Pricing intelligence for retail: lessons from Estée Lauder and Leroy Merlin

One-minute summary

In April 2026, a large home-improvement retailer in Santiago lost 8% of margin in a single quarter. Inflation was not the cause. The cause was that 23% of its SKUs were priced 4% to 11% above Sodimac, and the team learned about it through an outside agency PDF — once every 14 days.

This is the standard photo of LATAM retail in 2026. Competitor prices move every day, marketplaces rewrite tags several times a day, and retailers respond with spreadsheets. Pricing intelligence is the infrastructure that closes the gap: scraping → SKU normalization → ML models → rules → dashboard → automated action in the ERP.

Pricing intelligence for retail is not a cheap parsing trick. It is a system that runs three jobs in parallel: (1) see the competition in real time, (2) defend the brand's MAP policy against gray-market imports, (3) optimize own prices using category-level elasticity. Estée Lauder in LATAM works mostly on (2): controlling 12 brands across marketplaces. Leroy Merlin works on (1) and (3): daily monitoring of 50,000+ SKUs against Sodimac, Easy, Promart. Both cases are unpacked below, with stack, mistakes, and where this simply does not work.

• Pricing intelligence ≠ web scraping. Parsing is 10% of the work. The other 90% is SKU matching, ML, a rules engine, and ERP integration.
• Estée Lauder in LATAM: monitoring of 12 brands across 8+ marketplaces in 4 countries, promo-fraud detection, MAP compliance restored from 71% to 94% in 9 months.
• Leroy Merlin (Russian case, structure applies to LATAM): scraping 50,000+ SKUs × 3 competitors, daily refresh, margin uplift from +1.4 to +2.8 pp depending on the category.
• Stack: Scrapy/Playwright + residential proxies, PostgreSQL + ClickHouse, dbt, ML matching on embeddings, Metabase, Odoo as the action layer.
• Where it does not work: B2B with tenders, luxury fashion with unique SKUs, discounters with margin under 8%.
• Cost: SMB MVP runs USD 8,000-15,000 in implementation + USD 400-800 per month in runtime. Payback in 3-6 months.

Where LATAM demand came from

Between 2024 and 2026 three factors converged and turned pricing intelligence into a required part of the LATAM retail stack. For a broader regional view, see the country-by-country panorama in LATAM countries.

First — e-commerce growth. Aggregated figures from CEPAL and national statistics offices show the online share of general retail climbed from 6% (2019) to 19-24% (2025) in Mexico, Chile, and Colombia. In Peru, from 4% to 12%. Mercado Libre controls between 27% (MX) and 65% (AR) of online retail. Price transparency has shifted from "walk into three stores" to "five seconds in the browser."

Second — the marketplace effect. When your SKU sits on Mercado Libre, Falabella, or Amazon Mexico, the Buy Box algorithm ranks sellers mostly by price. One competitor cuts 3% and your SKU falls out of the Buy Box. The result: a 40% to 60% drop in daily sales for that SKU. The reaction has to be measured in hours, not weeks.

Third — margin pressure. Between 2024 and 2025 LATAM retail rode the inflation wave (Argentina peaked near 117%, Mexico ran 4-5%, Chile 3-4%) on top of higher import prices from Asia (container logistics up 18% YoY). Brands and retailers squeeze every margin point, and pricing intelligence shows where they lose it.

Timeline of key shifts 2024-2026:

• Q2 2024: Mercado Libre tightens the "Mejor Precio" program and rewrites Buy Box ranking weights.
• Q4 2024: Falabella announces a dynamic-pricing roadmap for its own marketplace.
• Q1 2025: AWS adds regional EC2 Spot capacity in Mexico and Brazil. Scraping gets cheaper.
• Q3 2025: Amazon Mexico launches the Repricer tool for third-party sellers. Pressure on average rev/SKU.
• Q1 2026: OECD publishes updated guidance on personalised pricing. European brands (Leroy Merlin, L'Oréal, Estée Lauder) pull their LATAM operations toward global standards.

Technical stack: what you actually need

Pricing intelligence is built from 6 layers. Skip any one and the system ends up as "a PDF of competitor prices once a month." The data engineering architecture for enterprise retail walks through how to connect layers 3 and 4 without piling on technical debt.

#1. Data collection (scraping / API)

Competitor prices come in three ways:

• Public website scraping (90% of cases) — Scrapy + Playwright/Puppeteer for JS-heavy sites. Proxy rotation is mandatory — Bright Data, Oxylabs, Smartproxy. In LATAM you need residential IPs in the right country: Mercado Libre and Falabella geo-block foreign IPs.
• Marketplace API (where available) — Mercado Libre Developers exposes competitor Buy Box prices; Amazon SP-API supports reactive repricing. It does not cover everything, but it is cheaper than scraping.
• Affiliate feeds and Google Shopping feed — less fresh, but legal and stable. They serve as backup and cross-check.

#2. Product matching

The hardest and most underrated piece. You have a SKU "Pintura látex blanco 4L Sherwin-Williams ref. SW-2003." Sodimac lists the same product as "Sherwin Williams Pintura Latex 4 Litros Blanco Mate ref. SW2003-04." Text match does not work. You need:

• Embeddings (Sentence-BERT for Spanish or multilingual) with cosine similarity above 0.85 for auto-match.
• Image matching (optional, via CLIP) as a second pass.
• A manual review queue for scores between 0.70 and 0.85.

At 50,000 SKUs × 3 competitors you get 150,000 pairs. Without an ML pipeline it is not solvable.

#3. Storage

• PostgreSQL for operations (current prices, alerts, rules).
• ClickHouse or BigQuery for historical price series (50k SKUs × 3 competitors × 365 days ≈ 55M points per year). Detailed comparison in ClickHouse vs BigQuery for retail.
• dbt for the transformation layer (normalization, KPIs, elasticity).

#4. Analytics and ML

• Price elasticity at the category level (not at SKU — statistically unreliable): regression with fixed effects or hierarchical Bayesian models.
• Demand forecast — Prophet or a statistical baseline. LLMs are not needed here.
• Anomaly detection on promo-fraud — z-score or Isolation Forest is enough.

#5. Rules engine

This is where dynamic pricing makes the call: "if the competitor is 5% cheaper, drop 3%, but never below the floor." Open-source — Drools, Camunda DMN. Hand-rolled YAML rules are the norm for SMBs.

#6. Action layer

The price must flow back into:

• ERP (Odoo, SAP, NetSuite) → stores + e-commerce.
• Marketplace via API.
• Dashboards for merchandisers.

Odoo with a custom module (along the lines of dynamic_pricing) is critical for syncing prices under rules, versioning (audit log), and rollback on error. The base mechanics of pricelists and discount formulas live in the official Odoo documentation. To map how rules and custom modules fit together, start from an Odoo audit.

SMB stack cost (50-200k SKUs)

Compare with ROI: 1.5 pp of margin on USD 5M of revenue equals USD 75,000 per year. Payback in 1.5 to 3 months.

When it works — when it does not

The system pays back under specific conditions. If your situation does not fit, overhead beats the benefit. Before sinking budget into a pilot, review the layer list in Odoo implementation for LATAM retail.

It works if:

• Catalog of 500+ SKUs and 3+ visible competitors. Less than that is overkill — manual monitoring twice a week is enough.
• A category with elasticity above 0.5 (electronics, household chemicals, DIY materials, basic cosmetics). The buyer is price-sensitive: a cheaper shelf moves demand.
• E-commerce or marketplace above 20% of revenue. Pricing intelligence is an online paradigm; brick-and-mortar gets the derivative effect through omnichannel sync.
• A merchandising team ready to react. If price changes go through a committee twice a month, daily refresh is wasted spend.

It does not work if:

• B2B with tenders or contract pricing. Price is negotiated per deal; competitor monitoring is irrelevant.
• Luxury fashion or unique SKUs. If your capsule collection runs 200 SKUs no one else sells, pricing intelligence is replaced by brand-positioning research.
• Discount retailer with margin under 8%. You are already on the floor; there is nowhere to go.
• Rigid supplier MAP with no flexibility. If the contract forbids going under MAP, dynamics collapse. The focus shifts to MAP-violation tracking (the Estée Lauder scenario).

Gray zone (depends on the detail):

• Pharma and regulated categories. Mexico, Chile, and Colombia have reference-pricing regulation; ANMAT in Argentina periodically freezes certain prescription prices. Pricing intelligence works on the non-regulated mix (cosmetics, OTC, non-essential medical devices).
• SMB with a single marketplace channel. If 90% of volume runs through Mercado Libre, the built-in Repricer covers around 70% of the problem; external pricing intelligence is justified only for merch insights and cross-channel view.
• Highly seasonal categories. During Buen Fin, Cyber Day, and Hot Sale, elasticity and competitive dynamics shift hard. Models trained on off-season data emit false signals.

Typical mistakes

Five years watching pricing projects in LATAM produced a repeatable list of mistakes. Each one costs 6 to 18 months of runway if it is not caught early. More examples in the research archive.

Mistake 1: Treating pricing intelligence as a "parsing project." Parsing is 10%. SKU matching, 30%. Analytics and rules, 40%. ERP integration, 20%. Teams that hire a single "scripter" end up with a CSV archive of competitor prices, not a working system.

Mistake 2: Daily refresh where hourly is needed. Electronics on Mercado Libre rewrite prices 4 to 6 times a day. Daily refresh means reacting to yesterday's price. For FMCG and DIY, daily is fine; for consumer electronics and flash-promo categories, 2-4 refreshes per day is the minimum.

Mistake 3: Ignoring promo as a separate signal. The competitor holds MSRP at USD 100 and at checkout offers an 18% coupon. Effective price = USD 82, listed = USD 100. If the scraper only takes listed, you see the market 18% higher than it actually is and lose conversion.

Mistake 4: Turning on dynamic pricing without floor or ceiling. Already covered in the callout: no guard rails takes you to USD 1 or to an absurd ceiling. Floor, ceiling, alerts, and rollback are mandatory.

Mistake 5: Computing ML elasticity on the aggregate, with no segmentation. Dog food in Mercado Libre Mexico has different elasticity than the same SKU in Mercado Libre Chile. ML on the aggregate produces a meaningless average. Segment by region, channel, and customer cohort.

Mistake 6: Treating pricing intelligence as a one-time project. Competitors restructure their sites every 2 to 4 months, which breaks scrapers. Without a maintenance budget (1 dev × 30% time), the system falls apart in 6 months.

Case 1: Estée Lauder — monitoring 12 brands in LATAM

Situation. A holding with 12 cosmetics brands (premium and mass-market) across 4 LATAM markets (Mexico, Chile, Colombia, Peru). Distribution via authorized retailers (Sephora, Falabella, Ripley, Liverpool), own e-commerce, and Mercado Libre/Amazon. Pain: gray-market importers (individuals, small online stores) dumping prices 25% to 40% below MSRP. Consequences:

1. Cannibalization of authorized-partner sales, generating distribution conflict.
2. Brand equity damage (premium sold at mass-market prices).
3. Cannibalization of promo campaigns (the promo price was already matched by gray stock before launch).

What they did. Pricing intelligence pipeline:

• Daily scraping of 8 marketplaces and ~120 e-commerce stores in 4 countries.
• Product matching via embeddings + manual QA queue for borderline scores.
• Anomaly detection: SKU priced more than 15% below MSRP → alert in Slack to the brand-protection team.
• Auto-notify marketplace ops (Mercado Libre, Amazon) via the MAP-violation request form.
• Cross-reference with the internal promo calendar: if the scraper sees a discount not on the approved plan → promo-fraud flag and investigation.

Result (anonymized figures, snapshot across 11 brands in the holding):

• Promo-fraud detection rate: 67% (vs ~12% before the project).
• MAP compliance restored: from 71% to 94% in 9 months.
• Remarketing ROAS: from 1.5x to 4.2x. Side effect: a cleaner price field → higher LTV → higher bids in paid media start paying back.
• 340+ unauthorized listings removed via marketplace ops.

The full case in Estée Lauder case: pricing & promo covers the MAP-protection stack detail and the actual forms used with Mercado Libre and Amazon.

Case 2: Leroy Merlin — dynamic pricing on 50,000+ SKUs

Situation. A home-improvement retailer with physical and online channels. Catalog of ~120,000 active SKUs, of which ~50,000 have 2+ visible online competitors. Pricing reactivity: 2-3 weeks via email and spreadsheets. Competitors (Sodimac in CL/PE, Easy in AR, Promart in PE) changed prices 1-3 times a week; the team saw those changes 14-18 days later. Margin erosion: ~1.8 pp YoY in high-elasticity categories (paints, tools, lighting).

What they did.

• Build in-house scraping pipeline (Scrapy + residential proxies for 3 competitors).
• Product matching via embeddings; ~80% auto-match, 20% to manual queue for merchandising.
• ClickHouse for price history; daily refresh for DIY, weekly for low-elasticity categories.
• Rules engine with 6 rule types: price-match, undercut-by-X, premium-hold, MAP-protect, clearance-accelerate, intro-margin.
• Integration with Odoo through a custom dynamic_pricing module. Prices update in POS and e-commerce nightly after rules run.
• BI dashboard in Metabase: SKU coverage, win/loss vs competitor, daily margin impact.

Result after 12 months:

• Reaction time down from 14-18 days to 1 day (95% of rules automated).
• Margin uplift: +1.4 pp on high-elasticity categories and +2.8 pp on private label.
• 23% of SKUs repositioned — some up (where they were undervalued), some down. It was not "everything cheaper."
• Traffic recovery of +11% in categories where they had been losing marketplace position.

The case is Russian by origin, but the structure transfers almost 1-to-1 to LATAM: home-improvement vertical, 3 visible competitors, marketplace effect, a category grid by elasticity that lines up. Sodimac plays the OBI role, Promart plays the Castorama role, and so on. The extended version with architecture lives in Leroy Merlin case: web scraping + dynamic pricing.

Download the blueprint: Pricing Intelligence Stack

The MVP ships in 8-12 weeks on a USD 25,000-60,000 budget. For SMBs with 5,000+ SKUs and 3+ competitors, payback runs 3-6 months. The "Pricing Intelligence Stack Blueprint" lead magnet is a PDF with the 6-layer architecture, Scrapy + matching snippets, an ROI calculator scoped to your catalog, and a sample rules-engine YAML. Download the blueprint (email-gated).

Frequently asked questions