source: §7 · p. 11, §12 · p. 15Takeaways & Limitations
TwinStore is designed to screen alternatives and sharpen hypotheses, narrowing the options that proceed to tests with real customers.
Source: §10 · p. 14Current validation moves purchases within receipt-derived trip sizes; price responses come from other retailers and meta-analyses, promotion lifts are observational, and the main test covers one US supermarket format.
Source: §10 · p. 14
Abstract
from arXiv · showhide
We present TwinStore, a source-available synthetic consumer platform for physical retail built on vision--language models. TwinStore turns a store into a digital twin---from 82 ready-made formats, a floor-plan image or phone photos of its shelves---and fills it with synthetic shoppers who plan a trip, walk the aisles and decide what to buy by looking at rendered shelves with packs, price tags and promotion signs. Every trip ends in a receipt, so a layout, price, pack or promotion can be tried on the same shoppers before it reaches the store. Around this core, TwinStore provides 3D playback, paired comparisons and layout optimization, shopper interviews, a trade-area model, counterfactual edits of real shelf photos, a marketing studio with eleven tools, and an assistant that turns a plain sentence into a drafted study or an answer grounded in a run's own numbers. In a held-out validation on eleven public datasets, simulated receipts approach the households' own natural variation (category-mix divergence 0.017--0.018 against a real drift of 0.012), price elasticities fall inside the range measured in the field, and three independent decision models agree. TwinStore runs on cloud or on-device models and is distributed with installers for Windows, macOS and Linux.
1 Introduction
Retail decisions are difficult to evaluate before purchase histories arrive: field experiments are slow and confounded, while demand models miss visual shelf cues and surveys measure stated preferences. TwinStore combines a store twin and vision–language-model shoppers who produce receipts, with tools for retail decisions and held-out validation against real purchases.
1 Introduction
- Field experiments take weeks, test few alternatives, and face confounds; demand models lack shelf-level visual inputs, while surveys measure stated preferences.Retail decisions are judged by the purchase histories they produce, which arrive only after the decision.Source: §1 · p. 2
- Prior synthetic consumers reproduce survey, behavioral-experiment, and willingness-to-pay responses, but read text and were mainly validated against stated responses [3], [1],, [24] [25].The passage contrasts those text-based agents with visual point-of-purchase decisions and purchase-log validation.Source: §1 · p. 2
- TwinStore combines a digital store twin with vision–language shoppers who see shelves, record trips as receipts, and support retail decision tools.The platform is evaluated against real purchase data in a held-out validation.Source: §1 · p. 2
- The vision-based shopper renders shelf faces with packs, price tags, and promotion signs so visual-merchandising changes reach the purchase decision.Source: §1 · p. 2
- TwinStore builds stores from 82 parameterized formats, 25 multi-floor buildings, floor-plan images, or shelf photographs.Source: §1 · p. 2
- The platform integrates playback, paired comparison, layout optimization, interviews, trade-area modeling, shelf-photo edits, eleven research tools, and a study-and-statistics assistant.Source: §1 · p. 2
- Held-out validation uses revealed purchases from eleven public datasets across six countries and three decision models.Source: §1 · p. 2
2 Design goals
TwinStore's design goals focus on making decisions from what shoppers see, comparing alternatives fairly, building store twins quickly, supporting use without a programmer, and checking shoppers against unseen purchase data. The section also states that known limits are reported.
2 Design goals
- Shoppers should decide from what they see because retail levers such as a new pack, extra facing, or promotion ribbon act through perception.Source: §2 · p. 2
- Comparisons reuse the same shoppers, random draws, and unaffected decisions across store alternatives.Source: §2 · p. 2
- Templates, floor-plan reading, and shelf-photo reading are intended to reduce store-building time to minutes.Source: §2 · p. 3
- TwinStore aims to make installation, model setup, experiments, and reports usable without a programmer.Source: §2 · p. 3
- The shoppers are checked against real purchase data they have not seen.Source: §2 · p. 3
- The report states that known limits are addressed.Source: §2 · p. 3
3 System overview
TwinStore turns real-store inputs into a twin, runs synthetic shoppers on one of three engines, and feeds their trips to decision tools. Its software combines an engine, web application, and launcher, and extends Parastore.
3 System overview
- Real-store inputs build a twin comprising the store, the day's situation, and the trade area; synthetic shoppers run on one of three engines and their trips feed decision tools.Source: §3 · p. 3
- The Python engine holds the domain model, simulation engines, and analyses, and calls language and vision models through a provider-independent layer.Source: §3 · p. 3
- A web application renders the store in 3D and hosts the tools, while a launcher installs and starts both on Windows, macOS, and Linux.Projects are stored on the user's machine; only configured model calls leave it.Source: §3 · p. 4
- TwinStore extends Parastore: the grid store model, persona and arrival generation, path planning, footfall maps, three-stage language engine, and 3D playback originate there.TwinStore adds the vision and rule-based engines, store generation from formats, floor plans and shelf photos, and several tools and models.Source: §3 · p. 4
4 Building the twin
The twin represents fixtures, products, and store operations; it can be generated from formats, floor plans, or shelf photos and configured with a day's conditions and a surrounding trade area.
4.1 The store model
- A store is a grid of fixtures with entrances and exits; fixtures have faces, levels, and slots holding products with prices and optional promotions.Opening hours and open weekdays are stored per store and floor, and uploaded product images appear on packs and shopper-facing shelf images.Source: §4.1 · p. 4
4.2 Formats and buildings
- TwinStore offers 82 furnished store formats and 25 multi-floor buildings, each with its own assortment and design, with configurable layout, assortment, hours, and other settings.A preview shows the floor plan, checks capacity, and can be compared with the format default.Source: §4.2 · p. 4
4.3 From floor plans and shelf photos
- Floor plans are read into walls, fixtures, and doors, while shelf photos yield editable planograms with product slots, names, and price tags.Reviewed shelf photos can be combined into a layout and used directly by vision shoppers instead of a render.Source: §4.3 · p. 4
4.4 The situation of the day
- A simulated-day situation combines conditions such as weekday, weather, events, crowding, stock-outs, promotions, and shopper notes.Nineteen situations are preset, and any combination can be composed.Source: §4.4 · p. 5
4.5 The trade area
- The trade area can be generated by neighborhood type and density or imported from OpenStreetMap, with visits allocated by a gravity model.The model also uses storefront and rival-shelf photos and accounts for linked trips from nearby venues.Source: §4.5 · p. 5
5 Synthetic shoppers
TwinStore creates shoppers from store or household information and offers rule-based, language, and vision engines for simulating their trips. Vision shoppers plan, inspect shelves, and make budget-bounded purchases, while paired runs keep unaffected decisions fixed for comparisons.
5.1 Personas
- Personas combine demographic and household attributes, budgets, visit reasons, and moods, and can be generated from store context or grounded in real households.Trade-area patterns and supplemental customer research can also shape the population.Source: §5.1 · p. 5
5.2 Three engines
- TwinStore offers rule-based, language, and vision engines, ranging from fast screening to rich reasoning and image-based decisions.The rule engine uses probabilities fitted to receipts; the language engine decides in three stages.Source: §5.2 · p. 5
5.3 The vision shopper
- A vision shopper plans at the entrance, follows a route, and uses rendered shelf images to select what attracts attention before a seeded draw determines purchase.The purchase draw respects budget and household trip size; captured shelf photos can replace rendered faces.Source: §5.3 · pp. 5, 7, Figure 3 · p. 6
5.4 Movement
- A path planner routes shoppers between stops, with walking speed and shelf time drawn from distributions matched to pedestrian and eye-tracking measurements.The 3D view animates shoppers walking, browsing, picking up products, and paying.Source: §5.4 · p. 7
5.5 Models
- TwinStore supports multiple model providers and can select an open vision–language model sized to the computer for local use.A run can assign shoppers across a weighted mix of models.Source: §5.5 · p. 7
5.6 Determinism and paired designs
- Per-shopper random streams and cached model responses keep decisions fixed across paired store runs, reducing comparison variance through common random numbers.New-seed replicates measure run-to-run noise, and paired bootstrap intervals accompany comparisons.Source: §5.6 · p. 7
6 Tools for decisions
TwinStore layers playback, comparisons, optimization, shopper interviews, shelf-photo editing, marketing studies, and a language-model assistant onto its simulated shoppers. These tools support examining trips and comparing proposed retail decisions using paired runs and run-specific statistics.
6.1 Playback and results
- Playback presents simulated days in 3D and reports sales, traffic, shelf-level choices, and integrity checks, with per-shopper results exportable for comparison with real receipts or sales.Views include footfall, dwell, and predicted congestion maps.Source: §6.1 · p. 7
6.2 Comparison and optimization
- Paired comparisons report metric differences and intervals, while an optimizer searches product or fixture arrangements toward a selected goal.The rule engine screens candidates on thousands of shoppers before vision shoppers judge finalists on a paired panel.Source: §6.2 · p. 7
6.3 Journeys and interviews
- Trip views show routes, shelf stops, and shopper thoughts, and individual or group interviews ground answers in shoppers’ personas and trips.Group interviews include four to forty shoppers from a segment.Source: §6.3 · p. 8
6.4 Shelf Lab
- Shelf Lab edits shelf photos to test changes such as product swaps, stock-outs, price tags, and promotions, then compares shopper choices across the same shoppers.It also identifies missing or misplaced products and produces budget-constrained restocking plans.Source: §6.4 · p. 8
6.5 Marketing studio
- The marketing studio provides eleven tools on shared synthetic consumers, and finished studies can be duplicated, rerun under new conditions, and exported.Its conjoint tool collects choices across profile tasks and uses a market simulator to recompute shares as products and prices change.Source: §6.5 · pp. 8–9, Table 2 · p. 9
6.6 Assistant
- The assistant routes plain-language requests to a suitable study or run query and drafts a setup for user review before anything runs.After a run, it answers from that run’s fact sheet and lists the figures supporting each answer.Source: §6.6 · pp. 9–10, Figure 5 · p. 10
7 Validation
On held-out shopping data, simulated receipts approach household-to-household and household-over-time patterns, while simulated responses to prices and promotions and shopper movement align with real-world references. Three decision models agree across the reported checks.
Protocol
- In The Complete Journey (801 households), settings are calibrated on history weeks 1–26 and targets 27–39, then frozen; the main test uses history weeks 1–39 and targets 40–53 on 150 age-stratified households excluded from calibration, plus 118 new households.Ten further datasets test single aspects, including price elasticity and walking speed.Source: §7 · p. 10
Buying the right things
- Category-mix divergence is 0.017–0.018 for simulated trips versus 0.012 for households’ own drift, and post-calibration receipt-classifier AUC is 0.56–0.62 versus 0.52 for household self-comparison.The validation summary covers held-out data, with ranges spanning decision models and modes.Source: §7 · pp. 11–12, Table 3 · p. 12
Reacting like real shoppers
- Simulated own-price elasticities range from −0.78 to −1.60, inside the field range of −0.70 to −1.76 and −2.62 across item-level sales and brand-level meta-analyses.Promotion ribbons lift products 1.25–1.7 times versus 1.4 times for an in-store display; effects of time and events correlate with real effects at 0.75–0.87.Source: §7 · pp. 12–13, Figure 7 · p. 12, Figure 8 · p. 13
Moving like real people
- Simulated shoppers’ median walking speed is 1.12 m/s versus 1.13 m/s across 13,159 pedestrian tracks, and their shelf time is 7–9 seconds versus 8.2 seconds per unplanned purchase.Source: §7 · p. 13
Replication
- Claude Haiku 5.5, Qwen 3.8 Flash, and GPT-6 Luna agree on every check with the same personas and settings.The mixed population usually falls between the best and middle model and performs best on product choice and demographic differences.Source: §7 · p. 13
8 Implementation and deployment
TwinStore is distributed with desktop installers and supports local or cloud model use. Its interface includes multiple languages and themes, while local credentials, cached calls, licensed assets, and automated tests support deployment.
Installation
- Releases for Windows, macOS, and Linux install required components on first start, and the built-in sample store can run without an account or key.Source: §8 · p. 13
Interface
- The application supports six languages, day and night themes, and layouts that keep playback and run controls accessible on phones and tablets.Source: §8 · p. 13
Models and cost
- TwinStore displays expected model-call counts, caches identical calls, stores credentials on the user’s computer, and can run offline after a local model is downloaded.Source: §8 · p. 14
Assets
- The realistic view uses MIT-licensed avatars and CC0 materials and photogrammetry scans, with asset sources recorded in the repository.Source: §8 · p. 14
Quality
- Several thousand automated tests cover the engine and application, including interface checks at phone, tablet, and desktop sizes.Source: §8 · p. 14
9 Related work
Prior work establishes synthetic consumers and simulated shopping trips, while TwinStore moves them into a visual store and validates them against revealed purchases.
Synthetic consumers
Shopping simulation
- Grocery-shopping models and instrumented-store path data established routes, stops and baskets as trip structure; generative agents added language-model reasoning in simulated spaces [24].In-store research links facings and position to attention and choice, while vision–language models make those visual cues available to synthetic shoppers [2] [20].Source: §9 · p. 14
- Digital twins provide a frame for trying store decisions on a mirror before implementation.TwinStore moves synthetic consumers into a visual store and validates them on revealed purchases.Source: §9 · p. 14
10 Limitations
The validation has boundaries in trip behavior, evidence sources and store settings, so TwinStore is positioned to screen alternatives and sharpen hypotheses before real-customer testing.
Scope
- Current validation keeps household trip size fixed from receipts, so interventions shift purchases within trips rather than testing effects mediated by additional store visits.Extending validation to additional-visit effects is identified as a next step.Source: §10 · p. 14
Evidence
- Price responses come from other retailers and meta-analyses, while promotion lifts are observational.Source: §10 · p. 14
Scope
- The main test covers one US supermarket format, and convenience-store situation effects rely on priors until public convenience-store receipts are available.The models’ inherited knowledge and biases also motivate using TwinStore to screen alternatives and sharpen hypotheses before testing with real customers.Source: §10 · p. 14
11 Availability and license
TwinStore is source-available under a license permitting research and non-commercial use, with installers for three desktop operating systems.
Availability
- TwinStore is distributed with installers for Windows, macOS and Linux.Source: §11 · p. 14
License
- The TwinStore License 1.0 permits research and non-commercial use; the full terms are in the repository.Source: §11 · p. 14
- TwinStore builds on Parastore, whose MIT License notice is retained.Source: §11 · p. 14
12 Conclusion
TwinStore models physical stores and visual shopping trips as receipts for configurations not yet deployed; held-out receipts approach households’ natural variation and field-measured response magnitudes across three models.
12 Conclusion
- Held-out receipts approach households’ natural variation and match field-measured response magnitudes for price, promotion, distance and day across three independent models.The platform turns store formats, floor plans or photographs into a twin where visual shoppers generate receipts for undeployed configurations.Source: §12 · pp. 14–15
A Reading the screens
- The interface makes the receipt-producing chain inspectable, complementing validation of how closely simulated receipts match real ones.The screens expose repeatable checks rather than asking users to take the mechanism on trust.Source: A Reading the screens · p. 17
B Building the twin
- TwinStore builds stores from furnished formats or plan images, and its 3D view renders the same records that shoppers use.The format library includes 82 options; plan images are read by a vision model with a geometric fallback.Source: B Building the twin · p. 18
C From a shelf photograph to a planogram and its counterfactuals
- Shelf Lab applies local photo edits—such as swaps, stock-outs, new packs, price changes and promotion signs—so only edited regions differ from the original.Sim2real compares the same shoppers and seeds across the photo, a simulator render and optionally the raw photo.Source: Figure 10 · p. 20, C From a shelf photograph to a… · p. 21
D The simulated day
- Playback combines recorded shopper thoughts and picks with footfall, predicted congestion and fixture-level revenue, buyers and stop-to-buy rates.The sample-store replay reports live totals, while the layers let managers compare store activity before and after a change.Source: D The simulated day · p. 21
E Configuring a run
- Run configuration exposes the engine, population and expected call cost, with seeds and cached responses supporting reproducibility and paired comparisons.The rule-based preview covers all 2,440 personas without model calls; vision runs can sample 30 shoppers, and advanced settings specify the day and shopper decisions.Source: E Configuring a run · pp. 22–23, Figure 12 · p. 23
F What a vision shopper sees
- Vision shoppers receive store and shelf images with persona and trip context, then return noticed products, a choice, interest rating and reason.Shelf promotions and price changes alter only the relevant image regions; cached full inputs keep shoppers who do not see the changed face identical across arms.Source: F What a vision shopper sees · pp. 23, 25, Figure 13 · p. 24
G Following one shopper
- The shopper view links routes and recorded thoughts to a receipt, letting users inspect whether narrated steps align with purchases.A single trip is also available as a per-shopper export row from which aggregates and validation numbers are computed.Source: G Following one shopper · p. 25, Figure 14 · p. 26
H From trips to results
TwinStore presents trip-level results, paired layout comparisons, marketing studies, trade-area estimates, and an interface for running and inspecting studies. Together, these screens show how simulated receipts are produced and compared with store data.
H From trips to results
- The results panels aggregate trips, and the sample run’s $21,032.06 revenue equals the spend total across its 2,440 exported shopper rows.The panels also show hourly visitors and running receipt totals during playback, then post-day results, product rankings, and diagnostics.Source: H From trips to results · pp. 26–27, Figure 15 · p. 27
I Comparing alternatives on the same shoppers
- Layout 1 has 2.0% higher spend per visitor ($8.62 vs. $8.45; 95% interval +0.8 to +3.2%) and 1.6% more units per visitor, without bringing in more buyers.Source: I Comparing alternatives on the same… · p. 29
- Pairing produced a spend-difference standard error of 5.2 cents versus 10.9 unpaired, and reaching the same precision without pairing would require about 4.3 times as many shoppers per layout.The spend correlation across layouts was 0.77; unpaired, the 2.0 % difference had z=1.5 and an interval from −0.5 to +4.5 %, while paired z=3.2.Source: I Comparing alternatives on the same… · p. 29
J The marketing studio
- A 30 % discount and an eye-level move each increased tea sales per 100 visitors, but neither beat the control on revenue per visitor beyond noise.The discount sold 0.8 more teas per 100 visitors (+50 %) and eye level 2.9 more (+181 %); all three arms had $7.37 revenue per visitor.Source: J The marketing studio · p. 29, Figure 17 · p. 30
K The trade area
- Weekday and weekend trade areas differ in traffic and visitor mix: weekend visitors fell to 1,031 from 1,366, while residents rose from 36 % to 55 % of visitors.On weekdays, 187 of 1,366 visitors came on linked trips; weekend customers came from farther, and the early-morning peak disappeared.Source: Figure 19 · p. 32, K The trade area · p. 33
L Working with the platform
- The six-language, mobile-friendly interface routes a typed question to an assistant-drafted study, which waits for user review before running.Users can choose a local vision model suited to their computer or use cloud providers whose keys remain on the machine.Source: L Working with the platform · p. 33, Figure 20 · p. 34
M From the screens to the validation
- The screens show how the twin and shelf images produce trips and paired comparisons, while validation measures receipt similarity against unseen real purchase data.Users can inspect why a result arose on their own store.Source: M From the screens to the validation · p. 34