You do not rank pages any more. You rank entities. Federico Fancinelli founded GeoSonar after two years running a GEO research lab, and that single shift is why AI search quietly favours established brands over the small players everyone assumed it would lift.

Key takeaways

• Bigger companies are entities by definition. Which is the uncomfortable finding: generative search is harder for small brands, not easier.
• Three layers, one signal. Infrastructure, narrative and authority, held together by coherence across every channel you own.
• Say different things in different places and you pay for it. A discrepancy between your site, your profile and your reviews is itself a signal.
• Prompt depth decides which sources get fetched. Generic prompts answer from the training cache. Deep ones reach into live retrieval.
• Sixteen million prompts analysed, and the advice is to watch about ten. Strategic categories beat volume.
• Prompt poisoning is real and narrow. Nudging a brand new site’s grounding works only before that brand is recognised.

On this page

• Ranking entities, not pages
• Coherence across the nodes
• Prompt depth and source tiers
• The accidental discovery
• Why Google publishes advice at all
• Speaking the machines’ language
• People, ideas and sources mentioned
• Questions this episode answers
• Go deeper

Ranking entities, not pages

It’s not ranking blue links, it’s not ranking pages, it’s ranking entities. And bigger companies are entities by definition.

— Federico Fancinelli

That second sentence is the whole problem. A large company already exists as a well-defined thing in the training data, with a Wikipedia entry, press coverage and a consistent name. A small company has to become one.

Federico has been in search since 2012 and built GeoSonar out of a research lab rather than a product idea, which is why the framing is structural rather than tactical.

Coherence across the nodes

The three layers are infrastructure, narrative and authority. Infrastructure is what a machine can read. Narrative is what you say about yourself. Authority is what everyone else says. The mechanic that binds them is coherence.

So if you are coherent among all of these nodes, then you have a higher chance of appearing as a suggested brand on the AI search results.

— Federico Fancinelli

Which makes the most common enterprise failure an organisational one. The website, the profile listing and the social accounts are owned by three different teams with three different descriptions of the company, and the discrepancy is the thing being measured.

Even the small can compete if everything is done right. So consistency is the key.

— Federico Fancinelli

Prompt depth and source tiers

Not every prompt triggers the same behaviour, which is why visibility tracking is so unreliable. A generic question gets answered from what the model already absorbed. A specific one sends it looking.

• Generic prompts answer out of the training cache, so recent work has no effect on them.
• Deeper prompts reach live retrieval, where the tier of source matters.
• Tier A is the very large and very trusted: encyclopaedic references, big forums, major publishers.
• Tier B is niche and local repositories, which is the layer a small brand can realistically enter.

His volume claim is sixteen million prompts analysed over roughly three months, and his advice from it is to stop tracking a thousand near-identical prompts and monitor around ten strategic ones instead.

The accidental discovery

Prompt poisoning turned up while building a bot-behaviour tracker. Repeatedly prompting models about a brand new website accelerated how quickly a grounding formed around it, which they watched happen server side with their own pixel.

The important caveat is that it only works before a brand is recognised. Once a model has a settled view of you, nudging it does nothing. It is a narrow window on new entities, not a lever on established ones.

Why Google publishes advice at all

Why would people trust what Google says about how to rank on Google? It doesn’t make any sense to me.

— Federico Fancinelli

So they are much more incentivized right now in controlling the narrative because who controls the narrative controls the behavior.

— Federico Fancinelli

He is not alleging deception so much as pointing at incentives. Guidance from a party with a commercial interest in your behaviour is a communications product, and should be read as one.

Speaking the machines’ language

So if you’re able to communicate with an agent in their own same language, they are more likely to choose you, not just because your products are better, but even before that.

— Federico Fancinelli

The practical version is structured, machine-legible information about what you sell, and he expects a formal integration protocol to become mandatory for software companies rather than optional.

SEO, in his framing, is not dead. It lives inside GEO, doing the same job it always did: making it possible for something else to find and understand you.

People, ideas and sources mentioned

Entity	What it is
GeoSonar	Federico’s AI-search visibility platform, built out of a two-year research lab
Coherence	Consistency across site, profile, social and reviews, treated as a ranking signal
Infrastructure, narrative, authority	The three layers of entity strength
Prompt poisoning	Accelerating a brand new entity’s grounding by how you prompt
Tier A and Tier B sources	Large trusted references versus niche and local repositories
Strategic prompt categories	Around ten meaningful prompts instead of a thousand duplicates

Questions this episode answers

• Why does AI search favour big brands?
• What does it mean to rank an entity rather than a page?
• What is coherence and how do I check mine?
• Do all LLMs work the same way?
• How many prompts should I actually track?
• What is prompt poisoning and does it work?
• Should I trust Google's own advice about ranking on Google?

Go deeper

Every link above goes somewhere different. These are the ones not already mentioned above.

• The full written recap of this conversation.
• The SEO Tactic Library, for the executable versions of the plays here.
• Federico Fancinelli on LinkedIn.

Find Federico Fancinelli

GeoSonar is the platform, and the research lab publishes separately.

Cite this episode

Fancinelli, Federico. Interviewed by Jeremy Rivera. “Entities, Coherence and Prompt Poisoning.” The Unscripted SEO Interview Podcast, 2026. https://the-unscripted-seo-interview-podcast.castos.com/episodes/geo-research-google-doesnt-want-you-to-see-entities-coherence-prompt-poisoning