How we pick the signals

This page explains, without jargon, how Marketing Pulse goes from thousands of public data points to the few signals you see every morning. It is written for people who run a small or mid-sized business and want to know what to trust and what not to. The examples are real, taken from data collected between 11 and 23 October 2026.

In short

1. Every morning the system collects fresh data from about thirty public sources.
2. An AI model gives each new headline a family: product, service, sector, event, sport, entertainment, current affairs or other.
3. Rules with fixed numbers decide what counts as a signal: for example "a product enters the Amazon top 10" or "searches grow by at least 30%". No made-up scores: either the threshold is passed or it isn't.
4. Out of the day's signals, the system picks at most 40 that matter to people who sell and hands them to a second model that writes the brief: three facts, each with something to do. An automatic check makes sure every number in the text really exists in the data; if one doesn't match, the brief is not published on its own.
5. The dashboard shows all of this exactly as it is written in the day's file: it recalculates nothing.

The market picker

At the top there is always a row of buttons: "All", "Global", "Italy", "Australia", "United States". You can switch on one or more.

• With one market on, you see only that market. With "Italy" every panel shows Italian data only: no rows from the United States, Australia or global ones mixed in. With "Global" you see only what applies to everyone, such as oil, exchange rates and shipping routes.
• With several markets on, the rows add up, and each one carries a small tag saying where it comes from: IT, US, AU or GLOBAL.
• "All" switches them all on. If you switch off the last market, you are back to "All".
• Everything follows your choice: the numbers at the top, the brief, the topics, ads, creative, Amazon, the calendar and the signal log. If the day's brief is not about the market you picked, you still see it, marked "global brief". The sector ranking is not split by market yet, and it says so.
• The map too switches markets on and off: click a market, double-click to go back to "All".
• Your choice is kept for the next visit and sits in the address (for example `#m=it,au`): send the link to someone and they will see the same markets.

1. Where the data comes from

All public sources. Some may be missing on a given day: when that happens, the day is put together with the ones available, without filling the gaps, and where it matters the dashboard shows the date of the latest data.

On 23 October the database held 8,549 headlines and measurements from the last 30 days.

2. What gets classified

An AI model reads new headlines in batches of 40 and, for each one, says:

• the family: product, service, sector, event, sport, entertainment, current affairs, other;
• whether it can be bought: yes or no (for example "ceramide cream" yes, "Juventus-Rennes" no);
• the topic in two or three words, about the specific case ("inter-shakhtar", not "football match"), so different headlines about the same case add up.

The answer is saved on the headline and never redone. If the model's answer is not clean, that batch stays without a family and is retried the next day: the system never makes up a family. Amazon products (they are already products), exchange rates, oil, ships and calendars are not classified.

On 23 October 2,219 headlines out of 8,549 had a family: 527 current affairs, 497 sport, 344 services, 283 entertainment, 232 sector, 176 product, 125 other, 35 events. Headlines not yet classified wait for the following days, in batches: every morning a share of them gets classified.

3. The signal rules

Every signal has a measure with its unit, a period and a source. Different units are never added together.

On 23 October most signals came from Amazon rankings (new entries in the top 10, price moves, climbs and exits); then attention spikes, upcoming official releases, topics talked about on several sources, two shipping straits and one oil move. The day's count is the number at the top of the dashboard.

3a. Two more readings: ads and creative

• Ad auction pressure. For each sector and market, the average price of a click on Google ads, in USD: sectors are listed from the most expensive, with the average of all sectors in that market as a reference, whether the sector sits above, below or in line and, for members, how much it moved since the week before. It is a market estimate, not what you would spend. If a sector has too few keywords with data, we say so instead of giving a number; the date of the data is always shown. How we measure it is explained below, with every term.
• Video formats that are working. Every day we look at two groups of TikTok videos in Italy, the United States and Australia: a sample of what TikTok shows in the "For You" feed (it isn't a chart: TikTok doesn't publish one) and the videos found with a few searches per sector, in the country's language (for example "dietro le quinte ristorante" or "ricetta veloce" for restaurants in Italy; eight searches a day per country, in rotation). We keep videos posted in the last 14 days and rank them by speed, that is views per hour since they went up. A model watches up to 14 videos per country a day, from the feed and from the searches, and describes how they are built (what happens in the first seconds, the steps, the text on screen, the voice or sound, the pace, who appears); then it groups them by how they are made, not by subject: the format is the structure of the video (for example "POV", "storytime", "sketch with a recurring situation", "before and after", "behind the scenes", "story with a twist ending"), even when the subject is sport, music or entertainment. A one-off video tied to a person or an event is not a format: a format only counts with at least 3 watched videos from at least 2 different creators with the same structure (with 2 watched and one inferred from the text we say so: "partly watched"); videos that weren't watched only serve as confirmation, and the name and description can't contain advice or numbers that aren't in the videos. Each format has a label, with one line explaining why: fits a business (a shop, a restaurant, a studio can redo it as it is), adaptable (the structure holds if you change the subject or tone) or creator-led (it works with a person putting themselves on camera). Each format says what you need to redo it, where the videos come from ("For You" feed or searches, with the sector) and whether the model watched the videos or inferred the format from the text (caption, hashtags and sound), when video analysis is off. Then we check every example: if one doesn't have the format's structure, the format is dropped. Video counts, views, speed and countries are counted by us. For each format, three videos to open on TikTok. Below, TikTok's viral sounds (Top 50 and Viral 50), which have no country. A format that grows fast (usually at least 5,000 views an hour, in at least 2 countries or from at least 3 creators) becomes a signal. TikTok hashtags are in "What people are talking about": they tell you what people talk about, not how a video is made.

3b. How we measure ad pressure

On the dashboard it is called "Ad auction pressure". The price of a click on Google ads (the CPC) is set in an auction: when more companies want to show up for the same searches, the price goes up; when there are fewer, it goes down. That is why a sector's average CPC tells you how much competition there is, right now, to get found by customers.

• One rule for every sector. For each sector and each market we take 4 terms: the service or product written the way a customer searches for it (in Italian for Italy, in English for the United States and Australia). To each we add 2 buying words for that type of sector, such as "prices" or "near me": 8 keywords per sector and market, 360 in all. No keyword is picked by hand.
• The average CPC. It is the average CPC of the 8 keywords weighted by monthly searches: a keyword few people search for counts little. Next to it we show the median. The value is in USD and is a market estimate, not what you would spend on your own campaign.
• The comparison. Each sector is compared with the average of all sectors in the same market: above, below or in line (within 10%). Global is the average of the three markets. A sector with fewer than 4 keywords with data is not ranked and we write so: "not enough data".
• The auction signal. Keywords are updated once a week, on Monday. Between one collection and the next, if a sector's average CPC in a market moves by 15% or more, an "Ad auction pressure" signal appears: going up, more companies compete for the same ads; going down, there is less competition. If the CPC holds still but searches move by 25% or more, the signal is about demand. At the top of the dashboard the "sectors with rising pressure" number counts the day's rising signals.

Buying words, by type of sector

The terms for each sector

Terms for Italy stay in Italian: they are the words Italian customers type.

Keywords are the term with its two buying words: for example "lawyer near me" in the United States, "avvocato prezzi" in Italy. On the dashboard the "How we collect" button shows them all, sector by sector, with how many had data in the latest collection.

Terms and keywords can be tailored to your company's services and markets in the Pro version.

3c. How we tell events from trends

Many things are talked about for one day only: a match, an award, a film release, a storm, the death of a famous person. Those are events: they start and end there. In the "What people are talking about" panel they sit apart, in the "Today's events" row, and we never call them trends.

• A trend is a topic that has been growing for at least 3 days in a row on at least 2 different sources (for example Google searches and TikTok videos). If it also grows in the way people live, not only in what they buy, we call it a cultural shift.
• For each trend we say how many days it has been growing, how big it is on each source, in words, and for how many hours we have seen it. Growth as a percentage only appears when it can really be measured.
• For each trend we say what people are saying: two or three sentences summing up the headlines about it, and below them the real headlines with source and date. We don't give advice and we don't write posts for you. If there are fewer than two headlines on the topic, we say so instead of summing up; on sensitive topics (health, deaths) there is no summary.
• Sensitive events, such as a death, are shown in grey with the "sensitive" notice.
• When no topic has 3 days of growth on 2 sources yet, the panel says so: "No new trends today". It means the system looked and found nothing above the threshold, not that it stopped.
• Emerging themes written by the model are in the brief and, for members, in the signal log: on the dashboard the conversation is all in "What people are talking about".

3d. How we measure sector growth

The "Who's growing, who's slowing" panel looks at the 15 sectors of the ad auction pressure in Italy, the United States and Australia, with the same keywords (you'll find them above, sector by sector).

• What we count. How many times a month people search for the sector's keywords. The current month doesn't count, because it isn't over: the top of the panel always says which month the data runs to.
• The comparison. The last three full months against the same three months a year earlier, relative to the market basket: the big number is the gap in points between the sector and all the keywords of that country. Google Ads searches are rounded into steps of about 20%, and in some years all the keywords of a market move up or down a step together: that is why the gap from the basket counts, not the change on its own. Keywords with an estimate jump (more than triple or less than a third in a year) are left out, and we say so. The year before comes from a second series, monthly and four years long, never mixed with the weekly one; where it isn't there yet, the comparison is between the last three months and the three before, and the panel says which markets those are.
• Growing, slowing, steady. Above +5% the sector is growing, below −5% it is slowing, in between it is steady.
• How much it agrees. Next to each sector we say how many of its keywords move the same way: if only one keyword out of eight grows, the number counts for less.
• Little data. If too few keywords have enough months of history, the sector is not ranked and you see it in grey, with "few searches".
• Several markets. With several markets on (or with "Global" and "All") each sector is the average of the countries, and the panel says so.
• What we don't use here. No Amazon rankings and no popularity indexes: only searches for the basket keywords.

3e. Seasons ahead and the year before

• Two search series, never mixed. Basket searches are updated every week with 12 months of history. A second series, once a month, brings 4 years of monthly searches for the same keywords: it is used only for the year before, for seasons over several years and to confirm jumps.
• A season goes up and then comes down. A rise of at least 45% on the average of the two months before, lasting at least a month and then returning close to the earlier level. If it stays up, it is a change of level and we don't call it a season.
• One year isn't enough to say "every year". With 12 months we say "last year, in the same month, it went up"; with four years we say in how many years it happened and how much it usually rises.
• Jumps. Searches are estimates rounded into steps and sometimes jump for no reason: from +200% up we don't use them, unless the second series shows the same jump in the same month (or the next one), in the same direction and within a factor of 3. Then we write it: "confirmed by a second series".
• The year before. By sector in "Who's growing, who's slowing" (the gap from the basket); by keyword in "What people are searching for", only when the keyword pulls away from its country's basket by at least two steps. Labels on the rise of the second-to-last month, only if the last month confirms it holds and the keyword rises more than the basket: "rising as every year", "rising more than usual" (more than in every year before), "early" (in other years it rose the month after).
• What we don't say. We don't tell you what to do: the line says what happened and when the month comes back.

4. How the brief and the topics are chosen

The brief has up to three facts. To write them the model receives at most 40 signals, picked like this:

1. out go sport and current affairs on topics that can't be bought;
2. exchange rates, oil, ships and calendars take at most one place in total: they matter, but a small business can do little about them in a day;
3. since 25/10 weight counts: first the signals with a real base and a readable comparison, that is buying demand with a long series (rising searches, seasons ahead, confirmed jumps), then Amazon rankings and prices, then video formats the model watched, then exchange rates, oil and the other measures with a base; last, as context, themes, trends and mentions. A small count ("8 mentions in 48 hours against 3", a few dozen searches) never opens a fact and never becomes the big number in the video;
4. every type of signal gets at least one place, so the brief doesn't talk only about Amazon.

The brief has as many facts as strong signals, three at most. A fact opened by a weak signal when a strong one was available turns the check red.

For each fact the model writes what happened and one thing to do. Then an automatic check compares every number in the text with the data: if a number isn't there or points the wrong way, the brief stays a draft and only the console sees it. On 23/10 the published brief was: "Isopropanol +87.3%, Chico Natale no. 1 in Books, gloves no. 1".

5. What gets dropped, and why

• Sport, entertainment and current affairs: they stay on the dashboard ("What people are talking about today", because they show where attention is going) but they don't become "talked about" or "rising news" and they don't enter the brief, unless the topic can be bought. On the first day, without this filter, the 9 topics "people talk about" were all sport and celebrities.
• Google's ranges: "over 500,000 searches" is a minimum stated by Google, not a count. It is used to put spikes in order, never to say that something is growing.
• Old rankings: an Amazon ranking older than 36 hours gives no signals; the first reading of a category has nothing to compare with, so no entries or climbs.
• Headlines without a family: if the classified share is less than half, "talked about" and "rising news" are not calculated and the system says so.
• In emerging themes, headlines in non-Latin alphabets (for example news in Chinese or Cyrillic) and, everywhere, mentions older than the week before: an old article read today is not today's news.

6. The limits, said plainly

• A ranking position is not sales data: Amazon shows the order, not how many units.
• Google searches are indexes and ranges, not exact volumes.
• Classification is done by a model: it can get a family wrong. It can be corrected by hand from the console, and the correction wins.
• Some sources may be missing on a given day: their panels then stop at the latest data; where it matters, the dashboard shows its date. On 23/10 Google Trends and monthly search volumes were stuck at 16-17 October: the sector ranking says so on every row ("data up to…") and gives less weight to stale data.
• The brief is written by a model: the numbers are checked, the "what to do" is a suggestion to weigh against your own experience.
• The public dashboard shows a part: the top signals by country and type, 7 days of curves. With a free account you see everything, with 30 days.

If a panel has no real data for the filter you picked, it doesn't appear: we prefer a shorter dashboard to a made-up number.