To get there, start in Ancona, take the highway to Rome, exit at Jesi. It feels like you're going into nowhere. In fact, you are going into nowhere. The valley of San Clemente, a fraction of Angeli di Rosora, a municipality of fewer than two thousand inhabitants nestled between hills that in spring smell of must and in September smell of grapes.

In that nowhere, in the middle of the valley, there's a factory. It doesn't look like a factory. It looks like a European university campus, with buildings scattered among the fields, large windows, a river — the Esino — that runs alongside and on which, at some point, someone installed four small hydroelectric plants.

That factory is the **Loccioni Group**. Four hundred and fifty employees, average age 34, half university-educated. They measure and control the quality of industrial products for the largest clients in the world — automotive, pharmaceutical, energy, aerospace. Every airbag that inflates, every insulin injector that doses the right amount, every engine that starts on a winter morning, has probably passed a test bearing their signature somewhere.

They are world leaders. And almost nobody, in Italy, knows their name.

This is the first installment of **#BraveHistories** of 2026, and we dedicate it to them for a precise reason: in an era where branding is measured in millions of euros of ads and CMOs that come and go every 18 months, Loccioni built a global brand **without almost ever doing advertising**. And they did it in a way that anyone who deals with branding should study by heart.

## 1968, the stable, the electric pump

To understand why their brand works, you need to know where they come from.

Enrico Loccioni — the founder — was born in a country house above a stable, in a valley that at the time still had no electricity or running water. He was fifteen when he received his first paid job: connecting an electric pump to bring water to a neighbor's animals.

That job is the literal prelude to everything. The prelude of the methodology (understanding a concrete problem, finding a technical solution), and the prelude of the identity (a technical company at the service of those who work the land, before that of industry).

In 1968, together with his wife Graziella Rebichini, he founds **I.C.I.E. — Impresa di Costruzione di Impianti Elettrici** (Electrical Plant Construction Company). It's an artisan company, making electrical panels for the workshops of the valley.

Fifty-eight years later, it is the **Loccioni Group**, world leader in industrial measurement systems, with clients in twenty countries.

Between one and the other, no marketing campaign. No celebratory rebrand. No hiring campaign on LinkedIn with *We are hiring* style photos.

How did they do it?

## Sharecropping as a corporate method

To answer, you have to understand something that is obvious to a Marche valley and a bit less so to the rest of Italy.

The culture of **sharecropping** — the agricultural system that shaped the Marche, Tuscany, Umbria for centuries — was not just an agricultural contract. It was a *shared risk model*. The sharecropper divided the harvest with the landowner. Good if the harvest was good, bad if it was bad, and in both cases everything depended on the season, the climate, the art of reading time and surviving uncertainty.

From that culture, Enrico Loccioni took something he put at the center of the company: **diversify risk without waiting for the next season**. If you're a sharecropper, the only way to survive is to have more plots, more crops, more sources of income. The same applies to a manufacturing company: not to depend on one customer, one sector, one trend, you must diversify skills, sectors, geographies.

Loccioni did it. In the '80s they worked for white goods (appliances). In the '90s they entered automotive. In the 2000s pharmaceuticals. In the last twenty years they expanded into renewable energy, electric mobility, water infrastructure.

Each new sector wasn't a *pivot* — startup term that would sound ridiculous here — it was an *additional plot*, according to a peasant logic applied to Industry 4.0.

## "Knowledge Company": their brand is inside, not outside

If you enter the Loccioni site, one of the first things you read is that they define themselves as a *Knowledge Company*. It's not a conference formula. It's a precise organizational architecture.

In a *Knowledge Company* (their version), **every employee is an entrepreneur**. Decisions are made as close as possible to expertise. Hierarchies are light. Relationships are based on trust. The company is a *school of competencies* — an expression they use literally — that trains people, in addition to producing products.

What does this have to do with branding, you might ask.

Everything. Loccioni's brand — the one that allowed them to become a world leader without spending on advertising — **doesn't live in advertisements**. It lives in the people who worked there, and who today go to conferences, write in newspapers, teach at universities, do consulting, get hired by customers. It's a brand that spreads *by emanation*: you work with Loccioni, you *are* Loccioni, wherever you go for the rest of your career.

In modern branding terms, it's an *employer brand* that has cannibalized the commercial brand in a positive way. The most powerful thing you can do. And for years it was debated whether it was an accident or a strategy. Going in person to Angeli di Rosora, you quickly convince yourself that it is a strategy. Perfectly conscious. Perfectly executed.

## "Impresa bene comune": when territory is also the product

Here comes the second axis, and for us it's the most contemporary.

Loccioni did a rare thing: they publicly declared that their **mission includes the territory**. Not in a rhetorical sense (local charity, sponsorships of the five-a-side tournament). In an operational sense.

Twenty years ago, they took two kilometers of the Esino river that runs near the headquarters under concession. They built four small hydroelectric plants on it. Today they produce more clean energy than they consume. They have an internal fleet of electric vehicles. They have sixteen *nomadic* technological laboratories powered by clean energy. They launched a project — *LOV, Land Of Values* — aimed at enhancing the Marche territory as a strategic company asset, not as a philanthropic hobby.

For a company that measures *quality control* of industrial products, it's an extraordinary branding move. Because it says something very specific to the market: "We *really* measure things, even those that hurt if you don't measure them — environmental impact, consumption, sustainability". It's positioning coherence applied to the business model.

Translated: **the brand isn't what you declare. The brand is what you do, told by those around you.** Loccioni understood this principle long before most Italian agencies.

## What they did NOT do

For methodological honesty, it's worth listing also the things Loccioni *avoided*.

**They never bet on the founder's name as an icon.** Enrico Loccioni is tellable, has a novel-worthy story, is the type of founder who on American social media would become a motivation meme. They deliberately kept him out of the center of the commercial brand. What exists is the *Group* Loccioni, not *Enrico* Loccioni.

**They never did a "big" rebrand.** The logo has been almost the same for decades. The forms change, but by *slow erosion* — as we saw in the May 21 piece on the [silent rebrand](/blog/silent-rebrand-5-italian-smes).

**They never sold the "Loccioni story" to international media in a tacky way.** When they appear in Sole 24 Ore or Financial Times, it's because someone sought them out. Not because they sent a press release.

**They never opened to private equity funds.** They've remained family-owned. This has taken away growth speed, but given them identity continuity. For a silent brand, it's what matters most.

## What we take home

Three lessons, one per quadrant:

**For startup founders.** Don't dream of marketing that explodes. Dream of *word-of-mouth that explodes*. Loccioni's brand lives in employees who become ambassadors, in customers who become case studies, in territories that become showcases. Build the organization to generate this word-of-mouth, don't expect your CMO to generate it.

**For those who do branding (and for us who do it for a living).** Brand isn't made in manuals. It's made in operational decisions: how you treat employees, where you put investments, in which territory you choose to be, what you refuse to do. A good brand manual reflects those decisions — it doesn't replace them.

**For Italian B2B SMEs.** Your competitive advantage is not the most modern graphic. It's the deep coherence between what you declare, what you do, and what others say about you. Loccioni has forty to fifty years of advantage on this axis, but forty to fifty years are built one at a time, starting today.

There's a final detail that's worth the whole piece. When you show up in Angeli di Rosora — and you should, at least once — you don't see a brand store, you don't see a marble reception hall, you don't see signs with values written big. You see people working. You see a river with four turbines in it. You see a parking lot with electric cars.

The brand is exactly that. There's nothing else.

And almost nobody knows it, in Italy. It's time to remedy.

Thursday, June 11 we go back to the AI front with a piece we've needed for a while: *The prompt isn't the problem. It's the context.* What's changed in seven months building BP and Tuken, and why 2024's prompt engineering is the right discipline solving the wrong problem.

---

**Do you have a brand the market doesn't see yet?** We work with Italian B2B companies on [substantial, long-term branding strategies](/branding-su-misura-strategia-marketing-aziendale). No plot twists, plenty of consistency. [Let's talk](/contatti).

In Angeli di Rosora, province of Ancona, a Marche family built an industrial measurement multinational without almost ever advertising. What their silent brand teaches us — and why we don't know it well enough in Italy.

#BraveHistories: Loccioni, the silent brand of 450 people measuring the future of industry

Imagine a boardroom. A vendor is presenting you a quote for "an AI solution on your corporate documentation". At some point they say a sentence like: *"We'll use latest-generation embeddings to index your documents in a 1,536-dimension vector space."*

You nod. None of the ten people in the room asks what it means. The word "embedding" passes, settles into a seventy-thousand-euro budget, and goes into production.

Six months later, the internal company search engine answers "nothing found" to the query "August vacation", and nobody knows why.

Welcome to the second episode of **#HardWords**, our series that explains in ten minutes the tech terms vendors sell you without ever telling you what they really are. Today, finally: *embeddings*.

## The metaphor that works

An embedding is the **translation of a sentence into coordinates.**

Imagine a map. Not of streets — of meaning. In this map, every concept, every sentence, every paragraph, every image, has a position. *Vacation* sits near *holiday*. *Rome* sits near *Italy*. *October* sits near *November*. *Father* sits near *fatherhood* (but also, surprisingly, near *man* and *son*).

This map doesn't have two dimensions like a street map. It has **thousands**. Literally — modern models use embeddings of 768, 1,536, or even 3,072 dimensions. Don't try to imagine them: no human can. Just know that in that multidimensional map, every sentence is a point, and *the closeness between points corresponds to closeness in meaning*.

That's all. An embedding is the coordinate vector of a sentence on that map.

## Why it exists

It exists because computers have a problem: they don't understand the meaning of words. They can compare letters, they can count. They don't know if "vacation" and "holiday" are synonyms.

For decades, we tried to solve the problem with vocabularies, grammar rules, ontologies. It worked poorly. Then, around 2013, someone did this thing: instead of teaching the computer what words *mean*, we made it read billions of texts and told it "find a way to represent them as numbers *such that* words that appear in similar contexts have similar numbers."

Result: the computer built the multidimensional map by itself. It doesn't know what vacation is. It knows that the word "vacation" statistically reminds it of "holiday", "August", "long weekend", and that's enough.

This is the trick. Everything we call "AI that understands language" — chatbots, semantic search engines, summarizers, translators — runs on top of this idea.

## What it's for in your company

In practice you need embeddings every time you have to search by **meaning instead of by exact words**.

**Case one: internal search.** Your employees search for "travel expense reimbursement" and don't find the document titled "Travel Expense Policy". Embeddings: the two texts are close in the space of meaning. They find the right one even if the words are different.

**Case two: conversational assistant on company documents.** A customer asks your website chatbot "do you ship to Sicily?". The chatbot doesn't have a specific page on this. But it has a generic logistics policy. Embeddings: the two things are close. The answer arrives.

**Case three: recommendations.** You have an e-commerce with three thousand products. A customer buys an item. Embeddings: the products closest to the one purchased (in the space of descriptions, images, tags) get proposed.

**Case four: clustering / segmentation.** You have two thousand support tickets from the last month. You want to understand the 5 most recurring problems. Embeddings: group nearby tickets, and clusters emerge by themselves without having to write rules.

In all these cases, the "old" alternative was: write rules, maintain them, update them every time the vocabulary changes. Embeddings do all this almost automatically. They cost you computing a vector for each piece of content, and saving it in a vector database (yes, that's another hard word; we'll talk about it later).

## When you should ask for them

Once you know what they are, the question to ask in the boardroom changes. Don't ask "do you use latest-generation embeddings?" — it's the wrong question, because anyone will answer yes. Ask:

- **Which embedding model do you use?** (Names you'll hear: `text-embedding-3-large` from OpenAI, `voyage-3` from Voyage, `embed-v3` from Cohere, open-source models like `bge-m3`.)
- **How many dimensions?** (More dimensions = more nuance, but also more cost. For most business cases, 768 to 1,536 is fine.)
- **How do you handle updates?** (If you add a new document, it needs to be re-embedded. If the embedding model changes, *everything* must be re-embedded. Who pays? Who monitors?)
- **What language was it trained on?** (Many embeddings are excellent on English and mediocre on Italian. If your documentation is in Italian, this is problem number one.)

If the vendor hesitates on one of these questions, they're selling you an idea, not a solution.

## When you should be wary

Three red flags, to hang on the wall:

**One**, someone tells you that "embeddings solve the problem". They don't solve it. They are *one layer* of the solution. On top there's always something else: how you split the documents (chunking), how you handle context (RAG, which we'll talk about), how you present the result (interface).

**Two**, they propose to "embed everything and then we'll see". It's the equivalent of "let's dump all our documentation into the database, and the AI will do the rest". Doesn't work. Embeddings without a thoughtful chunking strategy produce random results.

**Three**, someone tells you that "your embedding will be customized for your company". Almost always false. What's customized is *what* you embed (your documents), not *how* (the embedding model is a market standard). Fine-tuning an embedding model is possible but rare, expensive, and makes sense only in very specific cases.

## Summary in three lines

An embedding is a translation of a piece of language into a vector of numbers.

Things with similar meaning end up close in this numerical space.

Everything that does "search by meaning" under the hood does this.

Now, in the boardroom, you know what to nod to and what to challenge.

Monday, June 8, we return to the **#BraveHistories** series, with an Italian case that is an underrated B2B like few: a Marche company of precision measurements that built a silent brand without spending a euro on advertising.

---

**Are they proposing to you "an AI solution on your company documents"?** Under [artificial intelligence for business](/intelligenza-artificiale-per-il-business) we also audit technical quotes, in half a day, free until the end of June. [Let's talk](/contatti).

What a star map has to do with your company's internal search engine. The tech word that most often appears in quotes you receive and that almost no one really explains to you.

#HardWords: Embedding

In September 2024, Paul Graham published an essay on his personal site called *Founder Mode*. Just under two thousand words. Simple language. The argument was this: founders of startups are taught to run a company as if they were professional CEOs — delegate, trust, let go — and this ruins them, because founders have a knowledge of the product and a decision speed that no external manager can replicate.

For a month, the piece occupied public discourse in the global startup ecosystem. It was cited, rewritten, parodied, demolished. It had its backlash, its backlash-of-the-backlash, and finally its place as a *flag-concept* of an era.

Then 2025 happened.

Last year startups stopped making the news for their ideas and started making the news for their layoffs. The venture capital industry restructured. Three of the most-loved Italian scaleups ended up in emergency acquisitions. Another six died in silence. Those that survived, we saw again at the Italian Tech Week last October, and they told us a truth that's worth writing down.

The *founder mode* Paul Graham talked about in 2024, in 2026, in Italy, is dead. At least in that form.

And it's precisely for this reason that it becomes interesting again.

## What the essay really said (net of the parody)

To orient ourselves, it's worth remembering what Graham said net of the meme.

The essay was born from a conversation with Brian Chesky, founder of Airbnb. Chesky recounted that, after being pushed toward a *manager mode* model (delegate to teams, trust middle levels, climb up), his company had started losing energy. When he'd returned to *founder mode* — direct involvement on product decisions, small meetings, communication that skips levels — the company had started moving again.

Graham generalized. He said that the difference between a founded company and a managed company is structural. Founders aren't better CEOs, they're *different* CEOs. Replacing a founder with a professional CEO is almost always the moment a company stops being itself.

The essay was brilliant. It was also, reading it today, slightly romantic.

## What happened after 2025

The Italian startups that survived 2025 — a year in which Italian Series A rounds collapsed by about a third compared to 2023, and in which Italian *late-stage* practically stopped — aren't those that best applied *founder mode*. They are those that understood that *founder mode* wasn't a pose, it was an *organ of sense*.

Let me explain.

In 2024, around Graham's essay, a kind of mythology had spread: the founder who shows up at the senior team meeting and makes decisions in 15 minutes, the founder who asks to see operational metrics every Friday, the founder who skips six hierarchical levels to talk directly to a developer. Everyone wanted to do it. Everyone thought *that* was *founder mode*.

It was a caricature. And in 2025, under real pressure, it broke.

Founders who did *founder mode* as caricature — micromanagement dressed up as vision — burned their senior team. Capable people left. When the liquidity squeeze came, there was no one experienced in the company anymore to handle it. Game over.

Founders who did *founder mode* as **competence** are still here. They knew which decisions they had to make (product, positioning, strategic go/no-go) and which they didn't (compensation, HR processes, daily operational management). They knew when to skip a level and when, instead, to trust their VP. They knew how to recognize their areas of differential value and disarm the temptation to stick their nose in everywhere.

In other words: the original *founder mode* was a **diagnosis**, not a **prescription**. It was saying "if you feel like a manager, you have a problem". It wasn't saying "behave this way".

## What we've seen

We see the Italian startup ecosystem very closely — both as an agency, and now that we ourselves are a young product company with [OpificioAI](https://opificio.ai), Tuken, and three products in the pipeline. Allow us three pointed observations, gathered in the field.

**First observation.** Italian founders who held up in 2025 have one thing in common that is the opposite of the cliché. They're not control freaks. They're *surgical interventionists*. They know how to spend days outside the meeting room, months without intervening on a team, and then drop on a specific decision with total intensity, decide, and go back to doing other things. Their *founder mode* is a mode that turns on and off. It's not a tone.

**Second observation.** None of the strong founders of 2026 define themselves as *founder mode founders*. They stopped talking about it. When you meet those who do talk about it, usually, it's because they're using it as an excuse to justify things that aren't working.

**Third observation.** The opposite mistake — *manager mode* disguised as maturity — is equally widespread and equally lethal. Founders who, under pressure, stopped making product decisions, delegated the pitch to heads-of, dropped the presidency over brand identity to a new VP Marketing recruited from abroad. Same result: the senior team progressively substituted for them, and the product stopped resembling their head. Six months later, investors say "we don't understand what you want to do anymore". Right.

## The honest point on the Italian ecosystem

The *founder mode* discourse in Italy is, by the way, a slightly luxury discourse. To apply it, you need to have a company with at least fifty people and a formed middle-management team. Most of our scaleups aren't there. They're still at the previous step — the one where the founder *is* the company, and the only problem is figuring out how to stop being it *without killing it*.

For family SMEs that are transforming into *tech-enabled companies* — a trend we see growing — the problem is even different: there the *founder* is often the father or mother, and *founder mode* isn't discussed, *it's endured*. The real problem is not whether to apply it. It's **how to transition it to a new generation without the transition coinciding with the end of the company**. It's a huge, very Italian, under-told issue. We'll come back to it in another piece.

## The new diagnosis

If the *founder mode* of 2024 was a diagnosis ("you're not a manager, stop pretending"), the *founder mode* of 2026 is a finer diagnosis: **can you distinguish decisions that change the nature of the company from those that manage it?**

The first are yours. The second aren't.

The first are few, brutal, counterintuitive, and almost always include a "no" said to someone very capable who's asking you for a promotion, an investment, a different course. The second are many, repetitive, need processes, and almost always require you to stay quiet.

Confusing the two is the most expensive mistake of a founder. Doing it in both directions is symmetrically disastrous: those who treat process decisions as nature decisions paralyze the company; those who treat nature decisions as process decisions lose it.

The mature *founder mode* is this **ability to distinguish**. It's an art. You don't learn it from a blog post. You learn it by having lost something.

If you're lucky, 2025 taught you. If you're not, 2026 will.

Thursday, June 4, we go back to the tech glossary with a new episode of our **#ParoleDifficili** series. Topic: *embedding*. What a star map has to do with your company's search engine — spoiler, a lot.

---

**Are you navigating a scale or ownership transition in your company?** We help Italian startups and SMEs [define brand identity and strategy in transformation phases](/branding-su-misura-strategia-marketing-aziendale). [Let's talk](/contatti).

September 2024: Paul Graham writes a three-thousand-word essay that splits the startup ecosystem in two. Two years later, Italian startups that survived 2025 tell us a stranger truth than the one we expected.

Founder mode is dead. Long live founder mode.

Eight months ago, one night in September 2025, an agent we'd put into production on an internal prototype did a very interesting thing. It sent a confirmation email to a test client, read the reply, interpreted "Okay, confirm!" as "execute everything", and deleted an entire configuration table.

It wasn't a real client. It wasn't a production table. But it was close enough to the real thing that the next morning we sat around a table and made a very honest list of everything we'd done wrong.

That list grew over the next seven months. Meanwhile we put Tuken into production — the events and parking management platform — and we started building BP (the business plan assistant, launching in the coming months). Each of these products added a chapter to that list.

Today we pass it on. Four mistakes. They're not the sexiest. They're the most frequent, the least talked about, and the ones that really make the difference between an agent that runs well and one that produces silent disasters.

## Mistake #1 — Giving it too many powers because "we'll manage later"

The first agent you put into production, whatever framework you use, makes you want to configure it with broad permissions. Read everywhere, write almost everywhere, access to three or four external APIs, ability to send emails, ability to call payment tools in "test mode that's actually the same as production".

The logic sounds reasonable: *"Let's see what it can do, then we'll restrict."*

You never see. You never restrict.

What actually happens is that after two weeks you're used to seeing it work with those powers, you've built flows that take them for granted, and at that point restricting them means breaking them. So you don't restrict. And a month later the agent has access to half the system, nobody remembers exactly which half, and the only person who knew has resigned.

**The principle we apply now, on all agents, on Tuken as on BP:**

The agent starts with the minimum required to *not* do its job. Yes, you read that right: it must fail the first time you try it. Then you grant a permission, you try it, and only if it's really needed do you leave it on. Each new permission has an expiration date written clearly: if at that date nobody remembers why it has it, you take it away.

It's annoying. It costs you an extra day of work per agent. It saves you from a number of disasters that's hard to quantify because — beauty of avoided incidents — they don't happen.

## Mistake #2 — Nobody knows what changed in the prompt the day before yesterday

Prompts are not code. You know it. I know it. We told each other.

But then this happens. A colleague changes a line in the system prompt of an agent because "it was answering badly to a specific client's question". It works better for that client. It works worse for the other five. Nobody notices for three days. When we notice, nobody remembers exactly what changed, when, and why.

The subtlest bug of all time is the bug *you don't know exists*, and it's exactly what unversioned prompts produce. Meanwhile: the model behind the scenes has changed (the vendor updates without telling you), the context has changed (new documents indexed, old ones removed), a temperature parameter has been touched. Five things move together. Metrics worsen in obscure ways.

**What we do now:**

Every prompt is in a git repo, has a changelog, a version, a commit message explaining *why* it was changed. Modifications go through pull request like any other code — even when "it's just a comma". When we release a new version, an automatic battery of tests (a set of questions/expected answers) runs and tells us immediately if something regressed.

It's a practice the community calls *prompt-ops* or, more seriously, *LLM-ops*. It's the thing that changed our lives on BP, where a single wrong prompt can produce a financial plan that looks like a real one but with made-up numbers. There you can't afford that.

## Mistake #3 — Paying for inference, not measuring retries

This is the subtlest one, because it disguises itself well.

When you put an agent into production, the first metric you watch is *cost per API call*. It's proposed in dashboards, it's easy to graph, it's what everyone talks about. So you monitor it. How much does OpenAI cost us per month, how much Anthropic, how much Mistral.

That number, alone, lies.

The economic truth of an agent in production is: **how many times does a real user have to retry an action before getting what they wanted?** Call it *cost-of-retry*, call it *abandonment rate*, call it what you want. That number contains almost everything.

Real example, from an iteration on Tuken. An assistant agent helped an event manager configure pricing packages. Inference cost: super low, a fraction of a cent per request. Real cost: the manager had to ask the same thing on average 4 times before getting the right configuration. The *real* cost was the manager's time (a lot) and their trust in the tool (which was collapsing). Inference was irrelevant.

The remedy wasn't changing the model or writing longer prompts. It was *measuring retry*. Once measured, product decisions aligned themselves: less free-text where it didn't work, more guided choices where they were needed, "let me talk to a human" fallback where the agent wasn't reliable.

**The KPI: every critical interaction must answer "did this user get where they wanted to go?".** If you don't know, you're paying a cost you don't see.

## Mistake #4 — Measuring the agent, not the product

We come to the worst one, because it's the most seductive.

When you work on an AI agent, you want to measure it. Right. You build dashboards that tell you: median latency, tokens consumed, tool call accuracy, model success rate on a test battery, percentage of fluent responses. They're good numbers. They give you the feeling of being in control.

And they're almost always the **wrong measure.**

The agent is not the product. It's *inside* the product. What matters is not "the model answered well". What matters is "the user completed the flow they came to do". They're two different things, and the second is much harder to measure, so we often don't. We measure the first, congratulate ourselves, and then one day notice users aren't renewing and we don't know why.

On Tuken we fell into exactly this trap for the first six weeks. We had beautiful dashboards on the agent: accuracy, latency, intent distribution. All green. Meanwhile the metric "did the manager publish the event after the conversation with the agent?" — the one that really told us if the tool served — we looked at once a week.

Now we look at it before all others. Agent metrics are *secondary* to product metrics. If the agent is a Nobel-prize model but users don't finish the flow, the agent needs redoing. If the agent is mediocre but users close the flow, the agent is fine.

Translated into an operational rule: **define the product metric** *before* **you start building the agent.** Not after. Not in parallel. Before. If you can't define a clear product metric, you probably don't need an AI agent — you need a tougher conversation with your team about what you're really building.

## The checklist, distilled

For those in a hurry. Four questions to ask yourself before putting an agent into production.

1. **Permissions.** Does the agent start at minimum? Does each permission have an expiration?
2. **Versioning.** Is every prompt in git? Does every modification go through a PR? Is there a test battery running on each release?
3. **Cost-of-retry.** Do we know how many times a real user has to retry? Do we measure it? Do we graph it?
4. **Product KPIs.** The metric that counts is the user completing the flow, not the agent answering well. Have we defined it *before* starting?

If the answer to a single one of these four questions is "not yet", you're making the most expensive mistake of 2026: keeping in production a system that seems to work but is costing you in ways you don't see.

## A small preview

Mistake number two — the versioned prompt — is only half the story. The other half is that a prompt, even a perfect one, is worth little if the *context* you pass to it is wrong. What you put in the context, where it comes from, with what freshness, with what permissions: that's what we call *context engineering*, and it's the real discipline of 2026.

We'll talk about it on June 11 in a dedicated piece. For now, let's stop here.

Monday, June 1, we'll go back to the startup pillar instead, with a piece that's been brewing for a while: *Founder mode is dead. Long live founder mode.* Two years after Paul Graham's famous essay, what's left of it in Italian startups that survived 2025.

See you Monday.

---

**Got an AI agent that works but you don't quite trust yet?** Our [artificial intelligence for business](/intelligenza-artificiale-per-il-business) service includes operational audits of agents already in production. Half a day, operational output, no PowerPoint slides. [Let's talk](/contatti).

Permissions too broad, prompts without versioning, retries nobody measures, KPIs that measure the wrong thing. An operational checklist written after making all four of them on Tuken and BP.

AI agents in production: 4 mistakes we see (and have made) every time

In October 2025 we published a piece titled *"AEO: the new frontier of visibility Google doesn't want you to know about"*. I opened with a data point: 58.5% of searches ended without a click.

That piece became the most-read of the year on our blog. For two reasons:

The first is that we'd called it right. AEO — Answer Engine Optimization, the art of getting cited by LLM-based answer engines — was really becoming the new game of organic visibility. Those who noticed it immediately began building advantage a year ago.

The second is that, while writing it, we didn't yet know how many things we would understand backwards in the seven months that followed.

This is the honest post-mortem: what aged well, what aged badly, what's already to throw away. In order.

## What worked a year ago (and still works)

Let's start with the good news. Three indications we gave in October 2025 hold up well in May 2026.

**Pages structured as "list of questions with answers below" are still gold.** They were for the old SEO (featured snippets) and they are for the new (chat citation). What's changed is the entry volume: LLMs love this structure because it's exactly how they reason internally — question → atomic answer → source. Keep writing real FAQs, long, with a self-sufficient answer of 2-3 paragraphs.

**Entities matter more than keywords.** We already said this. Today it's crystal clear: LLMs don't search for "phrases", they search for *connections between concepts*. If your page on Collio wine *also* mentions "Friuli", "ribolla gialla", "Slow Food", "steel-tank vinification", you're inside the semantic network of the category. If it doesn't appear, you're invisible even if the keyword is repeated ten times.

**Inbound links from sources that LLMs "read often" are worth what links from Wikipedia were once worth.** Reddit, Quora (the parts that survived), stable thematic communities, sector blogs with few but loyal readers, podcasts with public transcripts. SEO link-building for the next 18 months looks very much like digital PR from 2015. Focus on the 30 sources that count in your sector. Forget the 300 where nobody listens.

## What aged badly

Here things start to hurt.

**"Optimize for Perplexity!" was useless advice.** We gave it too. The truth is that Perplexity, ChatGPT, Gemini, Copilot, Claude, Mistral all have different strategies for crawling, citation, and ranking. Optimizing for one doesn't help on another. And every two months someone changes algorithm, indexing scope, or inline citation schema.

What we've understood is that you don't optimize *for an engine*. You optimize *to be the best answer*. Period. The technical nuances weigh much less than the toolmakers led us to believe. Those who sold "AEO Audit Tool" targeted at Perplexity sold smoke.

**Schema markup is less powerful than we thought.** Adding `FAQPage`, `HowTo`, `Article` with all fields remains useful, but the delta of citations it produces is low. LLMs don't read schema. They read text. Schema helped (and helps) the Google SERP. For AEO, real content matters, not hyperstructuring. Save dev hours to write better.

**"AEO metrics" are still a mess.** A year ago we hoped that by spring 2026 we'd have a Search Console version for AEO — impressions, click-through, average positions in AI responses. Spoiler: nope. There are third-party dashboards that try to monitor what percentage of prompts you appear in, but they're all estimates, with different methodologies, and numbers that vary by 200% between tools. If anyone sells you an AEO report with two decimal places, take two steps back.

## What's already to throw away

Just one thing, but it weighs.

**"Write content designed to be CITED"** — we said it again and again, but with a nuance that today needs complete revision.

For months we recommended a very specific format: short paragraphs, self-sufficient sentences, numerical claims clearly stated, citation of authoritative sources inline. The logic was: if the LLM finds a perfect bite, it takes it and cites you.

It worked. For about nine months.

Then something happened: every SEO on the planet started writing this way. The public web filled up with identical "bites" — 50-word paragraphs, percentage data, source in parentheses. Result: LLMs started to *distrust* this aesthetic, because it signaled optimized content, therefore often empty. The signal is inverted: today pages that look too much like the "perfect AEO format" perform *less* than pages written like a normal article, well-structured, with a human introductory paragraph and reasoning that unfolds.

It's the usual story. SEO 2010: keyword stuffing → penalized. SEO 2015: H1 obligatory + meta keyword → devalued. AEO 2024–2025: atomic paragraphs with citation → devalued. Every two or three years the system rewards content that *doesn't look optimized*, because everyone else is optimizing the same way.

**The new pragmatic rule:** write content that an industry expert would gladly read like an email from a colleague. No more "high-information-density sentences". No more forced "TL;DR". Long pieces with a real voice are back, built to be read, not parsed. The engines understand — because the models have improved to the point of reading the way we do.

## A new thing we didn't know in October

There's also a phenomenon that was marginal in October 2025 and is central today: the **second-hand citation**.

It works like this. You publish a piece. The piece is read by a human who works at Wired, at Il Post, at a serious Italian sector blog, at a niche newsletter. That human writes something based on your piece, and cites you. From that moment, *the LLM cites the secondary source* (Wired, Il Post, the newsletter), not you. Your content is the seed. But your brand doesn't appear.

This radically changes the AEO economy. For a year we thought being well-indexed by LLMs was enough. In reality it's better **to be well-cited by those who are well-cited**.

Translated into Italian: digital PR toward vertical media — those 30 names mentioned above — is the real AEO channel of 2026. The articles on your blog aren't enough. You need articles *on your blog that become the source of other articles*. It's a longer loop. It's what really works.

## The honest point: we're building tools to measure better

Editorial confession. While writing this piece we realized that, talking about AEO as consultants, we were forced to use the same tools we were criticizing. So we started building one ourselves, inside [OpificioAI](https://opificio.ai): a system that runs standardized batteries of prompts on the main AI platforms and monitors over time whether and how a brand is cited — with transparent methodology, no magic.

We're using it internally on clients who ask for it and on Tuken (the first product that will soon need to answer the question "does ChatGPT know me?"). We'll talk about it again when it's ready to be used also by those who aren't clients.

## The operational rule, seven months later

If we had to distill what we've learned in a single line, it would be this:

**AEO isn't a new discipline. It's journalism applied to SEO.**

Write well. Build authoritative sources around you. Measure what you can (and distrust those who measure everything). Update pieces that work, not those that don't. Stop chasing the trend of "LLM-optimized formats" and return to reasoning.

In seven months we'll revisit even this. There are things today that seem like the rule that by Christmas will be superseded. What we can promise is: we'll tell you.

Thursday, May 28, we'll write about the four most recurring mistakes we see when a company puts an AI agent into production. Four mistakes we made too, on Tuken and on BP. Four you can avoid right away if you start now.

---

**Want to find out if ChatGPT, Perplexity and Gemini are already talking about you?** We work on [sustainable organic growth strategies](/growth-hacking-strategie-per-crescita-sostenibile) between classic SEO and AEO. [Write to us](/contatti) and we'll do a free AI citation audit on your brand.

In October 2025 we wrote a piece on Answer Engine Optimization that became the most-read of the year. Seven months later, some things we've understood better. Others we've understood backwards. One is already obsolete.

AEO one year later: what works, what changed, what already broke

The most expensive rebrand of 2025 lasted 48 hours. Cracker Barrel — American restaurant chain, over 600 locations — presented a new logo in September, burned about $100 million in market cap in the following week, and pulled everything back.

We wrote about it [in the October piece](/blog/heritage-branding-quando-la-tua-storia-vale-piu-del-logo-perfetto). Today we want to tell the story that sits on the opposite end of the spectrum — the one that doesn't make the news, and precisely for that reason deserves more attention.

There are Italian SMEs that change skin constantly, but they do it so quietly that even loyal customers don't notice it until they look at an old package, an old bill, an old business card, and think "ah, but it used to be like that".

It's called *silent rebrand*. It's harder to do than an event-rebrand. It's almost always more effective. And it's exactly the discipline that separates branding used as an asset from branding used as a marketing stunt.

## What we mean (and what we don't)

A silent rebrand is not "never having changed the logo since 1972". That's stagnation, not discipline.

It is instead the **continuous release of micro-updates** that, taken individually, are not newsworthy, but added up over time produce a completely different identity. Think of the lifecycle of a console or a messaging app: the 2026 product doesn't look like the 2020 one, but nobody remembers *when exactly* it changed.

In short: the silent rebrand preserves the **emotional equity** of the brand (that thing the customer feels before thinking) while systematically updating the **material form** (logo, packaging, tone of voice, digital touchpoints).

Five recent Italian examples, with the lesson each carries.

## 1. Loro Piana — when the graphic disappears, but everything remains

What's changed in the last five years: slightly narrower typography, thinner mark, packaging where the graphic identity has lightened, product photography that abandoned studio sets for natural environments.

What has NOT changed: the material palette (camel, beige, neutrals), the sense of fabric weight, the idea that the product should speak before the brand does.

**Lesson:** when the product is the promise, the brand gets smaller and smaller. Loro Piana's silent rebrand is a controlled erosion of the graphic mark in favor of the material mark. It can only be done if you have something so recognizable on the tactile and visual plane that you don't need to remind people verbally. It works for those who sell craftsmanship, for those who sell engineering, for those who sell serious food.

## 2. Aboca — phytotherapy that constantly changes and looks unchanged

Aboca is one of the most interesting Italian companies to observe precisely because every 18 months something changes: packaging has been lightened, the website has been redone several times, product lines have been reorganized by use-occasion instead of by ingredient, and what was a side editorial line — Aboca Edizioni, born in 2012 — has grown to become an essential part of the brand's positioning, shifting perception from "phytotherapy product company" to "house of thought on the relationship between humans and biology".

Yet, if you ask a customer, they'll tell you "Aboca is Aboca".

**Lesson:** broadening the semantic perimeter is a deep rebrand move, but if you do it with continuous thematic coherence, the customer doesn't perceive it as rupture — they perceive it as *natural evolution*. The keyword here is *natural*. It means: no launches, no "now we're also…", no press release with a headline. Just books arriving in bookstores, products migrating on the shelf, copy adjusting itself.

## 3. Pasta Mancini — the farm-to-pasta that never shouts about it

Mancini pasta is the most cited Italian case of the last five years in food branding, but it's worth revisiting through this lens. The 2014 pack and the 2025 pack are very different: typography, hierarchy, palette, photography. But anyone who knows Mancini pasta recognizes it on the shelf at first glance even today.

Why? Because the thing they've never touched is the **compositional structure**: the photography of wheat in the foreground, the sober typography, the clarity of the origin denomination. The grammar has remained the same for ten years; the vocabulary has updated.

**Lesson:** in a silent rebrand you change the elements but never the **visual grammar**. If your packaging reads like a book (big title at the top, central image, paragraph below), you can change font, color, illustration — but it must continue reading like a book. If you change the grammar too, you've done an event-rebrand. And you have to say it, you have to justify it, and you have to cross your fingers.

## 4. Loccioni — the most underrated B2B in Italy

Loccioni is an industrial group from the Marche region that makes precision measurement systems for automotive, aerospace, energy. It's not exactly the first name that comes to mind when talking about branding. Yet it's one of the cleanest Italian cases of silent B2B rebranding.

In recent years they've modified site, manuals, client presentations, institutional videos: everything has become visibly more contemporary. But the brand has preserved the essential typographic rigor, the family name well visible (Loccioni, not an acronym), and above all the tone that Marche entrepreneurs recognize — sober, technical, never grandiloquent.

**Lesson:** in B2B, 70% of the silent rebrand goes through how you write, not how you draw. Updating the tone of voice, eliminating obsolete *jargon*, simplifying how you describe services: it's invisible on the graphic plane, devastating (in a positive sense) on the perception of modernity. It can be done in six months with a good copywriter and a clear brief. It's the most underrated move in Italian B2B branding.

## 5. Pastificio Felicetti — the rebrand done from packaging up

Pastificio Felicetti has refined over the years a strategy we like very much: they stopped thinking of "rebrand" as a top-down exercise (logo → everything else) and started thinking of it as a bottom-up exercise (packaging → everything else). The new lines — organic, single-grain, gourmet — became the laboratory where identity updated, and from there it climbed back toward the mother brand.

When they touched the main brand, it was already coherent with six or seven years of experimentation on sub-brands. Changing the mother brand turned out to be almost automatic, and almost invisible, precisely because of that.

**Lesson:** if you have a wide range, **sub-brands are the gym of the rebrand**. Experiment there, where mistakes cost less, and then consolidate upward only when the direction is clear. It's the opposite of what many agencies advise you to do (logo first, applications after) and works much better.

## The five rules of silent rebranding

If we try to extract a method from these five cases, something like this comes out:

1. **Change the elements, never the grammar.** Font, color, photography: free. Composition, hierarchy, perceptual weight: locked.
2. **Update the tone of voice before the logo.** Costs less, shows more, is reversible.
3. **Use sub-brands as the laboratory.** Identity is tested where mistakes don't hurt.
4. **No announcement, no press release.** If you have to explain the rebrand to the customer, you've done an event-rebrand. The silent rebrand is noticed after, not during.
5. **Measure emotional equity, not just graphic recognizability.** The right question is not "is the logo visible?", it's "does the customer feel the same thing they felt three years ago, even if they can't tell you why anymore?".

## When the event-rebrand is right anyway

For honesty's sake: there are moments when the silent rebrand isn't enough. Change of ownership, merger, entry into a new market, radical repositioning, reputational crisis. In those cases the event is needed, the declaration is needed, the clear signal to the market is needed.

But those are cases where *something has really changed in substance*, not just in form. Cracker Barrel, for example, had done an event-rebrand without changing the substance: and it went the way it went. The question to ask before a rebrand is not "will the market understand it?". It's "*is there really something new to understand, beyond the graphics?*".

If the answer is no, the silent rebrand is almost always the right answer.

Monday, May 25, we'll come back to the most-read topic of 2025: **Answer Engine Optimization**. What's changed in a year, what still works, and what we've understood doesn't work anymore — also thanks to what we've seen building our own products.

---

**Want a silent (and honest) audit of your brand?** We work on [branding projects tailored to Italian SMEs](/branding-su-misura-strategia-marketing-aziendale) every day. [Write to us](/contatti) and let's grab a coffee, no obligation.

While the giants spend millions on spectacular launches (and sometimes pull them after 48 hours), the companies that really work modify their brand in small continuous steps. Five Italian cases that do it better than average.

The silent rebrand: 5 Italian SMEs that changed skin without making noise

The last article published on Opificio Bites is dated October 5, 2025. It was called *"Heritage branding: when your story is worth more than the perfect logo"*. Then: silence for seven months.

It wasn't a break.

In those seven months we launched **OpificioAI**, we built **Tuken**, and we put another three products into the pipeline — a business plan assistant, a conversational business intelligence layer, and an AI-native publishing platform — that will see the light over the coming months.

We're not saying this for self-congratulation. We're saying it for two very specific reasons, and then we'll get back to work.

## Why we stopped

There's a comfortable version of this story: "we were focused on client delivery". It's true, but it's not the real one.

The real one is that in September 2025 we looked at the blog — well-written articles, well-informed, well-SEO-optimized — and we realized we were writing *commentary* about a craft we weren't doing enough of in first person anymore. We were talking about AI agents as if we'd already seen them all, talking about prompt engineering as if it were a closed discipline, talking about AEO as if Google would stop shifting the ground under everyone's feet.

For a creative studio that introduces itself as a specialist in AI for business, that's a dishonest position. So we made a simple, annoying decision: stop writing until we had something worth telling *from the inside*, not *from the outside*.

Seven months later, we have something to tell.

## OpificioAI: the studio became a product too

[**OpificioAI**](https://opificio.ai) is the new name we use for the part of our business that produces tools — not just services. Under this umbrella live the products we describe below. It's not a spinoff completely separate from the agency: it's the same team, with a second hat.

Why is it news? Because until now *we did AI for clients*; from today *we also do AI that clients use directly*. They are two different jobs, with two different responsibilities, and two different speeds. An agency advises. A software company ships code that runs at night while you sleep.

We had to learn the second thing. It's much, much harder than the first. More on this below.

## The products we shipped (and the ones coming next)

Three things are already live, two are about to launch.

**Already shipped.**

- [**OpificioAI**](https://opificio.ai) — the portal that brings together our offer of tools and agents for businesses. It's the entry point, the identity of a product company. You'll already see it linked at the bottom right on the AI pillar page of the site.
- **Tuken** — operational platform for managing events and activities with booking flows, package management, and customer handling. We're already using it with first real clients. It's the product that taught us what it means to write software that runs without asking us anything.

**In the pipeline, launching in the coming months.**

- **BP — Business Plan AI.** Conversational assistant for building complete financial business plans: products, personnel, costs, investments, financing, multiple scenarios, balance sheet anomalies. Designed for founders and SMEs that today fill in endless Excel files or pay thousands of euros to a consultant for their first round. *Status:* core engine in place, interface in polish.
- **BI — Business Intelligence AI.** Conversational layer on top of company data: questions in natural language, answers with charts and tables, no dashboards to configure for months. *Status:* selective early access.
- **Blogger AI.** AI-native publishing platform: built for those who want to run an editorially coherent blog without writing prompts by hand every day. It's a direct child of this very experience — of the fact that running a blog takes more time than anyone will ever tell you. *Status:* private beta in stress-test phase.

We'll write a dedicated piece on each of these products in the coming months. Not today.

## The four things we learned (that change how we write from today)

Seven months of building instead of seven months of commentary shifts the perspective. Four things in particular caught us off guard.

**1. Prompt engineering was the easy part.**

The real problem, in production, is *context engineering*: what you put into the model's context, from which sources, with what freshness, with what permissions, with what fallback. Writing the right prompt is an afternoon. Building a system that puts the right information in the right prompt is where you win or lose. We'll write a dedicated piece in a few weeks.

**2. The cost isn't inference. It's the retries.**

Everyone watches the per-token price. Almost nobody measures how many times a user *has to retry* an AI-assisted action before getting what they wanted. That number — call it *cost-of-retry* — is where the real economic truth of the product lives. If it's high, the user abandons and you don't know. If it's low, you've built something that delivers value without asking for too much.

**3. Agents don't replace people. They replace interfaces.**

When an AI agent works well in production, the effect is not "we laid someone off". The effect is "we removed three menus, two forms, and a dashboard". People remain. It's the interface that dissolves. This radically changes how you design a user experience — and for those coming from the design world, it's one of the most underrated stories of the moment.

**4. Compliance arrived before maturity.**

The AI Act enters new phases, GDPR intersects with training data, enterprise clients demand guarantees on where the model runs and what it does with data. The boring part arrived before the interesting part. For Italian SMEs this is bad news disguised as good news: it means that whoever starts today starts already with a structural cost of compliance. It's not insurmountable, but it must be budgeted.

## What changes on the blog from today

Three things, concretely.

**More from the inside, less from the outside.** Pieces on artificial intelligence and technology will be written more often from what we see inside our products — anomalies, errors, architectural choices, pricing decisions — and less from Andreessen Horowitz posts. Translated: fewer *trend pieces*, more *workshop notes*.

**More Italian cases, fewer translations of American cases.** This applies to all pillars — branding, growth, startup, AI. For every two articles, at least one with a recognizable Italian case. It's the section in which the blog can make the difference against the background noise.

**Honest cadence.** Two articles a week — Monday and Thursday — until mid-August. Published means published. If we skip, we say so. If we change our mind on a topic, we say so.

The editorial plan for the next three months is already written: 26 articles, distributed across the four pillars (artificial intelligence, branding, growth & SEO, startups), with two recurring series — *#ParoleDifficili* for the tech glossary and *#BraveHistories* for case studies of brave brands.

## A note to those who've been reading us for a year

If you've been here since the first Opificio Bites article was about manatees and slow scalability, we thank you with the sincerity it deserves: coming back to write without ever having asked you to stay was an act of trust.

If you arrived now — maybe because you clicked the OpificioAI banner at the bottom of the page, maybe because a client forwarded one of our pieces — welcome. The house rule is one: here we talk about marketing, branding, AI and business growth without trying to sell you anything we wouldn't sell to ourselves.

Thursday, May 21, we really get going again, with a piece on a phenomenon we're underestimating in Italy: the *silent rebrand* of SMEs. Five cases, no theory.

See you there.

---

**Want to see what we're working on?** The OpificioAI portal is online at [opificio.ai](https://opificio.ai). To talk to us about our [artificial intelligence for business](/intelligenza-artificiale-per-il-business) services, [write to us](/contatti).

Since October we haven't written here because we spent 7 months building OpificioAI and Tuken, and putting BP, BI and Blogger into the pipeline. What we learned, why we're telling you now, what changes on the blog from today.