Third episode of the series. After unpacking [embeddings](/blog/hardwords-embedding), today we tackle the most-sold acronym of 2026: **RAG**, *Retrieval-Augmented Generation*.

If in the past year you received a quote for "an AI solution on your corporate documentation", 80% of the time they're selling you a RAG. Understanding what it really is — and when you really need it — is the difference between spending 30,000 euros well and spending them just to tell your customers you've done AI.

## The metaphor

A RAG is **an assistant that, before answering you, leafs through your archive**.

When you ask a question to generic ChatGPT, the model answers based on what it learned during training — a mix of Wikipedia, books, public articles, up to a certain date. It knows nothing *about you*. It doesn't know your contract with vendor X, your 2026 vacation policy, the installation manual for product Y.

A RAG fills exactly this gap. Here's how it's done:

1. Take your documents (PDF, Word, web pages, databases).
2. "Chunk" them into small pieces (chunking).
3. Transform each piece into an [embedding](/blog/hardwords-embedding) — a vector of coordinates in "meaning".
4. Save them in a vector database.
5. When the user asks a question, the system searches the database for the 5-10 pieces *most similar* to the question, and passes them to the AI model as *context*.
6. The model generates the answer based on those pieces.

Hence the name: *Retrieval-Augmented Generation*. Retrieval (from your documents) that augments generation (by the AI model).

## Why it exists

Three reasons, in order of importance.

**1. Factuality.** The AI model without RAG can invent ("hallucinate"). With RAG, answers are anchored to your documents — not to what the model learned from Reddit in 2024. If the source is your manual, the answer will be faithful to your manual.

**2. Updatability.** Retraining an AI model costs millions of dollars. Updating a RAG costs the time to re-index new documents — minutes, hours at most. When your price list changes, the RAG knows the new price. The "pure" model doesn't, until it's retrained.

**3. Permissions.** A RAG can respect access rights to documents. If Marco can see document X but Sara can't, the RAG can filter accordingly. A pure model can't: once the document is in its weights, it knows it for anyone who asks.

## When you really need it

Three typical cases where RAG makes sense. More precisely: where RAG is *the only sensible solution*.

**Case 1 — Customer support on complex product/service.** You have 200 pages of technical documentation and your customer care agents spend time searching them. RAG. It works, reduces average response time, is easily measurable.

**Case 2 — Internal help desk.** Employees ask recurring questions about company policies (vacation, reimbursements, HR procedures). Many questions, few HR people. RAG. Time recovered, frustration lowered.

**Case 3 — Pre-sales assistant.** You have a complex offer, multiple configurations, different use cases. An assistant that guides the customer through questions, retrieving relevant information from your brochures, is real value.

In all three cases, the *return on investment* is clear: less human time spent searching for information, more time spent doing things that require judgment.

## When you don't need it (and they're selling it to you anyway)

Three operational red flags.

**They propose a RAG to "help you make strategic decisions".** A RAG doesn't make decisions. It retrieves information. If your problem is decisional, RAG alone isn't the solution: a more complex system (agent, workflow) is needed, or maybe you don't need AI at all.

**Your documentation is disorganized, incomplete, or contradictory.** A RAG amplifies the quality of your documentation. If your documents are confused, the RAG will produce confused answers. It doesn't solve the problem; it automates it. Before a RAG, sometimes a content *due diligence* is needed.

**You have fewer than 50 documents.** Below a certain volume, a RAG is overkill. An AI assistant with documents loaded into context statically does the same job with a tenth of the complexity.

## What to ask the vendor

Six questions, in decreasing order of importance. If the vendor can't answer one of these without hesitation, take time before signing.

1. **How does chunking work?** (Do they split documents into pieces of how many characters? Do they maintain section context? Do they handle tables and images?)
2. **Which embedding model do you use?** (And does it know Italian? See the [embeddings piece](/blog/hardwords-embedding).)
3. **How do you handle permissions?** (If a document is confidential, does the RAG know certain users can't see it?)
4. **How do you handle updates?** (When I add a new document, how quickly is it available? Is there version control?)
5. **What does the assistant do when it can't find the answer?** (Says "I don't know" or tries to invent? The right answer is the first.)
6. **How do you test answer quality?** (Do they have a battery of test questions? Do they run it on every update?)

Three well-formulated answers, and you can proceed. Three hesitations or "we evaluate case by case", and they're probably selling you an idea, not a system.

## Summary in three lines

A RAG is an AI that, before answering, leafs through your documents.

Solves the problem of hallucinations and updatability — *if* your documents are of quality.

You really need one only if you have a lot of documentation and a lot of recurring questions. Otherwise they're selling you an expensive trend.

Monday, July 13 we change register: *The founder's summer. 4 things to do in August that really weigh in September.* No "recharging the batteries". Four low-volume, high-impact jobs.

---

**Have a RAG in mind for your company and don't know if it makes sense?** Under [artificial intelligence for business](/intelligenza-artificiale-per-il-business) we audit RAG-quotes in half a day. We explain what you're buying, before you buy it. [Let's talk](/contatti).

All AI vendors are proposing 'a RAG on your corporate documentation'. What it really means, when you really need it, and when they're simply selling you yet another trend.

#HardWords: RAG (Retrieval-Augmented Generation)

There's a narrative genre that in July repeats in every Italian company: the strategic meeting where it's decided "in September we'll do something serious on AI". A manager is nominated, consultants are mentioned, a budget is estimated, it gets postponed.

September arrives, priorities change, "something serious on AI" becomes "something serious on AI in Q1 2027". And so on.

This piece is to sabotage this cycle. Five **real** experiments, launchable in 5-7 days of effective work. No consultants needed. No enterprise budget needed. What's needed: one person from the company with time, an account on a commercial LLM (ChatGPT Team, Claude Pro Team, or Gemini Workspace), and the discipline to close the pilot within the week — not open it to close it at Christmas.

Three rules of engagement.

**Rule 1**: one experiment = one concrete problem. No "let's see what AI is good for". Yes "let's see if AI solves this specific thing".
**Rule 2**: declared effort budget. 5-7 days of one person. If more is needed, the pilot failed (and that's fine — you know that road doesn't work).
**Rule 3**: binary success criterion. Yes or no. No "interesting but worth investigating".

The five experiments, in order of increasing ease.

## Experiment 1 — Analyzing 100 customer emails

**The problem**: in your commercial inbox request emails arrive from potential customers. Every email has signals inside — sector, company size, urgency, specific problem — but nobody classifies them. They get read one by one.

**What you do in a week**: take the 100 most recent emails, anonymize personal data, and feed them to the model in 5 batches of 20 with your classification template (sector, size, urgency, request type). The model returns an Excel to you.

**What you discover**: probably, two or three dominant segments you hadn't noticed. A concentration of requests in a specific sector, or a recurring problem you didn't know was so widespread. It's the raw material to reposition September's communication.

**Success if**: at least two actionable insights emerge that you didn't have before.

## Experiment 2 — Site content audit

**The problem**: the site has 50-200 pages. Some are excellent, others outdated, some overlap. Nobody has ever done a *real* audit because it would take a senior person's week.

**What you do in a week**: download page content (even just from a sitemap), give the AI precise instructions — *"evaluate each page for: clarity of message, freshness of data, overlap with other pages, alignment with the attached brand voice"* — and receive a report.

**What you discover**: usually 10-15 pages to revise immediately, 3-5 to eliminate, 2-3 obvious overlaps generating SEO confusion.

**Success if**: the report identifies at least 5 concrete actions to take before September.

## Experiment 3 — Dynamic FAQ for help desk

**The problem**: you receive repetitive questions in pre-sales or support. When an email arrives, the answer is practically always the same. Your customer care team spends half their time copy-pasting.

**What you do in a week**: take the last 200 customer support conversations, have the model categorize recurring questions, generate a dynamic FAQ of 20-30 entries. Then configure an AI assistant on the site that answers these 30 questions, with escalation to human if the question doesn't fit.

**What you discover**: between 60% and 80% of requests fit into these 30 questions. The chatbot doesn't replace customer care, but frees time for real requests.

**Success if**: after a week of testing on the site, at least 50% of users who used the chatbot didn't need to talk to a person.

**Warning**: before publishing the chatbot, read our [piece on AI Act phase 2](/blog/ai-act-phase-2). There's a transparency obligation.

## Experiment 4 — Automatic briefing for calls

**The problem**: before a call with a prospect, you should do 30 minutes of research on them: site, latest LinkedIn posts, sector news. Nobody ever does it.

**What you do in a week**: create a prompt that, given the customer's name and site URL, returns: what they do, positioning, latest public announcements, typical problems of their sector, a draft of 3 smart questions to ask. Keep it in a favorite. Before every call, launch it.

**What you discover**: the quality of commercial calls rises 30% without investing *an hour* more per call.

**Success if**: after a week, people using it say it's "uncomfortable not having it".

## Experiment 5 — Generating variants for social ads

**The problem**: the creative team produces 3-4 variants of a social campaign. They're few. Meta's bandit algorithms need 8-15 variants to learn well.

**What you do in a week**: take the variant that worked best in recent campaigns, use it as a *seed* to generate 10 variants with the AI model — text + image description (images are made by generative tools, but even as a brief for the internal designer it works).

**What you discover**: 2-3 of the new variants perform better than the original. It's not automatic, it's statistical. More test volumes = more probability of finding a winner.

**Success if**: the average CPA of the campaign drops at least 10% after two weeks.

## The principle behind all five

Did you notice the common structure?

- Specific problem, not abstract
- Material you already have (emails, pages, conversations, briefings, ads)
- Weekly effort, not quarterly
- Binary success

This is the **AI pilot discipline** for 2026. Companies doing AI seriously don't spend two quarters "evaluating the vendor market". They do five experiments in five weeks. Three fail, two work. On the two that work, they scale.

Everything else is theater.

What do you really need to start? An hour of a decision-maker — you — to choose which of the five to begin with. And a person willing to do the 5 days of effective work. Total cost of the operation, on the books: a week. In licenses: 30-50€ (a team account on an LLM). Risk: zero, or close to it.

Now you have no more excuses. Now, maybe, you also have time — it's July, the company runs at half speed, it's the perfect moment.

Thursday, July 9 we return to the glossary: **#HardWords: RAG (Retrieval-Augmented Generation)**. Why all AI vendors are proposing "a RAG on your documentation" — and when you really need one.

---

**Want one of the five experiments done well, four hands?** Under [artificial intelligence for business](/intelligenza-artificiale-per-il-business) we do AI pilots with fixed time, fixed price, fixed output. [Let's talk](/contatti).

Five tactical pilots, quick ROI, no enterprise budget. For Italian SMEs that want to do AI seriously before September without convening a seven-person committee at a Lake Garda resort.

5 AI experiments you can launch in a week (even from under the umbrella)

Tradition. In August 2025 we published a similar piece, and many of you arrived — more than a summer article usually gets, a sign the material interests beyond seasonal limits.

We do it again, but differently. Shorter. Sharper. A non-hagiographic sequel — the summer of 2026, in Italy, is a strange season: excessive heat, post-Easter market still soft, nervous wait for the September return. The brand that goes through it well comes out stronger. The one that goes through it badly comes out confused.

The three promised lists, in order.

## The ten mistakes we see every year

In quick sequence, with minimal annotation.

**1. The autoresponder that closes everything.** *"I'll be back in office on September 1. For urgencies, contact [colleague]."* Period. No brand, no personality, no useful alternative. Eight words of vacuum while your customer is writing because they need help *now*.

**2. The "discover September's news" in the autoresponder.** The idea that your vacation is an occasion to do promotion. It's jarring. Vacation is a *stop*, not a *marketing trigger*.

**3. The silent socials for four weeks.** A brand that disappears in August and reappears September 1 makes noise. In the sense: the audience notices the absence. Not always positively.

**4. The socials full of "happy holidays!" fake-enthusiastic.** Opposite variant. Posts of sun, sea, umbrellas, essentially identical between competitors. Noise without signal.

**5. The newsletter that skips August and returns with an "energetic September".** If your newsletter loses a week in August, nothing happens. If it loses four, users forget about it. Frequency is a pact.

**6. The autoresponder that reveals too much.** *"I'm on vacation in [Exotic Location], I don't read emails."* For a client dealing with a real crisis, it's gas on the fire.

**7. The empty blog from mid-July.** Typical mistake of companies that publish "when inspiration arrives". Summer is perfect to *schedule* in advance, not to *stop*.

**8. The homepage that doesn't update the summer message.** The site that in mid-August still shows the May promotion. Small, but it's the sign that nobody really looks at the site.

**9. The return newsletter titled "We're back!".** Serious variant: the newsletter that on September 1 tells *that you're back*, rather than *what you did* or *what's changing*. Whoever reads you didn't look for you: they opened you out of habit.

**10. The automatic replies with grammar errors.** They appear everywhere. They are the saddest case of all: written hastily in July, applied to everything, forgotten.

## The five brilliant ideas (that really work)

**1. The "real August" newsletter.** A special edition that doesn't talk about your brand, but about a collateral topic your audience loves: five books, three podcasts, ten places to walk to in your region. It's a gift, not a marketing action. And precisely for that, it works really well.

**2. The founder's social.** Photos of the founder doing normal things — an event, a market, a walk, a conversation. No promotion. It's a summer seen through the founder's eye. It creates a level of intimacy no paid campaign can buy.

**3. The declared pause.** A post at the top of the site or a pin on socials: *"This summer we stop from August 5 to 25. We'll be back on the 26th with [thing]"*. It's honest. It's useful. It's almost always more appreciated than fake "keep going".

**4. The summer backstage.** Photos of who works in July when many are already gone. Not vindication, story. Companies made of people have people who work in summer, and telling them values them.

**5. The pillar content of mid-August.** What few do. Publish a serious and long article on Ferragosto (mid-August holiday). Little noise around = very little competition, and the readers who are there are disproportionately active (and bored). It's one of the best times of the year for impact.

## The Italian case: Crodino

Just one case. Deliberately.

Crodino is one of the few Italian brands that has transformed summer into product identity. Not a "summer launch" that repeats. A **structural seasonality** that becomes character. Crodino is the aperitif you realize you want to drink when the heat arrives. For ten, eleven months of the year it's a background presence. In June-July-August it returns to the foreground, with a coherent language (the yellow, the glass, the solitary or convivial aperitif), and then returns to rest.

They don't do a "summer campaign". They do their job, always, and summer amplifies it.

It's the most important lesson of the piece. **A brand on holiday isn't a brand that changes tone for summer. It's a brand that, in summer, shows the most amplified version of itself.** If you don't have a version of yourself that amplifies naturally in August, the problem isn't the July editorial plan. It's the brand upstream.

Monday, July 6 we enter the full summer season with a tactical piece: *5 AI experiments you can launch in a week (even from under the umbrella).* Quick pilots, fast ROI, no enterprise budget.

Happy holidays to those who leave. Good work to those who stay — here we're still writing.

---

**Want a brand that works even when you're not thinking about it?** We build it together on [substantial branding projects](/branding-su-misura-strategia-marketing-aziendale). [Let's talk](/contatti) — even in July.

Summer is the litmus test of branding: those in good shape prove it even in August, those who aren't discover that the account manager's out-of-office is not enough to make a brand work. The summer appointment of the blog returns.

Brand on holiday 2026: 10 mistakes, 5 brilliant ideas, 1 Italian case

For twenty years, "being visible online" meant one thing: ranking on Google. It was a solid geographic metaphor — *the first page* — and a solid hierarchical metaphor — *position 1, 2, 3*. There was a map. There were rules. It worked.

Today we're at a point where the map has tripled. There are three different types of engines, each with different rules, each with a different path to be on.

And almost all the content marketing plans we see in 2026 are still written for a single engine — the previous one — applying rules that held in 2018.

This is the piece we'd been wanting to write for months. The new map of organic visibility, in three steps. A long-form pillar that will serve to revisit content strategy in autumn. Put it away.

## The three engines that coexist today

Let's start with a simple observation. When a user searches for information online today, they can do it in three profoundly different ways, on three different types of engine. Each produces a different consumption behavior, a different advertising, a different citation/visibility model.

### Engine 1 — Classic search engine (Google, Bing)

The query is a string. The result is a list of links. The user clicks, lands on the site, reads.

**What it rewards**: relevant keywords, inbound links, technical structure (Core Web Vitals), schema markup. It's the SEO we've known for twenty years. It still works, but the *traffic volume* it produces for many sectors has dropped 30-50% from 2023.

**Where it still matters a lot**: e-commerce (purchase decisions almost always pass through a click), B2B with long cycles (the user opens 8 tabs and reads), local content ("restaurant near me" is still Google).

**Where it matters less and less**: informational questions ("how do you calculate DSCR?"), comparison questions ("difference between A and B?"), synthesis questions ("summarize the history of company X").

### Engine 2 — Answer engine (Google AI Overviews, Bing Chat, and Perplexity in its more traditional mode)

The user makes a query. The system generates **a synthetic answer** directly in the page, citing some sources. The user, in 60-70% of cases, *doesn't click*. The answer is satisfying, the user moves on.

It's the AEO — *Answer Engine Optimization* — we've talked about for a year (remember the [2025 piece](/blog/aeo-la-nuova-frontiera-della-visibilita-che-google-non-vuole-che-tu-conosca) and [the update from a month ago](/blog/aeo-one-year-later)).

**What it rewards**: content structured in question-answer, topical authority (sites that return frequently to a topic), citations from recognized sources, clear declarative language.

**The problem**: they don't send you the visitor. They cite you. The value of classic traffic decreases, the value of *brand mention* increases.

### Engine 3 — Generative engine (ChatGPT, Claude, Gemini conversational, Perplexity in Pro mode)

The user doesn't make a query: they have a **conversation**. They ask for advice, evaluate options, want a plan. The model builds a complex answer integrating sources, pre-trained knowledge, and the conversation context. Citations are optional and often secondary.

This is the real frontier of **GEO — Generative Engine Optimization**. The rules are different because the "engine" isn't searching for documents, it's *building an answer* inside a dialogue.

**What it rewards**: presence in the **training data** of the model (so: having been published, cited, discussed *before* the training cutoff), well-defined entities (your brand is a recognizable entity, with clear and differentiating characteristics), digital reputation (mentions in sources the model treats as reliable).

**The problem (big)**: you can't optimize in real time. What the model knows about your brand, it knows at the moment of training. If you weren't there, you aren't there. And model update cycles are months.

## The three-circles rule (and where your brand can really be)

The companies that ask us *"how will we be visible in the next three years?"* often want a single answer. The real answer is: **choose which of the three circles you're in, and optimize mainly there**.

The three circles overlap but aren't identical. The critical point is they require different strategies, different content, and — especially — *different metrics*.

**Circle 1 — Search.** Keep doing SEO seriously if: you sell products (e-commerce), you have a strong geographic localization (restaurants, local services), you have a long B2B funnel where the user compares 5 vendors before choosing. Here your KPI remains qualified organic traffic.

**Circle 2 — Answer.** Invest in AEO if: you sell consulting services, technical expertise, thought positioning. Here your KPI is the **citation**, not the traffic. ChatGPT cites you 1000 times a month? It means you're inside the discourse. Traffic is secondary; brand recognition is primary.

**Circle 3 — Generative.** Invest in GEO if: your client's purchase decision happens *after a long conversation with an AI*. Think of real cases: those looking for a strategic consultant, a training school, a legal consultant, a branding consultant (yes, us too). Here your goal is to be **cited as an option** in conversations where a decision-maker evaluates partners. It's become a serious channel in 2025-2026.

For almost all companies, **one circle is dominant**. For some (the most mature), two circles coexist. Very few companies really work on all three — and rightly so, because spreading thin costs.

## The seven practical rules of GEO (that we actually use)

Here's what we learned in the months we spent making [OpificioAI](https://opificio.ai), Tuken, and the products coming, recognizable inside the AI conversations of our potential clients.

**Rule 1 — Define your entity.** AI must know who you are. Three clean sentences: what you do, for whom, what makes you different. Repeated with small variations on all channels. Site, social, professional profiles, podcast appearances. Extreme consistency.

**Rule 2 — Become citable by third-party sources.** The more you're cited by sources the model treats as authoritative (vertical newspapers, stable sector blogs, podcasts with public transcripts, GitHub, Wikipedia if possible), the more you appear in AI conversations. It's the digital PR of 2026.

**Rule 3 — Publish "canonical definitions".** For every key concept of your business, have a page that defines it clearly, authoritatively, citably. It's what LLMs search for when they have to explain a term.

**Rule 4 — Curate your linguistic footprint.** If they call you in three different ways (e.g. *"OLA"*, *"Opificio Lamantini"*, *"Opificio Lamantini Anonimi"*), AI struggles. Decide your canonical name, use it always, and on main pages make aliases explicit.

**Rule 5 — Measure citation, not traffic.** Tools exist (including the one we're building) that run batteries of prompts on main AI platforms and see what percentage you appear in. That's the new *Search Console*. Start looking at it.

**Rule 6 — Update old content, don't just create new.** A 2022 article of yours that ranked well on Google in 2023 may today be your biggest GEO asset — because it entered training data. *Updating it*, not replacing it, is the move SEOs don't expect and LLMs appreciate (freshness matters, but continuity matters more).

**Rule 7 — Build a POV.** Companies without a clear point of view get ignored by LLMs, because there's nothing to cite. Companies with a clean *POV* (even uncomfortable) get cited by opposition, confirmation, articulation. Having an opinion is the single most underrated thing of GEO.

## The three most expensive mistakes of 2026

To close, what we see being mistaken most frequently.

**Mistake 1 — Continuing to measure only organic traffic.** If your Google traffic drops 20% but your mentions in AI chatbots rise 50%, *you're winning*. The old metrics don't tell you this story. Update the dashboard.

**Mistake 2 — Trying to optimize for *all three* engines simultaneously.** It's the perfect trap. You produce content that tries to please everyone and ends up working for no one. Choose a dominant circle, excel there, gather the others as side effects.

**Mistake 3 — Thinking GEO is "a slightly different version of SEO".** It isn't. SEO optimizes for algorithms searching for *match*. GEO optimizes for models searching for *meaning*. The two require different *minds*: a good SEO is not automatically a good GEO specialist. The new specialist looks more like a digital PR + an editor than a positioning technician.

## One last thing

If you're wondering *"which circle should I be in?"*, here's a quick diagnostic question: *how does your target client discover their vendors?* If they search on Google and click, you're in circle 1. If they ask the browser chatbot and read the synthetic answer, you're in circle 2. If they have a long conversation with an AI and name you, you're in circle 3.

For most companies in 2026, the truth is: **one dominant circle and one support circle**. Classic SEO continues to bring *purchases* (circle 1), but circle 2-3 brings *consideration* — the phase in which your brand enters or doesn't enter the shortlist. And without consideration, even the best Google SEO isn't enough.

We'll talk about it again. For now we say goodbye here.

Thursday, July 2 we enter the summer era with an affectionate, quick piece: **Brand on holiday 2026** — 10 mistakes, 5 brilliant ideas, an Italian case. Sequel to our summer classic.

---

**Have a content plan that worked in 2023 and today yields less and less?** Under [growth hacking and sustainable organic growth](/growth-hacking-strategie-per-crescita-sostenibile) we do full SEO→GEO reconfigurations. [Let's talk](/contatti) before September.

Search, Answer, Generative. Three different engines, three different strategies. How we're reconfiguring client content to stay visible inside ChatGPT, Perplexity, and Gemini — without throwing away what works on Google.

Benvenuto in Opificio: il luogo dove i Lamantini sono pronti a far emergere idee, brand e innovazione al tuo fianco.

Scopri i nostri percorsi

Chi ci ha dato fiducia

Gli ultimi Bites

Studio creativo a Monselice (PD): branding, AI e marketing per aziende

Cosa facciamo

Dove operiamo

Come lavoriamo