METHODOLOGY
How We Measure
Every number IntellX produces passes through five layers of ingestion, parsing, matching, weighting, and correction before it reaches a dashboard or a boardroom. This page explains each layer. If you are evaluating IntellX as a data source, this is the document your research director needs.
PANEL DESIGN
Designed for the market, not adapted from another one.
Most purchase panels in MENA are localized versions of methodologies built for Western European grocery markets. IntellX was designed from the ground up for the structural realities of Saudi consumption: 40% traditional trade, a 68/32 Saudi-Expat household split, six distinct regional consumption cultures, and a rapidly growing out-of-home channel that no legacy methodology was ever designed to capture.
Recruitment and Stratification
Panelists are recruited across the demographic dimensions that define the Saudi consumption universe: region (10 strata), socioeconomic class (4 tiers: AB, C1, C2, D), nationality (Saudi, Expat Arab, Expat Non-Arab), and household size (1-3, 4-5, 6+). Recruitment quotas are set against national population estimates from the General Authority for Statistics (GASTAT) census and Ministry of Human Resources labor reports. Labor camp populations are excluded from the residential household universe, producing a 68/32 Saudi-Expat household ratio rather than the raw 55/45 census split.
Dual-Mode Capture
Every panelist operates in two capture modes. Household mode records purchases made for the home, weighted against a universe of over 8 million Saudi residential households. Individual mode records personal consumption occasions outside the home: gas stations, cafes, delivery apps, convenience stores. The distinction is identity-based, not location-based. A panelist buying coffee for themselves at a gas station logs it as individual. A panelist buying milk at that same gas station for the family logs it as household. This dual-mode architecture is what enables all-channel coverage without conflating in-home and out-of-home consumption behaviors.
Weighting
Raw panelist counts do not equal national consumption. Iterative Proportional Fitting (IPF) projects panelist behavior onto the full Saudi residential household universe by adjusting weights across all four demographic dimensions simultaneously. The algorithm converges in 5 iterations. Weights are capped at 3x the median (5x hard ceiling) to prevent any single household from dominating category-level estimates. Effective sample size is monitored continuously: ESS must exceed 60% of raw panel count. Design effect targets are held below 1.7.
THE OOH GAP
25-35% of FMCG spend is structurally invisible to legacy measurement.
Traditional purchase panels were designed to measure what enters the home. In-home diary panels capture grocery trips and supermarket purchases with high accuracy. But they were never designed to capture the coffee bought at a gas station, the snack from a convenience store during a commute, the meal ordered on a delivery app and consumed at the office, or the water bottle grabbed at a pharmacy checkout.
In Saudi Arabia, 25-35% of FMCG consumption happens outside the home. This is not a marginal channel. It is growing at 15-25% annually, driven by delivery app adoption, fuel station modernization, and a generational shift toward out-of-home occasions. Legacy panels cannot capture it because their methodology was not designed for it: diary panels record items brought home, retail audits measure what stores sell. Neither captures what an individual consumes on the go.
How we solve it
IntellX captures out-of-home consumption at the item level through a consumer-side mobile app. Panelists photograph products, scan barcodes, and capture receipts at the point of consumption. Seven AI agents parse the multi-lingual transaction data. Every OOH occasion is tagged with context: who was there, why the purchase happened, and whether it was consumed immediately or taken home. The result is a complete consumption picture that includes the channels legacy panels structurally cannot reach.
THE 5-STEP PIPELINE
From raw signal to boardroom-ready intelligence.
Every purchase captured by a panelist passes through five processing stages before it becomes a data point in any analysis. Each stage solves a specific problem that has historically required manual intervention, specialist knowledge, or both.
Ingest
Capture from any channel, in any format.
Four capture methods operate simultaneously: barcode scan (packaged goods), receipt photo with AI extraction (grocery trips), product photo with AI identification (unpackaged or out-of-home items), and manual quick-log (fallback). Consumer-side collection means no retailer partnership is required. The system captures from modern trade, traditional trade, e-commerce, delivery apps, gas stations, cafes, vending machines, and pharmacies. Every channel, every format, one ingestion layer.
Parse
Arabic-native NLP built for MENA commerce.
Seven AI agents process the raw capture. Receipt OCR handles mixed Arabic-English text, abbreviated product names, and non-standard formatting across 15+ delivery platforms and hundreds of retailer receipt formats. Entity recognition resolves the same product written three different ways into one canonical record. The NLP stack was built from scratch for the Arabic-English spectrum of MENA retail. No open-source model and no commercial API handles this language mix at the accuracy required for panel-grade data.
Match
Semantic resolution against a 50,000-product knowledge graph.
Every parsed transaction is matched against a vector-embedded product knowledge graph containing 50,000+ FMCG products with brand, manufacturer, category, segment, variant, and pack-size attributes. Matching operates in 1,024-dimensional embedding space with HNSW indexing. Semantic similarity handles misspellings, transliterations, and abbreviations that exact-match systems fail on. The knowledge graph enriches with every transaction processed: new products are identified, classified, and added to the graph within 48 hours.
Weight
Statistical projection to national population.
Iterative Proportional Fitting (IPF) across multiple demographic dimensions: region, socioeconomic class, nationality, and household size. Calibrated against GASTAT census data and Ministry of Human Resources labor reports. Weight capping prevents over-representation. Effective sample size monitoring ensures statistical reliability. Multi-layer correction factors account for category-specific diary under-reporting, regional purchase behavior bias, and channel-specific capture rates. The weighting methodology was designed for the unique demographic structure of Saudi Arabia, not adapted from a Western market template.
Analyze
40 engines, every standard and advanced framework.
Weighted data flows into 40 composable analytical engines. Penetration decomposition. Loyalty segmentation. Brand switching dynamics. Growth attribution. Trial-and-repeat tracking. Heavy-Medium-Light buyer classification. Competitive share analysis. Demographic profiling. Channel migration. Each engine encodes methodology refined over a decade of FMCG measurement across three of the five largest consumer markets in MENA. The engines compose: any question can be filtered by channel, demographic, geography, and time period. Output is delivered through interactive dashboards, exportable presentations, or API.
VALIDATION
How we know the numbers are right.
Consumer intelligence is only valuable if the buyer trusts it. Trust requires transparency about both strengths and limitations.
Internal Consistency
Every pipeline run is checked for internal consistency: value shares sum to 100%, penetration rates fall within demographic plausibility bounds, weight distributions remain within capping thresholds, and effective sample sizes exceed minimum reliability standards. Anomaly detection flags statistical outliers before they reach any client-facing output.
Cross-Category Benchmarking
Category-level KPIs are benchmarked against published industry data where available: trade press reports, manufacturer annual results, and publicly disclosed market share data. When IntellX penetration or share estimates diverge from external benchmarks by more than 5 percentage points, we investigate the cause and document the explanation.
Panel Health Monitoring
Panel quality is monitored continuously across four dimensions: response rates (are panelists capturing regularly?), capture completeness (are they logging all purchases, not just large trips?), demographic stability (does the panel still represent the target universe?), and diary fatigue (are capture rates declining over time within individual panelists?). Attrition is tracked and replacement recruitment is triggered when any demographic stratum falls below minimum thresholds.
What we disclose.
No measurement system captures everything perfectly. Disclosing limitations is how we build the kind of trust that legacy providers avoid by simply never mentioning theirs.
Small categories with low penetration (<5%) produce estimates with wider confidence intervals. We flag these in every output.
Out-of-home capture depends on panelist compliance at the moment of consumption. Capture rates for impulse purchases (vending, counter snacks) are lower than for planned purchases (grocery, delivery).
Traditional trade pricing is less consistent than modern trade. Price imputation methods are documented per category.
Panel recruitment in remote regions (Northern, Southern provinces) is more difficult. Weighting corrects for under-representation but does not eliminate the underlying sample limitation.
The product knowledge graph covers 50,000+ FMCG products. Highly localized or artisanal products may require manual classification, which introduces a 24-48 hour delay.
Methodology matters. We make ours visible.
Request a technical briefing. We will walk your research team through the full pipeline, share our validation framework, and answer every question about how the numbers are produced.
Request a Methodology Briefing