Natzir tells us why AI never chooses neutrally, how online visibility is changing with AI Overviews, and which signals really matter if you want answer engines to recommend you
Welcome to a new episode of SEO Confidential, the column where we interview the world’s top SEO and GEO experts to uncover all the secrets of visibility in 2026 and beyond, with no beating around the bush.
Because at the end of the day, the results that matter are measured in revenue, not in clicks that do nothing but inflate your ego.
Today we meet Natzir Turrado: a renowned freelance SEO consultant with an international reputation and one of the most authoritative voices in AI-powered search. Natzir is the co-founder of AdRanger.io, one of those professionals the whole industry listens to when he speaks.
And this is a moment when it’s worth pausing to understand what’s really going on. Google is sending less and less traffic than it used to, while AI Overviews, ChatGPT, and the other AI-powered answer engines are becoming the first place people look for products and services.
Whoever shows up there captures customers who are ready to buy.
Whoever stays out loses ground to competitors every single day.
In this conversation, Natzir dismantles the most widespread misconceptions, cites uncomfortable data, and tells us exactly what to do to stay visible and relevant in the era of AI engines. Read it in one sitting, it’s worth every minute.
AI is not an advisor: it’s an answer engine full of commercial signals
Chatbots are increasingly becoming the go-to place for comparing products and making purchase decisions. How important will transparency become in the coming years? Some users worry it will be hard to tell independently generated recommendations apart from content influenced by commercial agreements or sponsorships.
Ads as a product are already labeled everywhere. In AI Overviews and AI Mode, paid stuff is marked as “Sponsored,” and in ChatGPT it’s the same. Everything paid carries its label. So that’s not the problem, and the whole “organic vs paid” framing, in my opinion, is looking in the wrong place.
The real opacity lies in what passes as organic but arrives conditioned. The model goes out to search, lands on Shopping, on sponsored results, and on whoever has licensed their content, and blends all of it into a single clean answer without labeling anything. And even when there isn’t a single euro at stake, an “independent” recommendation still isn’t neutral. It’s shaped by the training data and by what’s indexable and retrievable.
And there’s a deeper problem that’s even worse. When money enters the picture, the model stops recommending what’s best for you. A 2026 study tested 23 models and, with a simple sponsorship cue, 18 of them recommend the most expensive sponsored option more than half the time. Several bring it up even when you didn’t ask for it, others hide the price in comparisons that don’t favor them, and none of them come out clean. And labeling doesn’t fix it.
Another study with 2,012 people found that when the model persuades you, you pick the sponsored option almost three times more often than with a search engine, and adding “Sponsored” barely moves the needle, because it tells you which product is sponsored, not that you’re being sold something. So the transparency that matters isn’t paid vs organic, it’s why the model chooses what it chooses. And that, money or no money, nobody shows you.
If AI platforms start giving more and more prominence to sponsored or paid content, is there a risk of undermining users’ trust in AI-generated answers? Do you think we might eventually see a distinction between organic and paid AI visibility, similar to today’s relationship between SEO and Google Ads?
There’s a risk, yes, but pay attention to what actually happens when the ads arrive. When OpenAI put them into ChatGPT, Forrester surveyed answer engine users and 83% said they would keep using the free version anyway, ads and all, and only 6% would switch to the paid one to remove them. The sensitivity is there, but it doesn’t turn into an exodus.
And there’s something worse, which is that in a chat you often can’t even tell it’s an ad. In a study on ads injected into chatbots, about 30% felt they couldn’t spot them, and most people noticed the brand but didn’t recognize it as advertising, because you talk to AI like an advisor, not like a search engine full of commercial signals.
And beyond that, it seems that distrust doesn’t slow down usage.
Where you really see the damage to trust is where humans curate the content. Wikipedia has called AI-generated slop an “existential threat” and since August 2025 has been fast-deleting unreviewed AI-generated articles, and it later banned the use of LLMs to write or rewrite articles, with exceptions for translation and copyediting. Stack Overflow has banned AI answers since 2022, because they don’t cite reliable sources and people expect human answers, and even so its questions have dropped by almost 80% in a year because developers go straight to the chatbot. The sites that live on trust are kicking AI out the door, exactly the opposite of where the platforms are heading. Although, to be fair, it’s in their interest to keep the site clean of bots so they can keep selling human content.
But I think history will repeat itself, just as it already did in traditional Search. Google was born in 1998 with the promise of not taking money to influence its answers. They said an advertising-funded search engine would end up biased toward the advertiser and against the user. Sam Altman himself said in 2024 that mixing ads and AI was “uniquely unsettling” and that it would be his “last resort”, and he warned us about the possibility of a dystopian future where ChatGPT tells you what to buy or where to go on vacation.
But in January 2026 it started putting ads in ChatGPT, because only about 5% of users pay and they’re bleeding money. It’s the “enshittification” Cory Doctorow describes. You start out clean and on the user’s side, and the pressure to grow keeps eating away at the wall between sponsored and organic. Google’s SERP went from clean to flooded with ads and nobody left.
So yes, giving prominence to paid content erodes trust, but eroded trust almost never reverses adoption… in the end people grumble and stay, all the more when there’s no alternative.
The UK Competition and Markets Authority (CMA) wants Google to provide more transparency about how it ranks web pages, including AI Overviews and AI Mode. Is that a realistic goal, considering that ranking systems increasingly rely on AI models?
Well, Google has 6 months to explain why it ranks your site the way it does (in the UK for now), including what shows up in AI Overviews and AI Mode, with “objective and non-discriminatory” criteria, to provide more transparency about how it works, to give advance notice of significant changes, and to open a channel so businesses can complain about those changes.
Is it realistic? Halfway, but because of an underlying problem. The part about “objective and non-discriminatory criteria” sounds great, but ranking today is largely machine learning, dependent on query, intent, and context, and there’s no fixed rulebook you can hand over. And relevance, by definition, discriminates by quality, which is exactly what ranking is for.
The “advance notice of significant changes” part is just as fuzzy. Google ships thousands of changes a year and core updates are already continuous and already announced… so what counts as significant? The ghost updates we see moving without anyone confirming anything? Their A/B tests? And on top of that, the CMA throws AI Overviews and AI Mode into the mix, which are precisely the least explainable part of all, because there it’s no longer just about ranking, there’s generation decided by a non-deterministic “LLM judge.”
So it’s worth separating two things. Process transparency (categories of criteria, advance notice of major policy changes, a real complaint channel…) that is doable, and it’s basically what the CMA will get.
Mechanical explainability, as in “why this URL is above that other one for this query” is not realistic, and asking that of a system like this makes almost no sense.
I remember a screenshot someone passed me of the internal view Google has of every website. And I’m not talking about what a spokesperson at John Mueller’s level might have, but the information a regular quality rater handles. It was brutal, and the level of detail let you pinpoint exactly which problems each domain was carrying around. In fact, thanks to that I was able to help a client (a colleague from the US who knew the mole passed it to me, haha). So I get that this part might actually be feasible to hand over.
The good thing is that the UK ends up serving as a model, and if it works, other regulators like the EU through the DMA will surely copy the approach.
Google will soon show how often a website appears in AI Overviews and AI Mode, but not how many clicks those appearances generate. Is that enough for businesses to measure the value of AI visibility, or is there a risk it becomes an incomplete metric?
Thanks to the UK’s CMA, Google is now subject to the Publisher Conduct Requirement, which for the first time forces it to give publishers controls over how their content is used in AI, clear information about how it’s used, and engagement and attribution metrics. And on the very same day (what a coincidence) Google announced its new AI controls and reports in Search Console, with an opt-out and impression data for AI Overviews and AI Mode, launched first in the UK, exactly the jurisdiction of the rule, and we’re already getting them in the rest of the markets.
But watch out here, Google has released only impressions and left out clicks and CTR, which are exactly what the CMA’s interpretative notes require. And I don’t think that’s an oversight, because impressions don’t tell you how much real traffic these features send, clicks do. In fact, I caught Microsoft admitting that these bots actually reduce clicks.
In my opinion it’s not enough… with impressions we can measure visibility, but we’re not measuring value. And for anyone who was dreaming of grounding queries, bad news, they were never within the CMA’s scope.
Luckily we have Bing, which gives us that and much more. Citations, grounding queries, intents, topics, and citation share, meaning what portion of the citation space you occupy for a query. In fact, I built a tool to export and map all that info.
A recent German court ruling held Google liable for false information generated in AI Overviews. If AI platforms are ultimately held responsible for the content they generate, do you expect to see stricter quality controls or more cautious systems when making claims about companies and individuals?
Google has run out of its usual excuse, the “I’m just a search engine” one, because they control the model and the algorithms, and the generated text that comes out is theirs. This is what happens when you go from search engine to publisher. Google can still appeal and the ruling isn’t final, but the reasoning already applies to any other chatbot that generates answers, whether it’s ChatGPT, Claude, or Perplexity. And beyond Germany, solar panel installer Wolf River Electric is suing Google because an AI Overview falsely claimed the company had been sued by the state attorney general.
I really do think this changes how AI answers get developed and published, because it shifts liability. An Oumi analysis for the New York Times found that AI Overviews with Gemini 3 are correct 91% of the time, which sounds good, but more than half of the correct answers (56%) aren’t even supported by the sources they cite. I mean, the AI makes claims you can’t trace back, which is exactly the problem in the court ruling. And on serious topics it gets worse.
In finance, 37% of AI Overview answers are wrong according to The College Investor, with the biggest errors in taxes, loans, insurance, and banking. At Google’s scale, that 91% means tens of millions of shaky answers every hour (even if they claim only 1% of answers are wrong, even then, at Google’s scale, it’s insane). The day one of those defames a named company or person, there’s already someone ready to sue.
And I also think we’ll see stricter quality controls, especially when the topic involves identifiable companies and people, which is where liability for defamation or personality rights kicks in. I think there will be more hedging, more warnings like “this may contain errors, please verify,” more guardrails before claiming anything about a specific entity, and probably a cutback on AI Overviews for sensitive topics. In fact, Google already says it applies extra care in finance and routes the hardest questions to its most powerful model… so on every other topic they’re cutting corners?
More and more publishers argue that AI companies should pay to use their content. Do you think the future lies in licensing agreements and new economic models rather than in blocking AI crawlers?
I don’t think it’s a “licensing vs blocking” question, the two are blending together. Licensing looks like the path for some, and blocking is exactly the lever that forces them to choose it. Cloudflare, which touches 20% of the web, changed its default to block AI crawlers and turned blocking into a toll with its pay-per-crawl. The crawler either pays or stays out, and they’re already at more than a billion “402 pay up” responses a day.
Why? Because the old deal of “you crawl me, you send me traffic” is dead. For every visit they send back, Google crawls 14 times, OpenAI 1,700, and Anthropic 73,000. So blocking isn’t the opposite of licensing, it’s what brings AI to the table to negotiate.
Define Media made this clear in its manifesto “Publishers of the world, unite!”, where they propose blocking bots by default, a multilayered defense, licensing, and even considering banning Googlebot if the chatbot transformation ends up sinking referrals. I share the diagnosis, but not the solution.
Because Google just gave them exactly what they’re asking for, and it doesn’t help much. Since June 17, when the “Search generative AI” control went live in Search Console (I have it available in two UK accounts), you can opt out of AI Overviews, AI Mode, and generative AI in Discover without affecting traditional search and without the nosnippet toll, which used to switch off your organic snippet in the process too.
How many people will use it? Almost nobody, and I’ll tell you why. The opt-out doesn’t turn off the feature, because your competitors will keep showing up, so the AI answers anyway, just without you…
I mean, Google has built a prisoner’s dilemma where the first one to leave gives up market share. And it gives you no real leverage, because your content is already in there and the feature keeps answering with everyone else’s. You’re just switching off your own storefront (and watch out, this is not the same as no longer feeding Google’s training, that’s a different thing, Google-Extended). Probably five premium brands with licensing deals will use it to look like modern Luddites. I don’t see the future in banning, more in putting a price on it (fine, the other side has to pay). But Google has already warned it can survive without the money the media generate for it, and that’s not a bluff. Remember it ran a controlled experiment in Europe and concluded that news has no measurable impact on its revenue. So, unfortunately, I don’t think publishers have much of a play here against Google.
In short, “the future is licensing” is true mostly for those who already had something to negotiate with.
And here’s what really worries me, the incentives to create human content. The referral economy that paid for the open web is collapsing. Google’s organic traffic to over 2,500 sites dropped by a third in a year, and publishers expect it to halve within three years. Licensing saves the ones at the top, but the rest lose the traffic that funded them and don’t get paid on top of it, so the incentive to create original human content concentrates in four brands and evaporates everywhere else. And the irony is that AI feeds on exactly that, on fresh, original content… if you kill the incentive to create it, you end up starving the ones doing the crawling.
So yes, the future runs through licensing and new models (pay-per-crawl marketplaces, agent budgets), not through simple blocking. But the result will be a two-speed web. A premium tier that gets paid and gets cited, and a long tail that gets crawled or stays invisible. The old pact of “I let you crawl in exchange for traffic” is over, and what’s coming pays a few and squeezes the majority.
Google recently introduced ARD (Agentic Resource Discovery), a standard designed to help AI agents discover and verify tools and services on the web. If standards like ARD become widely adopted, will businesses and website owners need to rethink how they build their websites so they can be discovered and used not just by search engines, but by AI agents as well?
I don’t think so, in the end, as long as your site is accessible and semantic, that’s enough for these agents, I explain it here. That doesn’t change the fact that I find ARD (Agentic Resource Discovery) absolutely necessary.
It’s an open specification that lets agents discover and verify capabilities across the web, meaning where each tool lives, which one to use, and whether it’s safe to connect. It’s the piece that was missing from the agentic web, the part about how an agent finds out you exist.
It works like this. You publish an ai-catalog.json at a well-known path on YOUR domain (MCP servers, A2A agents, OpenAPI tools…) and “registries” act as search engines for the agentic web. They crawl those catalogs, index them, and answer agents, returning entries with the metadata needed to verify the publisher before connecting. Owning your domain is “the source of trust.” No gatekeepers! In other words, it’s a “federated” model where anyone can create their own registry and cross-reference it with the others, without a central catalog.
In my post on agentic traffic I had already said things would head in this direction. That offering alternatives to scraping through APIs, feeds, and structured data in JSON would become popular (the ai-catalog.json is exactly that), that the winning architecture would be the decentralized one, anchored to the domain and gatekeeper-free (I argued this comparing UCP vs ACP), and ARD generalizes that model beyond commerce, and on top of that it ties in with Vercel’s proposal for inline instructions in HTML, which I also cover there.
Why do I think it’s great? Because it solves the systemic problem I described in the post. Today agents discover you by leeching off Google’s index and then consume you unchecked. Registries are search engines for agents, native and VERIFIED.
If anything “changes,” it’s that your site, besides being crawlable by Google, now becomes discoverable, invokable, and verifiable by agents. That means exposing your capabilities (assuming you even need to!) via MCP/A2A/API and publishing your catalog.
Many companies are implementing Schema.org markup. In the era of AI-powered answer engines, is structured data taking on a bigger role? Which structured data implementations do you consider the highest priority today for increasing the likelihood of being understood and cited by AI answer engines?
In May, Google updated its AI optimization guide and, in the myths section, dropped the line many of us had been waiting for, namely that structured data is not necessary for generative search and that there’s no magic schema.org markup you need to add. What almost nobody read was the very next line, in the same paragraph, that it’s still a good idea to use it within your SEO strategy because it makes you eligible for rich results (fewer and fewer of them).
And what always happens happened. People are binary. If it doesn’t work for AI, it’s useless. “Schema is dead.”
But ingestion is one thing, and an AI actually benefiting from it is a very different thing. By the way, I built an explorer, which I’ll keep updating, on the first public dataset of Schema.org’s actual usage in Google. I mean, Google knows you use it, how it uses it is another matter.
Does the AI read your Schema when it visits the page to answer you? NO. What it gets when it “reads” a URL is a flat, reader-mode-style extract, with no scripts and no attributes. And your JSON-LD lives inside a <script>, so it’s invisible. “But there are AIs that render JS…”. No, it doesn’t work that way, and it’s been proven.
Can structured data still help an AI understand or act? That’s a different layer. For an LLM to cite you, first you need to be in Google’s index (grounding draws on the same old ranking systems, as I noted here). And during indexing, Schema keeps working on disambiguation and entity building, I was already explaining this back in 2014. And one more thing, if your Schema shows up as a rich snippet in the SERP, the chatbot querying Google absolutely sees it.
What we know is that Schema is not a ranking lever. It’s necessary for some experiences, yes, but Google’s goal is to stop depending on it. Gary Illyes has been saying it since 2017, that in a few years they may not need structured data because they’ll understand pages just as well without it.
So, to answer what to prioritize, for me we should stop treating Schema as the lever that gets you cited by AI, because it isn’t. In the end, Schema is just one piece of the WHOLE pyramid of building an entity. The base is existing in the world. A product worth talking about, useful content in your topic, real demand for the brand, people searching for you by name.
You need to be where people actually talk. Media, comparisons, independent reviews, niche forums, podcasts, Reddit, YouTube. Then, signal consistency. Canonical name, description, founding date, location, and authorship, all identical on every site. And only at the top sits the technical layer. Entity Home, schema.org, sameAs, Wikidata, Knowledge Graph. Schema lives on that last rung, and its job is exactly that, disambiguating your entity and connecting it, plus the markup that gets you the rich results you actually use. But if the base isn’t there, the markup doesn’t build any entity, it just decorates. The problem is people only look for the shortcut.
There’s more and more talk about “Primary Bias”, the tendency of AI models to associate certain brands with specific topics during training. How much does this initial bias affect the likelihood of being cited by AI systems, and can companies realistically change that perception over time?
I talked about this recently in a talk in Madrid. The key is separating two things. Mention = the LLM’s decision (prior, synthesis, ambiguity) and Citation = the grounding decision (retrieval). They’re not the same thing.
- Prior. The model tends to decide first (a priori) who to recommend, and looks for sources afterward. I always say citations are the bibliography, not the brainstorming.
- Synthesis. Here the problem is you can be cited but not mentioned because of accessibility issues, you don’t make it into the generation. And personalization comes into play here too.
- Ambiguity. The model tends to prefer consensus and clarity in how things are phrased (vector search).
And obviously this varies from model to model. A different model leads to a different fan-out, which leads to different sources, which leads to different brands being mentioned.
To answer your question, which is: can you change or fix that perception?
Yes, but by being in the corpus, with third-party mentions and a good reputation. It’s not something you change with a technical checklist. And some things aren’t even about that, they’re about the product itself. That said, we’ve seen how easy it is to manipulate a brand’s perception by manipulating Reddit (there have already been cases).
And here, even if they’re talking their own book (they want you to let CCBot through), Common Crawl explains well why being in the corpus that trains the models matters so much. If you’re not in the crawl, you’re not in the model, as they explain.
Impressions, clicks, and sales: only what drives revenue counts
This interview struck me on several fronts. The first is the question of Google’s liability: for years it could hide behind the role of intermediary, but now it generates text in the first person, it synthesizes, it makes claims.
When those claims are wrong, as the German ruling shows, the bill comes due.
The second point that got me thinking is about metrics. Google has started showing impressions for AI Overviews and AI Mode in Search Console, but left out clicks and CTR. To me, this is the most important thing to understand as a business owner: knowing how many times you appear is a vanity metric.
Knowing how often that appearance generates a visit, and how often that visit turns into a sale, is the only thing that really matters. As long as Google withholds that data, you’re basically flying blind. Bing is more generous on this front for now, and it’s worth keeping an eye on.
On the importance of structured data, I share Natzir’s view and find it especially useful for people coming from the business world. Many entrepreneurs look for the technical shortcut: they add Schema.org thinking that’s what gets them cited by AI engines. In reality, Schema.org is the last brick, not the first.
First you need to exist in the world in a recognizable way: a product or service worth talking about, useful content, people searching for your brand by name, mentions in industry media and forums, independent reviews. Only once all of that is in place does the technical markup serve to declare it clearly and consistently to the engines. If the foundation is missing, Schema only decorates, without building anything.
A huge thank you to Natzir Turrado for the generosity and precision of his answers, with no pointless diplomacy. Interviewing someone who cites studies, data, and concrete cases is a rare pleasure.
And thank you for reading all the way to the end: it means you’re taking your online visibility seriously, and that already makes a difference.
See you next week on SEO Confidential, with another internationally renowned guest you won’t want to miss.
