Search intelligence: reading demand before the season
Search intelligence is the practice of reading what people are looking for — at scale, over years — to anticipate demand that has not shown up in sales data yet. It is not keyword research. Keyword research returns volumes for words; search intelligence builds themes, puts them on a multi-year timeline, and separates a cycle from a trend from noise.
We have been doing this work since 2015, for clients including Iper La Grande i, one of Italy’s large supermarket groups. Long enough to have been wrong a few times, which is the part that taught us the method.
Table of Contents
Why sales history is the wrong place to look
In large-scale retail the important decisions are taken months ahead: the December assortment is closed in summer. By the time a customer walks into the store, everything has been decided — and whoever decided it was looking at data that, by construction, looks backwards.
Sales history tells you what was bought. It does not tell you what people looked for and did not find, and it does not tell you what they will start looking for in three months.
That gap is the whole job. For Iper the service was contractual, with the pace written into it: topics agreed three months ahead of the period they covered. In practice the lead time was often longer — the cured meats study for Christmas 2019 is dated 2 August. Spending August on Christmas dinner is a peculiar way to run a summer, and it is when those decisions are actually made.
The team was led by Raffaella Roani, who runs the same line of work today.
What we actually looked at
Not sales. Searches.
For each theme we took every relevant query — hundreds of them — and rebuilt twenty-four months of monthly volume for each one. Then, and this is where the value sat, we aggregated them into semantic clusters: not “how many people search for one product name”, but how much weight sits on recipes, on nutrition, on provenance, on local festivals.
A single keyword is noise. A cluster with two years of history is a behaviour.
The product that only exists in December
August 2019, cured meats, Christmas season. One detail in that study is worth the entire method.
Italian prosciutto crudo is flat all year: between 4,400 and 6,600 searches a month, with no Christmas peak at all. People buy it constantly, and the demand has no season.
Spanish ham does the opposite:
| October 2018 | November 2018 | December 2018 | |
|---|---|---|---|
| Spanish ham | 880 | 6,600 | 4,400 |
| serrano ham | 210 | 390 | 880 |
| pata negra | 320 | 480 | 720 |
From 880 to 6,600 searches in thirty days. A product that for eleven months of the year barely exists in the market’s mind, and that becomes a subject in November. The ham, needless to say, was there all along. What moved was not the product: it was the moment people started thinking about it.
The part that is not in the numbers
This is where the article could end, with the tidy conclusion that you look at the history and put the peak in the calendar. That would be convenient — particularly for us, since we could sell it as a method — and it would be false.
Because the interesting part is not the peak. It is how it grew.
Christmas 2017: 1,600 searches. Christmas 2018: 6,600. Four times as much in a year. Not a stable season to note down and repeat annually — a trend forming while we were watching it.
And that is the actual craft. In any time series the oscillations are countless: holidays, weather, passing fashions, statistical noise. The machine shows all of them, with equal confidence. Telling which one is becoming a trend, and which will be gone by the following season, is a judgement built by watching a market move for years.
An algorithm cannot distinguish a peak that will return from a peak that is already over. It only knows there was a peak. The difference between those two things is the entire purchasing decision.
How to know the method is not inventing patterns
The right question, faced with a number that quadruples, is whether the method is measuring a real season or an artefact.
The same study contains the control. Prosciutto e melone — a summer dish — has exactly the inverted season: 3,600 searches in June, 70 in November. And Italian prosciutto crudo, as noted, does not move at all.
Three different behaviours, measured with the same instrument on the same archive: one summer, one Christmas, one flat. When a method correctly recognises what is not seasonal, what it says about seasonality starts to be worth something.
It is a principle that holds well beyond cured meat: a model that finds patterns everywhere is not finding patterns. That was true in 2019 and it is more relevant now, when the tools that find correlations have become very fast — and speed has never made a correlation true.
What has changed, and what has not
The value was never owning the data. Even then, anyone could open a keyword tool and look at volumes. The value was aggregating it into themes, putting it on a timeline, recognising which oscillations mattered — and above all delivering it while it was still useful, which means months before, not once the season had started.
Today that work is called Visirank® Demand, and Raffaella Roani still runs it. AI has made it considerably faster: what took days of manual aggregation now takes hours, across wider archives.
But the difficult part was never reading the data. It was deciding what to do with it, in time — and that remains a human decision, taken by somebody who answers for the outcome.
Frequently asked questions
What is search intelligence?
Search intelligence is the analysis of search demand — the queries people type — aggregated into themes and reconstructed over multi-year timelines, used to anticipate demand before it appears in sales figures. It is used for assortment planning, production planning, editorial calendars and seasonal campaigns.
How is search intelligence different from keyword research?
A keyword research tool returns volumes for individual words. Search intelligence builds thematic clusters, reconstructs multi-year time series and compares them against each other to separate cycle, trend and noise. The underlying data is public and available to everyone; the work is in what is done with it.
How many months of search history are needed to identify a season?
Twenty-four months is the minimum to distinguish a genuine season from an isolated event. Recognising a trend that is still forming requires more than that — and requires somebody who knows what they are looking at, because a time series contains dozens of oscillations and the software presents all of them with equal weight.
Does AI make human analysis of demand trends unnecessary?
It makes the slow part — aggregation and comparison — considerably faster. It does not replace the judgement about which oscillation matters: an algorithm cannot tell a peak that will return from one that is already over, and that is precisely where an error costs a season.
Does this only apply to grocery retail?
No. It applies to anyone who has to commit months in advance: assortment, manufacturing, editorial planning, seasonal campaigns. The logic is identical; only the themes change.
If you have to decide now something that will only be visible in three months, this is the moment to look at the data.
→ Talk to us about demand data
