For years, SEO ran on a simple promise: rank the page, win the click. In the GEO era, the unit that gets ranked has changed — and so has who has the advantage. We sat down with Federico Fancinelli, co-founder of GeoSonar, on the Unscripted SEO Podcast with host Jeremy Rivera, to talk about what two years of research into AI search visibility actually found.
The surprising finding: you rank entities, not pages
Jeremy Rivera: There's nothing like research when it comes to SEO, because we have plenty of anecdotes but we're very sparse on actual data. What's your most surprising conclusion?
Federico Fancinelli: The most surprising thing is that I expected the opposite of what we found. I thought it would be easier for smaller brands to rank in AI answers — a generational opportunity, like what happened twenty years ago when Google went mainstream and the brands that did SEO made a fortune out of it. I assumed the same shift was happening now, and that small players finally had a shot against the big ones.
Then the lab proved me wrong. In most cases it's the opposite. The core difference between SEO and GEO is this: it's not ranking blue links, it's not ranking pages — it's ranking entities. And bigger companies are entities by definition. They don't need to prove they're entities to the web — they just are. They organically have more citations, more backlinks, more of everything that helps you rank in AI search results.
Jeremy Rivera: So in traditional SEO you had a bigger chance to compete, because it wasn't your whole entity being measured against other entities — it was just a single page.
Federico Fancinelli: Exactly. The lab actually started as a pro bono project to help small businesses rank faster on AI. What we realized instead is how much work is required to become a recognized entity that can compete against a multinational corporation or a much bigger competitor.
What actually defines an entity
Jeremy Rivera: Is entity recognition just an accumulation of third-party citations across data sources, or is there a heuristic layer where these AI tools are building their own knowledge schema — deciding internally what counts as an entity?
Federico Fancinelli: We need to separate the technical term "entity," as in something you can map on a knowledge graph, from the broader idea: the space you occupy in your market, the thing you represent on the web as a whole. That's the entity you need to define in order to be recognized first, and ranked second.
What the research found — and this isn't opinion, it's the data — is that as you try to define what you are, you're working across three layers: infrastructure, narrative, and authority. It's not just external mentions. It's the combination of external signals and what you say about yourself on your own channels. Entity recognition for LLMs isn't built from one source — it's a combination of factors that form the idea of your entity in their grounding, and from that grounding they generate an answer.
If you're coherent across all of these nodes, you have a much higher chance of appearing as a suggested brand in AI search results. If what you say on your website contradicts your LinkedIn page, or differs strongly from what your reviews say, that discrepancy penalizes your brand heavily. Coherence is one of the keys in entity forming. If you're coherent among all the channels relevant to your business, you have a real chance of positioning yourself in the results.
Coherence across the layers LLMs pull from
Jeremy Rivera: Is that coherence being checked in real time — through a query fan-out process where the LLM runs live searches and mines results — or is it coherence baked into the model's training data? Or a combination of both?
Federico Fancinelli: It depends on the prompt. That's one of our key findings. If the prompt is generic, the LLM usually pulls from the highest layer — the training data it already has, like a cache. It's easier to access, and LLMs have to be economically efficient, so they default to the readily available pool. But if you make a longer, more detailed, structured prompt, the model — depending on which one — tends to dive deeper into other pools of sources: tier A, tier B.
Tier A is usually something like Wikipedia, Reddit, Forbes. Tier B is more niche — your specific field, or if you're a local business, local repositories relevant to your area. So assuming a deep, structured prompt, you need coherence across all of those layers. That's the key.
Every LLM is a different beast
Jeremy Rivera: Is there a meaningful difference between Google's AI Overviews and the other LLM tools — talking with Matt Brooks of SEO Teric on this topic — ChatGPT, Perplexity, Grok, Claude — or is Google's advantage just its dominant search engine sitting on top of the same process?
Federico Fancinelli: It's absolutely not the case that they're the same. Every LLM has its own methodology for sourcing the information it uses to generate answers — and the differences are significant. Even within the Google ecosystem you have three different systems: AI Overviews, Gemini, and Google AI Mode. AI Overviews and Gemini behave similarly in our findings, since AI Overviews partly relies on the Gemini engine. Google AI Mode is a different beast — it blends traditional search results with the Gemini engine, so it overlaps more with classic Google rankings than with Gemini or AI Overviews alone.
Move outside the Google ecosystem — Anthropic, Perplexity — and you're in a completely different world, with different reasoning and different source interpretation. We have a large dataset across our clients, and we consistently see brands that are present in OpenAI's ecosystem but absent from Claude or Gemini. Optimizing for one LLM does not grant you the same ranking in all the others. They're genuinely different in how they search for information.
From keyword analysis to prompt analysis
Jeremy Rivera: I've seen stats suggesting you get a different answer as much as 80% of the time for the same prompt, depending on the user or session. How do you build reliable research on top of that kind of variance?
Federico Fancinelli: Big struggle. This is really about moving from keyword analysis to prompt analysis — even a small change in how a prompt is worded can produce a very different answer. And this we've actually mitigated with AI visibility mapping, a dashboard that shows if you are actually present in AI answers. What mitigated this for us was moving from quantity to strategic quality. We took inspiration from classic SEO: branded keywords, long-tail keywords, generic keywords — and translated that into prompt categories. We've analyzed tens of millions of prompts, and what we've seen is that once you monitor the right categories, the exact number of prompts becomes almost irrelevant — the picture balances itself.
You can monitor just a handful: a branded prompt ("What can you say about [brand] and why is it the best?"), a long-tail prompt specific to your niche, and an intent-based prompt that isn't even product-related — just "I'm about to launch, what should I do?" Map those strategic categories and it doesn't matter whether you run ten prompts or a thousand nearly identical ones. Curiously, if you reduce the number of prompts and make them more strategic, the results you're mapping actually become more consistent.
The observer effect and "prompt poisoning"
Jeremy Rivera: There's a concept in physics where particles behave differently when observed versus unobserved. We saw something similar in SEO when Google removed the ability to see 100 results at once and search visibility tools reported artificial drops. Is there evidence that LLMs adapt based on how often — or how — they're being asked about a brand? I believe Google or Microsoft has referenced something like "prompt poisoning".
Federico Fancinelli: Now we're entering black-hat GEO territory a little. It's actually very relevant to one of our latest findings. We were building a tool — not public yet — to measure how AI bots behave when they scrape a website. In the process, we were inadvertently prompt-poisoning the models: we made specific prompts about brand-new websites to test bot behavior, and we tracked everything server-side with a pixel installed for that exact reason.
What we saw is this: if you ask questions repeatedly in a certain way across all the LLMs, you're actually influencing the answers. We have proof of it — a mathematical proof, and video of it happening in real time. It only seems to work in the early stage, though. We're not confident it works once a brand is already recognized and mapped in the AI's grounding or training data. But on a brand-new website, prompt-injecting or "poisoning" the various LLMs can produce very fast results. It wasn't intentional research, but it happened.
Why you can't trust Google's advice
Jeremy Rivera: Have you seen public statements from the big platforms that feel like noise designed to distract from what's actually happening?
Federico Fancinelli: Think about Google's business model: what incentive does Google have to tell marketers exactly how to rank in their own system? Zero. Their business model depends on giving users the most useful answer, because that's what keeps people using the system, which is what sells ads. Imagine Google as a bank manager. They would never tell you where the cameras are positioned so you can rob their vault. Why would anyone trust what Google says about how to rank on Google?
Jeremy Rivera: I can't advise trusting Google either — we heard for a decade that they didn't use SERP behavior as a ranking signal, and then the 2023 DOJ lawsuit revealed it was one of their core ranking inputs all along. They're extremely good at saying things that are technically true and functionally worthless for anyone trying to optimize against them.
Federico Fancinelli: A company the size of Google never moves without intention. If they release information about how their ranking system works — the thing they should protect the most — there's a reason. I believe it's narrative control. As LLMs grow more prominent, Google is more incentivized to control the narrative around search, because whoever controls the narrative controls behavior — and the behavior they want is people staying on Google, so they keep selling ads. Releasing more information about how to "rank" in generative results lets them shape what's considered relevant, and it pre-empts an independent player — a GeoSonar, an Italian startup built from nothing — from figuring out the real ranking parameters and causing them real trouble.
SEOs as courtiers in the AI court
Jeremy Rivera: There's an analogy I like: SEOs used to be courtiers at the court of a single king — Google, with Bing off in the corner. Now there's a whole court: ChatGPT, Anthropic, Copilot. You're whispering to more players in the room, trying to build connections and get your patron recognized.
Federico Fancinelli: An oligarchy. It's a good analogy. And don't you think the king is getting a little jealous of the emerging sovereigns right now?
Jeremy Rivera: Definitely — and I think it's great. I'm enjoying the more open, wild-west scenario, because competition breeds innovation.
Featured snippets to context-dependent AI answers
Jeremy Rivera: Classic featured-snippet optimization was about query diversity — subheaders that directly answer a question, a snippet right after, and multiple answer formats (image, video, table, bullet list) under one H2 to improve your odds of being selected. Is there anything similar in how LLMs choose between different answer formats?
Federico Fancinelli: Absolutely — and AI answers are far more context-dependent than classic search. LLMs tend to prefer more specific content — even much more than the featured snippets did. If someone's looking for kitchen renovation inspiration and wants images of countertop colors, the AI fetching process looks specifically for image-heavy, consistent content across the whole ecosystem — not just one website, but your site, your LinkedIn, all your channels — and rewards the most visually consistent source. It's very context-dependent and very format-dependent.
Agentic commerce: speaking the machine's language
Jeremy Rivera: At the enterprise level we're starting to see agents running around with wallets, purchasing on behalf of users from vendors set up to transact with them. Does being "agent-ready" — e-commerce tooling that bots can transact with — give a brand a competitive edge in being recognized and surfaced for commerce queries?
Federico Fancinelli: Absolutely — we did a deep dive on this when we built the e-commerce branch of our tool. Simply put: if your website speaks the language of the machines, they prefer you. If an agent can transact with you in its own language, it's more likely to choose you — even before product quality comes into play, because there's less of a language barrier. It's like an empty highway: more likely to be driven on. That gap will close fast once everyone does this, but right now there's a real opportunity for e-commerce brands to position themselves early.
Why MCP is now table stakes
Jeremy Rivera: As a SaaS owner, do you just have to accept that something like MCP is now table stakes? I signed up for Open SEO, an open-source SEO tool, and my first question was: is there an MCP? And do you expect knowledge-base MCP structures to become standard on the front end of websites, for both humans and bots?
Federico Fancinelli: MCP is going to be mandatory for SaaS going forward, if it isn't already the actual state of things. We implemented it because I don't believe in a world where not being able to ask a tool questions directly is sustainable for a web service. API connections, and MCP especially — it's not something you can allow yourself not to have. It's fundamental now. Instead of opening an app, navigating to the right section, and analyzing the data, you just ask a question.
The hopeful close: SEO lives inside GEO
Federico Fancinelli: I want to end on a hopeful note. Even small players can compete if the fundamentals are right. Consistency is the key: get your three main pillars of signals — infrastructure, narrative, authority — organized correctly, and you have a real advantage over your smaller competitors and a genuine shot alongside the biggest ones, because very few brands are doing this properly yet. You don't even need to move into gray-hat or black-hat territory — stick to the actual principles and you're in a strong position.
Jeremy Rivera: I used to picture SEO and GEO as overlapping circles in a Venn diagram, but I've changed my mind — I think SEO sits at the center, and GEO wraps around it entirely. Everything we do in SEO is included, and then there's more. Accurate? I'm working on Project Kilby, a data center power solution, and there's so much more beyond the website to optimize for in this ecosystem.
Federico Fancinelli: Precisely — spot on. At its core it's still the same discipline: helping a brand be found online. What changed isn't the name or the discipline, it's the channel — it's not just the Google results page anymore, it's the answers generated by LLMs or by Google itself in AI Mode. Before, it was just the website; now it's the whole ecosystem. The website is one part of what you need to optimize, but there's a much bigger external system you need to take care of before you're recognized as an entity and suggested in these results.
Jeremy Rivera: Reputation matters more than ever, too — LLMs seem to hold a grudge for repeated negative signals in a way classic search never did.
Federico Fancinelli: Brands need to be careful. In the old SEO era, you could just publish another page. Now, it's your entire reputation at stake every time you make a mistake, because of entity recognition instead of page recognition. A recurring negative review is much harder to mitigate than it used to be — it stains your reputation.
[Button CTA] Want to see how your brand ranks as an entity across ChatGPT, Gemini, and Google AI Mode? Try GeoSonar
Final note
Want more research like this, straight from the GEO Academy?
Name
Email
CTA: Get GEO Academy updates