Apple Intelligence has arrived on every updated iPhone, and Siri now uses external AI models to answer users' questions — including your potential customers, who use iPhones more than any other device. Almost no one is working on this visibility yet, which means whoever moves now starts with a huge advantage over those who wait. There are concrete actions you can take today to get found by Siri AI, and whoever takes them first claims a space that will only get harder to win.
A developer downloads Llama 3.1 and runs a fine-tuning on their company’s documents. In that fine-tuning, does your brand show up or not? If not, it’s an invisible B2B channel — because the same model running on that developer’s laptop is the technical cousin of what Apple is integrating inside iPhone, Mac and iPad with Apple Intelligence and the new Siri AI.
Let me explain. When Apple opened the season of on-device AI, it did two things: it put its own model inside the operating system, and it hooked Siri up to external models for complex queries. For your brand, that means one thing: the richest user base in Europe is about to ask questions of an assistant that can cite you or ignore you. And no one is measuring how.
Let me walk you through a test I ran on Llama 3.1 8B locally, with cultural-tourism queries about Tivoli, Villa d’Este and Villa Adriana. It’s the most honest way to understand what happens when an “open” AI model becomes the foundation of an assistant that answers in place of Google.
What Apple Intelligence really is for anyone selling a service
For your reader, Apple Intelligence is “Siri that works better”. For you, the one running marketing, it’s a new layer of intermediation between the site and the customer. When someone asks Siri “what’s the most beautiful historic villa near Rome to visit in a day”, Siri no longer opens Safari on Google. It answers. And in that answer you’re either there or you’re not.
Apple chose a mix: its own on-device model for the simple things, third-party models for complex queries. Among the models the market is evaluating as a foundation is the Llama family, which Meta released as open source precisely to occupy the layer on which OEMs and developers build products.
Meta compared Llama 3.1 with two closed systems via API (think ChatGPT and Claude) and with another open-source model hosted in-house. The difference is this: open models you can take, host, optimize, refine. Closed ones you can’t.
The consequence for your business: the content Meta used to train Llama 3.1 is a fixed asset. If your brand was in the pretraining, you’re inside the model that runs on-device on tomorrow’s iPhone. If you weren’t there, you have to hope for the retrieval layer that searches the web in real time.
Why Siri AI is a different channel from ChatGPT and Perplexity
In previous articles I talked to you about E-E-A-T for AI and about author entity recognition. Those mechanisms apply to every AI engine. On Siri there’s an extra layer.
Siri starts from an on-device intent classification: it figures out whether the question is solvable locally, whether it needs to search the installed apps, whether it needs the web or whether it should be handed off to an external model. Your brand can be intercepted in three places: in the base model (pretraining), in the site’s structured data (schema.org has always been the language Siri reads best), and in Apple Maps and Apple Business Connect for local queries.
In plain terms: recent open models make fewer mistakes but more often refuse to answer when they aren’t sure. From this follows one thing: if your brand isn’t anchored to recognizable entities (Wikidata, schema.org, citations in sources the model trusts), Siri AI may simply not name you rather than risk being wrong. Invisibility here is caution, not a penalty.
Optimizing only for Google and thinking that’s enough.
The test I ran on Llama 3.1 8B locally (Tivoli, cultural tourism)
Transparency: I can’t directly test Apple Intelligence from Italy with public data. I ran the closest test — I installed Llama 3.1 8B locally (an open model from the family many OEMs are evaluating as a foundation) and queried it cold, with no web connection, about historic villas and cultural tourism in Tivoli.
Why Tivoli: two UNESCO assets (Villa d’Este and Villa Adriana), an ecosystem of local operators (guides, B&Bs, restaurants, tours from Rome) and stiff competition with the capital. A typical case of a “town 30 km from Rome fighting for visibility”.
I ran 12 test queries, like:
- “What are the most important historic villas near Rome?”
- “What to visit in Tivoli in one day?”
- “Where to stay near Villa Adriana?”
- “Best tour guides for Villa d’Este”
Indicative results (test on a single model, small sample, signal not study):
- Villa d’Este and Villa Adriana always named, with accurate descriptions. The pretraining digested Wikipedia and institutional sites well.
- Across 12 queries, zero mentions of private local operators (B&Bs, independent guides, restaurants). Even on “boutique hotel in Tivoli with a view” the model answered with generalities like “search on Booking”.
- The Rome-based tour operators that sell excursions to Tivoli show up only if they have decades of history. Businesses founded in the last 5-7 years: invisible.
An honest limitation: Llama 8B is the small model. Apple Intelligence on the iPhone uses more optimized models, and Siri has access to real-time web search that my test doesn’t simulate. But the signal is clear: if your brand wasn’t in the pretraining, on-device you’re mute. And the pretraining is by now frozen for that generation.
OpenELM is Apple’s model. The ecosystem is fragmented but everyone draws from the same public corpora. Operational consequence: if you’re inside the “common source” — Wikipedia, Wikidata, authoritative sources — you’re inside all of them. If not, outside all of them.
Check your presence on Apple Maps and activate Apple Business Connect.
The mistakes I see most often when a brand thinks “I’ll optimize for Siri”
Optimizing only for Google and thinking that’s enough. Siri reads structured data but by its own rules. An Organization markup that works on the Rich Results Test can be ignored by Siri if specific fields are missing (phone number, Apple-friendly address, sameAs pointing to official social profiles).
Neglecting Apple Business Connect. It’s Apple’s equivalent of Google Business Profile, and many Italian operators don’t know it exists. For a B&B in Tivoli, a pastry shop in Modena or a dental practice in Padua it’s the first step to existing on Apple Maps and therefore on Siri.
Incomplete schema markup. Schema.org is Siri’s native language. If you run a cultural-tours site and you don’t mark up your packages as `TouristTrip` or `Event`, you’re speaking a language Siri doesn’t understand. Open the Rich Results Test, paste the URL of the offer page, see whether the schemas are recognized.
Thinking Apple Intelligence is “the future”. It’s available in Italian on the iPhone 15 Pro/Pro Max and across the entire iPhone 16 line, and on Macs and iPads with the M1 chip or later. Whoever is invisible today will stay invisible for years, because the pretraining of the next generation of on-device models is being done on today’s data.
What to do concretely this week
Three operational actions, in order of priority:
- Check your presence on Apple Maps and activate Apple Business Connect. Search for your brand on Maps from an iPhone: if you’re not there or the info is wrong, you’re out of half of Siri’s local queries.
- Open the Rich Results Test and check your homepage. Look for the `Organization` schema. If it’s missing, start here. For a tourism business, also add `LocalBusiness` or the specific sub-classes.
- Check your entity on Wikidata. It’s the structured source that models like Llama 3.1 digested massively in pretraining. If you don’t have a Wikidata page, you’re not a recognized entity — and on Siri AI you’ll be a “non-result” even when your customer is literally searching for you.
Three entry-level checks: they help you understand whether you’re out of the game. The real analysis — how the different AI assistants cite you, on which queries, against which competitors — requires professional tools.
The thread that ties it all together: visibility in AI answers
Apple Intelligence and Siri AI aren’t “just another search engine”. They’re the first serious attempt to move AI from the cloud to the device, with everything that entails in terms of who ends up in the model before the customer even asks the question. Visibility in AI answers isn’t bought on the spot, it’s built in the months before.
In previous articles I showed you how tokenization works and how AI recognizes entities in the knowledge graph. For Siri both mechanisms apply, amplified by the fact that on-device the margin for error is zero: either you’re a clear entity or you get skipped.
In the next articles in the series I’ll tell you how the other mobile AI assistants compare and how to structure a cross-platform strategy — Siri, Gemini and Copilot are converging on similar user experiences but with different citation rules.
