SEO Confidential – Our Exclusive Interview with SEO Expert John Campbell on How to Stand Out in AI Answers

Between authority, reputation and control of the narrative, brands are now playing for traffic, trust and conversions inside AI-powered answer engines

Millions of searches no longer reach their destination. They never land on any website, never generate a click, never drive traffic. They are intercepted, synthesized and delivered to the user directly inside an answer generated by artificial intelligence.

ChatGPT, Perplexity, Gemini, Claude: the new gatekeepers of online visibility don’t show ten blue links. They show an answer. And in that answer, either you are there, or there is no alternative.

For brands, this is an epochal shift and, if you read me regularly, you know how carefully I weigh my words.

So the question is no longer just “how do I rank on Google?”, but “can I influence what a language model says about me when a potential customer asks for a recommendation?”

Welcome to a new episode of SEO Confidential, the space where I sit down with the most authoritative and internationally recognized search professionals to understand, with no filters, where this industry is heading.

Today I have the pleasure of talking with John Campbell, Head of Innovation and AI at ROAST, a British integrated performance media agency, and one of the sharpest voices in the debate on AI search and Generative Engine Optimization.

In this interview we cover everything an entrepreneur or marketer needs to understand right now to avoid disappearing from AI answers: how to build authority that language models recognize, how to measure the real business impact, how to defend your own narrative from unreliable third-party sources, how to read share of voice without being fooled by the numbers, and what to make of the ads coming to ChatGPT.

No abstract theories. Only concrete answers, from someone who works on these problems every day with real clients.

Enjoy the read. 👇

Against Negative GEO: owning the narrative with authoritative, proprietary content

More and more searches are now satisfied directly inside AI-generated answers, without users ever visiting a website: what should a brand concretely do to avoid disappearing and truly manage to influence those answers?

In my view, brands should focus on three key areas. The first is building solid, authoritative content on their own website: this is where language models draw from directly, giving the brand the chance to influence the generated answers and earn citation links.

The second concerns the brand’s presence on third-party sites cited by language models, such as news outlets, review sites and blogs: it is essential that the brand is mentioned in a positive, accurate and consistent way.

The third is maintaining an active presence on social platforms, since language models are increasingly drawing on content published on social networks as well: it is far better for the brand’s official content and accounts to be picked up, rather than leaving that space in the hands of others.

Many are hunting for “tricks” to get cited by artificial intelligence, but without a solid technical foundation, structured data and real authority, brands don’t even enter the radar of language models: in your experience, what clearly separates a brand that gets selected by AI from one that gets ignored?

The brands that get selected by artificial intelligence are the ones with a strong, well-established association with the products or services they offer. This is nothing new: it has always been the case. When people think about vacuum cleaners, they think of brands like Dyson, Shark or Henry Hoover. That association has been built over the years through marketing activity, from above-the-line campaigns all the way to digital. Language models simply reflect this reality.

That’s why, while it is important to fix the technical side and implement structured data, the brands that get recommended are, in essence, the ones that have already built a deep association with their product. If that association is weaker than your competitors’, in the short term you can try to lean harder on LLM-specific tactics, but in the long run that alone will not be enough.

There is a lot of talk about “being present in AI answers”, but few connect that visibility to concrete business results: how should companies actually measure its impact on revenue?

There are three main groups of KPIs to monitor.

• The first concerns visibility and citation metrics tracked through an AI monitoring tool, such as share of voice or citation rate. These are important indicators, but they should be read with caution, given the non-deterministic nature of language models, which means results can vary from one query to the next.
• The second concerns website traffic coming from AI citations, that is, the trackable clicks generated by cited sources. There is, however, a behavior that is harder to monitor: the case where a user discovers a brand through ChatGPT and then searches for it directly on Google, making attribution genuinely complex.
• The third concerns conversion attribution. In B2B this means, for example, adding an option like “How did you find us? Through ChatGPT” to lead generation forms, as well as making sure the sales team asks prospects directly. In e-commerce, it means analyzing sessions coming from AI sources through first-click, last-click and multi-touch attribution models.

A brand’s presence in the answers generated by language models is often described in terms of “share of voice”: how can it really be monitored, and what signals indicate that a brand is gaining ground on its competitors?

Share of voice is a useful metric, but it needs to be monitored alongside other indicators, such as clicks and conversions attributed to AI search, and not looked at in isolation.

An important aspect to keep in mind is the relationship between the number of competitors in a given market and how many of them language models actually tend to recommend. For example, if a category has ten main players and ChatGPT typically suggests five, the fundamental question is just one: is the brand among those five?

This is exactly where share of voice as a metric starts to show its limits, though. If, say, there are only five companies offering a certain service and ChatGPT tends to recommend five, visibility will naturally look high, but that number, on its own, tells you very little.

In these cases you need to go beyond mere numerical inclusion and focus on the position occupied within the answer, as well as the sentiment with which the brand is described. In other words, share of voice must be interpreted carefully, always in the context of how competitive the category is and how many brands language models normally tend to surface.

The market is full of promises about how to “get into ChatGPT”: what are the most serious mistakes you see companies make when they chase these shortcuts?

There are a couple of mistakes that stand out clearly. The first is the obsession with LLMS.txt files. Until one of the major models formally confirms that it actively uses them, their value remains limited, even though they are simple to create and carry no particular downsides. The underlying logic doesn’t fully hold up either: the idea that models should rely on a specific file created by the site owner, when they could simply crawl a well-structured site, seems technically unconvincing. If the information is already on the site, the crawler will find it anyway. Just think of how hard it has been, even for Google, to push site owners to create robots.txt files and sitemaps: encouraging the creation of yet another separate file risks adding complexity with no real benefit.

The second mistake is the overuse of listicle-style articles on your own site, particularly those where the company ranks itself first and pads the ranking with commentary heavily skewed in favor of its own brand over competitors. There are already documented cases where this strategy has backfired. Since Google is actively working to penalize this kind of content, it is likely that its influence on language models will follow the same direction.

That said, comparative content can work well in specific contexts, especially when the comparison involves your own product and a direct competitor that is very close in features, and the goal is genuinely to help the customer make an informed decision. In those cases, the key is to stay balanced and fair, basing the comparison on features and hard data rather than opinions or biased judgments.

With advertising entering chatbot experiences, the line between answer and promotion is becoming increasingly blurred: how much could this affect, over time, user trust and brand credibility?

We’re already seeing it… Anthropic has positioned Claude’s ad-free experience in direct contrast with the direction ChatGPT has taken. From a trust perspective, the impact could be significant. People tend to share far more personal and sensitive information with AI assistants than they ever shared with a traditional search engine.

That’s why, when ads start to appear, or when users begin to suspect that their conversations may be used to profile them for advertising purposes, the erosion of trust risks being stronger than what happened in the past with Google Search, which always felt more transactional and less personal.

The only element that could soften this effect is that the users who will actually see the ads will most likely be those on free plans, that is, people generally less invested in the platform and perhaps more used to ad-supported experiences. It remains to be seen whether that will be enough to contain a broader reputational damage, but it is certainly an interesting dynamic to watch.

Advertising inside ChatGPT guarantees visibility as long as you pay, but leaves no lasting value: in your view, when does investing in ads truly make sense, and when does it risk turning into an expensive dependency?

It is still too early to establish with certainty what the real impact of these ads will be, but the most accurate mental model, for now, is probably to think of them as a display channel: broad reach, brand awareness and the ability to put your message in front of a wide audience. Targeting precision will likely be limited, and control over the prompts or the users next to which the ads will appear will be minimal, at least in an initial phase.

There is also a structural limitation that is particularly relevant for B2B marketing: if the target audience is made up of professionals who use AI tools seriously, it is very likely that they have paid accounts, and therefore will not see the ads at all. In that scenario, the channel simply risks not reaching the right people.

More generally, brands are already fully aware of the risks of over-reliance on Google Ads, and that lesson is shaping how they look at new platforms. Few will be willing to build another expensive dependency from scratch. If anything, this concern is accelerating the push toward organic visibility in so-called GEO, that is, the ability to win space in AI-generated answers through real authority and content quality, instead of paying for a presence that disappears the moment the budget stops.

Data shows that when a site loses visibility on Google, it also tends to quickly lose presence in language model citations: have you observed this phenomenon, and if so, what are the first warning signs?

Yes, I have observed this phenomenon too. Across several clients, comparing average citation scores with average organic rankings on Google Search and Microsoft Bing, a very strong correlation emerges. When a site ranks well in organic results, citation scores in AI answers generally tend to follow the same trajectory.

The nuance, however, lies in how these signals show up. When a Google algorithm update hits a site, the effect is immediately visible across the entire keyword set inside Google Search Console. LLM visibility monitoring, on the other hand, is usually based on a much smaller number of prompts, perhaps a few dozen. Although these prompts are correlated with organic rankings, they do not map onto them perfectly. That’s why the correlation doesn’t always appear so clearly, especially if the set of monitored prompts is too narrow.

What has been observed, however, is that citation scores can also drop for reasons completely unrelated to Google, and this opens up a new category of risk. In one specific case, a significant drop occurred when Perplexity changed the number of URLs it displayed for each result. The client habitually appeared with six or seven URLs per prompt; when the platform reduced that number, the citation score fell sharply. It had nothing to do with content or Google rankings: it was purely a platform-level decision. This is exactly what brands need to prepare for: LLM-specific algorithmic changes, independent of traditional search, that can materially affect visibility with no warning at all.

When a brand doesn’t clearly control its own narrative, AI tends to fill the gaps using third-party sources that are often unreliable: what can a company do to prevent this kind of “alternative narrative”?

There are several approaches. The first is making sure your website offers complete, accurate and authoritative content, because every information gap inevitably tends to be filled by others. People turn to platforms like Reddit and the like precisely when they can’t find an answer on the official site. At that point someone else provides an answer, which may be inaccurate or completely wrong. If that content gains visibility and engagement, it ends up becoming part of the narrative that language models absorb over time. Getting ahead of this risk with clear, reliable, proprietary content is the most direct defense.

The second aspect concerns affiliate programs, a topic that is often underestimated. If a brand doesn’t have an affiliate program, third-party publishers and review sites have little or no incentive to include it in their recommendations, because they would get no financial return from the traffic they send. This means that precisely those sources, often cited by language models, will end up excluding the brand from the conversation.

As a result, affiliate strategy is no longer just about generating direct conversions. Today it is also a strategic lever for influencing which sources talk about the brand, how they describe it and, ultimately, whether or not the brand will manage to appear in AI-generated answers.

Language models can absorb and repeat false or defamatory content, presenting it as reliable: how real is the risk today of reputational attacks through Negative GEO, and how can they be detected early?

This is a genuinely complex issue. On the detection front, there are tools that can help, such as fact-checking layers capable of verifying the claims contained in the answers language models generate about the brand. This makes it possible to catch and prioritize the prompts where inaccuracies emerge, giving a clearer picture of where the narrative is starting to drift.

The point, though, comes back once again to content. If models look for information about a brand and can’t find it on the company’s official properties, they will inevitably draw on whatever is available elsewhere. So the defense remains the same: proactively own your narrative with accurate, complete and authoritative first-party content.

The harder truth concerns prompts that deliberately look for negatives, such as “what are the weaknesses of product X” or “what are the complaints about brand Y”. In those cases there is limited room for intervention, because the model is explicitly searching for critical content. The most honest answer is almost brutal: don’t have a bad product. Easy to say, much harder to do, because product and process improvements take time.

In the meantime, the most effective approach is to make sure that any legitimate criticism is offset by a much larger volume of accurate, authoritative and positive content, so that even when the model goes looking for negatives, it does so within a solid, well-structured information base.

Traffic from language models is still limited in volume, but it tends to convert far better than other channels: why are these users so ready to act?

It’s a very interesting phenomenon and it probably comes down to several factors.

The first is that users put more effort into crafting the prompt than they ever did with a regular search query. That extra effort creates a kind of sense of ownership over the result: there is almost a satisfaction in having asked the right question and having been pointed toward something genuinely relevant. That attitude then carries over into the on-site session, influencing engagement levels and propensity to act.

The second factor, perhaps even more important, is that language models seem to do a better job of connecting users to what they actually need. The old behavior of opening the top three organic results, scanning them quickly, going back and refining the search is largely compressed or eliminated. By the time the user lands on the website, a good part of the discovery and selection process has already been completed. In other words, they arrive at a more advanced stage of the decision journey.

The third element, harder to measure today, concerns the role of memory and personalization. If language models are using a user’s history, preferences and context to shape recommendations, that would help explain why the resulting traffic is so well qualified. There is still no full visibility into how much this mechanism is applied and how consistently, but if the level of personalization is high, it accounts for a significant part of the explanation.

YouTube, Reddit and first-hand experience sources are becoming increasingly central in language model answers: from this perspective, how is SEO work changing for brands that want to capture high-value traffic?

On the YouTube front, SEO work is bringing search professionals ever closer to the video production process itself. Today the brief increasingly starts from research: the video is made because it answers a specific query that surfaces both in traditional search and in language model answers. This is an important shift from how many companies have historically treated YouTube, that is, mainly as a brand awareness or product showcase channel. What is changing, in essence, is the intent that drives content creation.

As for Reddit, there is a very concrete risk: doing what SEO sometimes tends to do, which is finding a channel that works and progressively ruining it through over-optimization. If teams start populating discussions with brand-friendly content designed to influence language models, the community notices, the quality of the conversation drops and the entire signal loses value.

The smarter approach is to treat Reddit as an observation tool rather than a distribution channel. Listen to the conversations unfolding in relevant communities, understand the authentic questions users are asking and use that information to build content on your own website. In this sense, Reddit should guide your editorial strategy, not become a space to manipulate. A useful tip is to use tools like Reddit Pro and Reddit Answers to better monitor conversations and capture valuable insights.

For a smaller subset of brands that already have an established subreddit, there is certainly a legitimate space for involvement, but it should sit much closer to community management and customer care than to artificial content influence. The moment the presence looks clearly engineered, it immediately loses effectiveness. SEO often has a tendency to over-optimize what works until it becomes irrelevant: this is one of those contexts where restraint and a sense of proportion will make the difference.

Stand out in AI answers without depending on ads

Talking with John and exchanging opinions and points of view with him is always an extremely valuable experience.

The debate around AI search is full of hype, easy promises and tactics that sound great in slide decks but barely survive contact with reality. What emerges clearly from this conversation is instead something far more solid and, in some ways, more reassuring than you might expect.

The fundamentals still matter. If anything, they matter more.

Language models are not reinventing the rules of visibility: they are amplifying them. Those who have built, over time, a real association between their brand and a product or service, through authoritative content, consistent presence and earned reputation, are already at an advantage.

Those who chased shortcuts and surface-level optimizations, on the other hand, find themselves today exactly where they were yesterday: invisible.

This doesn’t mean there is nothing new to learn. It means the starting point remains the same as ever: being truly useful, truly relevant, truly trustworthy. And then working to make that emerge in the new contexts too.

What I find particularly valuable in John’s approach is his ability to distinguish between what is urgent and what is important. llms.txt files, self-serving listicles, last-minute tactics: useful, perhaps, at the margins, but never a substitute for a serious content strategy and a consistent presence.

Then there is one theme that struck me more than the others: the unattended narrative. Every information gap a brand leaves open gets filled by someone else, by Reddit, by third-party reviews, by inaccurate or biased content.

And over time, that content becomes part of what language models absorb and give back. Owning your narrative with accurate, complete first-party content is no longer just good SEO practice: it is a form of active reputational defense.

Finally, the matter of ads on ChatGPT. Personally, I remain convinced that the real competitive advantage in the coming years will not come from those who spend the most on paid visibility across AI channels, but from those who manage to build organic authority in a genuine, lasting way.

Advertising buys presence.

Trust cannot be bought.

A sincere thank you to John for the clarity and precision with which he shared his perspective: concrete, free of rhetoric and, above all, immediately actionable. Exactly the kind of conversation SEO Confidential exists for.

If you found this episode useful, share it with someone facing the same challenges. And if there is a topic you would like to see covered in the next episode, write to me: the best themes often come from readers.

See you at the next episode of SEO Confidential. 👋

#fullsteamahead