The short answer: AI personalisation for a Saudi retailer comes down to four jobs: recommending the right products, returning the right results when shoppers search in Arabic, English or Arabizi, pricing within clear guardrails, and forecasting demand around a calendar shaped by Ramadan, Eid and the Hijri year. Research suggests the prize is real but bounded, with personalisation most often lifting revenue by 10-15% according to McKinsey. In Saudi Arabia it also has to respect the Personal Data Protection Law, which requires consent before personal data is used for direct marketing. Start with the first-party data you already hold, pick one measurable use case, and prove it against a baseline before scaling.
The personalisation opportunity
Saudi shoppers are almost universally online. The General Authority for Statistics' ICT Access and Usage survey 2025 found that 99.0% of individuals aged 15-74 use the internet, 98.4% go online from a mobile phone, and 96.8% are online at least once a day. The same survey found that 71.0% bought goods or services online (80.2% among Saudis), led by clothing, footwear and sports goods, bought by 88.8% of online shoppers, and cosmetics at 40.2%. Those are catalogue-heavy categories where sizing, style and repeat purchases make relevance matter. Yet many Saudi retailers still serve the same generic experience to every visitor.
AI personalisation techniques
Product recommendations
Collaborative filtering analyses purchase patterns across your customer base to suggest products. If customers who bought item A also bought item B, the system recommends B to new buyers of A. Content-based filtering analyses product attributes (brand, category, size, colour, price band) and matches them to customer preferences, which helps with new products that have no purchase history yet. Most retailers need both, plus a simple fallback such as best-sellers in the shopper's city for first-time visitors.
Arabic-native search
Saudi shoppers search in Arabic, English and Arabizi (Arabic written in Latin characters). Good search normalises Arabic spelling variants, such as the different forms of alef, taa marbuta written as haa, and optional diacritics, maps transliterations like "abaya" and "عباية" to the same products, and learns synonyms from real queries. Track the share of searches that return no results; it is the quickest way to see what your catalogue is missing or mislabelling.
Dynamic pricing
AI models can adjust prices based on demand signals, competitor pricing, inventory levels and customer segments. A returning loyal customer might see a different offer than a first-time visitor, trading conversion against margin. Set guardrails before switching it on: price floors and ceilings, limits on how often a price can change, and a rule that promotional prices shown in marketing are honoured at checkout. Shoppers notice inconsistent prices quickly, and trust is harder to rebuild than margin.
Predictive inventory
Machine learning models forecast demand at SKU level, accounting for seasonality, promotions and external factors like Ramadan, Hajj season and Saudi National Day. Because the Hijri year is about 11 days shorter than the Gregorian year, Ramadan and Eid move earlier every year, so a model trained only on Gregorian dates will misplace the peaks. Build Hijri-calendar features into the forecast to reduce both stockouts and overstock.
The data you need
| Signal | Typical source | Used for |
|---|---|---|
| Order history | ERP, POS and e-commerce orders | Recommendations, demand forecasts |
| Browsing and search events | Website and app analytics | Session recommendations, search tuning |
| Product attributes | Product catalogue or PIM | Content-based filtering, new-product cold start |
| Stock levels by location | Inventory system | Hiding out-of-stock suggestions, forecasts |
| Consent status | CRM or preference centre | Deciding who may receive personalised marketing |
Most of the effort goes into joining these sources on a stable customer and product ID. A recommendation engine fed with duplicated customers or inconsistent SKUs will confidently recommend the wrong things.
Privacy and consent under the PDPL
Personalisation runs on personal data, and the Kingdom's Personal Data Protection Law applies to any processing of personal data about individuals that takes place in the Kingdom (Article 2). Article 26 allows personal data to be processed for marketing only if it was collected directly from the individual and they have consented, and it excludes sensitive data, which the law defines to include health data. That matters for pharmacies and wellness retailers: purchase histories can reveal health conditions.
SDAIA's Implementing Regulation adds the operational detail in Article 29: obtain consent before direct marketing, make opting out at least as easy as opting in, identify the sender clearly, and stop immediately when consent is withdrawn. Article 24 requires the controller to notify the regulator within 72 hours of becoming aware of a personal data breach that could cause harm. In practice, store consent flags alongside the customer profile and have every campaign and personalised email check them before sending.
SDAIA's AI Ethics Principles apply to everyone who designs, develops or uses AI systems in the Kingdom, and their transparency and explainability principle asks that people affected by AI decisions can understand the rationale. A short "why am I seeing this?" note on recommendations is a simple way to meet that spirit.
Implementation with Odoo
Odoo eCommerce already has the display slots personalisation needs. According to Odoo's documentation, optional products are suggested when a shopper clicks Add to cart, accessory products appear in a Suggested Accessories section at the Review Order step, and alternative products appear at the bottom of the product page. Out of the box, each list is maintained by hand on the product form. An AI model can fill those fields automatically from co-purchase and browsing data and refresh them nightly, so the storefront shows computed suggestions without theme changes. The model can run alongside your Odoo instance, so order data does not need to leave your environment, and the same outputs can feed email campaigns.
How to evaluate and measure
- Pick one use case first. Product-page recommendations or search tuning are usually the fastest to ship and measure.
- Keep a holdout group. Show a share of traffic the non-personalised experience so you can separate the effect of personalisation from seasonality and promotions.
- Measure the right numbers. Recommendation click-through, conversion rate, average order value, zero-result search rate, return rate and stockouts.
- Avoid common pitfalls: recommending out-of-stock items, popularity bias that only ever promotes best-sellers, testing Arabic search with made-up queries instead of real logs, and personalised messages sent to customers who never consented.
What the research shows
We do not publish a universal uplift figure, because results depend on catalogue size, traffic and data quality. The most widely cited benchmark is McKinsey's Next in Personalization 2021 research: personalisation most often drives a 10-15% revenue lift, with company-specific results ranging from 5% to 25% depending on sector and execution. The same research found that 71% of consumers expect personalised interactions and 76% are frustrated when they do not get them.
Before launch we baseline your conversion rate, average order value and inventory position, then measure results against that baseline.
How Mantiqi can help
Mantiqi builds AI personalisation for Saudi retailers across Odoo ERP, custom storefronts and mobile apps. Our AI and intelligent automation team handles recommendation models and Arabic search, our data analytics team sets up the baselines and dashboards, and our web development team ships the storefront changes. See the wider retail and e-commerce picture, read how online stores handle ZATCA Phase 2 e-invoicing and where AI agents fit in Saudi businesses, or contact us to discuss your retail AI roadmap.