A brand manager opens a dashboard and sees what consumers in five countries started talking about last week — before the trend reaches a shelf. A route planner sees which neighbourhood freezer will sell more next month. A planner knows how much chicken the market will have in the next quarter. Those are the three intelligence systems I built inside Pronaca's marketing function.
Consumer insight arrived as an annual study — accurate, expensive and already historical by the time it landed. Campaign tracking told us what had happened to our own advertising, which is a different question entirely. Nobody could answer the one that mattered to a food company: what is the consumer starting to want, before it shows up in our sales?
So we built the answer instead of buying it. Two systems, two different kinds of signal: one linguistic, one geographic.
A monitoring platform for digital consumer information, built on text mining and text analytics with natural-language-processing algorithms, designed to anticipate consumption trends rather than report on them.
topics under continuous monitoring
Spanish-speaking markets: Ecuador, Colombia, Peru, Mexico, Spain
of historical data brought into the model
source families: social, RSS, blogs and open web
Not campaign tracking. Not advertising tracking. The distinction was written into the scope on purpose — the moment a listening platform is judged on whether it flatters the last campaign, it stops being intelligence and becomes reporting.
Ingest social networks, web, blogs and news across the five markets.
Transform and enrich the text so the machine can work on it — the unglamorous step where most projects quietly fail.
Extract the key words and read the meaning around them, using dictionaries — word associations built for the category.
AI models over text, not over numbers.
Consumption-trend insights, ranked.
Data storytelling — let the data speak to a committee that has twenty minutes and a budget to allocate.
Natural language processing is the set of computing and AI techniques that analyse human language, to extract the information and knowledge held inside documents. Take three words:
Multiply that by forty topics, five countries and two years of history, and you stop guessing what the category is about to do.
The platform's job was to surface which conversations were gaining interest before they peaked, so the content calendar could move ahead of the trend instead of reacting to it. A food brand that arrives second to a consumption trend is paying full price for somebody else's momentum. Arriving first is the difference between setting the conversation and buying into it.
Architecture: sources feed a cloud data lake; analytics runs across cloud and on-premise. The phasing mattered as much as the technology — a listening platform that arrives complete in month nine is a platform nobody has learned to use.
Ecuador's protein business does not run through supermarkets alone. It runs through thousands of small neighbourhood shops with a company freezer in them. We had the sell-out data and almost no geography on it.
How much the point actually moves.
Maturity of the outlet, not just its size.
Traffic, independent of ticket.
What one shopper is worth in that neighbourhood.
Freezer master data plus sell-out per point.
ETL — addresses become coordinates, duplicates collapse, dead points drop out.
Clustering across the four variables, projected onto maps.
Results good enough to act on? If not, the loop runs again before anyone decides anything.
Where the next freezer goes, which route is under-served, which cluster deserves trade investment.
The control gate is the part worth keeping. A geo model that always produces a map always gets believed; one that is allowed to say "not yet" is the one you can plan against.
Poultry and pork are the core of the business and the least forgiving thing to forecast: a live product, a short shelf life and demand that moves with weather, holidays and price. Getting the prediction wrong costs money in both directions — waste on one side, empty shelf on the other.
Data pipelines built to pull, clean and reconcile the sources before anybody trained anything. The unglamorous half of the project, and the half that decides whether the other half works.
Machine learning applied to chicken and pork specifically — the categories where the portfolio holds roughly 40% and 48% brand power, and where a forecasting error is immediately visible in the P&L.
The reason this sat with marketing at all: promotion, price and campaign timing are inputs to the forecast, not noise around it. A prediction model that cannot see next month's promotional plan is guessing.
This project and the churn model share one rule I applied to every analytics investment: quality of data before volume of data, a working prototype in three months, and a named business decision the model has to change. A model nobody acts on is an expensive opinion.
Machines read the open web; people answer questions the web never asks. Running in parallel with the platform, four survey waves in five weeks profiled the community to identify consumer perception, habits and preferences.
Wave 1 · 27 January
Wave 2 · 3 February
Wave 3 · 11 February
Wave 4 · 15 February
of respondents outside the brand community named the company spontaneously — unaided recall, across a total of 6,195 people profiled.
The trends model moved the conversation in the marketing committee from "how did the campaign perform" to "what is the category becoming". The geo model moved trade investment from intuition about neighbourhoods to clusters with sales, maturity, traffic and ticket attached.
Both were built to the same standard I hold anywhere: a model that predicts is unfinished. It has to become a decision, and the decision has to become cash.
What consumers say online, clustered by topic, feeding content and product.
Demand forecast, location model and trend model, run by your team.
Continuous intelligence, at a fraction of the cost of a panel.