Amid bias, cookie-cutter content, and models trained on fragile data, Amanda King explains why brands should focus less on algorithms and more on the uniqueness of their value
Hello and welcome back to SEO Confidential.
These days, everyone is talking about generative AI, about how to “optimize for ChatGPT” or get noticed by chatbots. With some people pitching last-minute tactics and others flooding feeds with bulk AI-generated content, you risk becoming invisible, generic, just like everyone else.
That’s why today’s interview is so valuable, especially if your goal isn’t likes or clicks for their own sake, but real conversions.
Our guest is Amanda King, an SEO consultant active since 2010 and based in Sydney, principal consultant at Floq, an independent SEO consultancy focused on business-driven growth strategies.
Over the course of her career she has built expertise spanning technical SEO, content strategy, conversion rate optimization, data auditing, and reporting. An international speaker and respected voice in the industry, she has appeared at major conferences such as Ahrefs Evolve, SMX Munich, and Search Marketing Summit, and has contributed to leading publications such as Search Engine Land.
In this conversation we tackled very practical topics for your day-to-day work, including: why artificial intelligence, left to its own devices, tends toward homogenization, producing the “average” answer, while it is human intuition, the unexpected connections, that generate real value and differentiation.
Hers is an ode to brand uniqueness, against the rampant workslop.
Then there’s an even more uncomfortable, and fundamental, topic we addressed: who actually trains these models?
Amanda takes us behind the scenes, telling us about the underpaid workers, often in precarious conditions, who validate AI answers (the so-called “digital sweatshop,” a kind of digital factory where people do repetitive, poorly paid work to train and correct artificial intelligence).
Work done in a rush, under pressure, which inevitably turns into distorted, bias-laden data, the very data the models we use every day are built on.
Understanding this mechanism is the key to knowing how, and how much, to trust these tools in your work.
Ready? Enjoy the read, and get ready to see AI in a whole new light.
The risk of “Model Collapse” and the need to bet on trust, authority, and real value
Today many clients seem obsessed with the new generative AI tools and with how to “rank” on ChatGPT or Perplexity. In this new landscape, how do you manage companies’ expectations, and what are, in your view, the metrics that really matter today compared to the old vanity metrics?
There are two main things to consider. The first is measurement: the metrics available on these platforms are still relatively new and, in many cases, not entirely accurate or truly representative of the phenomenon as it’s being interpreted.
The second is the actual weight of LLM-driven traffic: right now it still accounts for a very small share of overall Internet traffic and isn’t making a meaningful dent in Google’s market share. The SparkToro study is a useful reference for digging deeper into this topic.
I often see consultants or agencies who, as soon as they take on an SEO project, jump straight into “solution mode”: they pull out keyword lists, propose technical audits, and start tweaking meta tags. In your experience, is this purely tactical approach a mistake? And if so, what are the real core-business questions any strategy should start from in order to succeed?
For me the answer is yes, always: starting from tactics is a mistake, because even though best practices can work for any company, they won’t necessarily generate the same impact, momentum, and confidence in organic growth.
I always start with the business, because understanding where a company wants to be in the next five years can radically change today’s priorities. That’s almost always where I begin: where do you want to go? What’s your goal?
I also ask what their current indicators of success are, where their advertising and marketing budgets are going, what customers are telling the sales and support teams, how busy the operational teams are over the next six months, and even how people talk about the company internally, meaning the actual language used every day. Before looking at the numbers, I try to build a picture made of context, nuance, and an understanding of the people involved.
SEO often requires complex technical work that can take time and investment, while companies have immediate financial goals. How do you handle SEO best practices when they clash with budget limits, time constraints, and business priorities?
Exactly like that: I weigh the operational and financial aspects within my Eisenhower matrix and in how I prioritize the work.
(The Eisenhower matrix is a method for prioritizing work by dividing tasks into four categories based on two criteria: urgency and importance.
It works like this:
- Urgent and important → do it right away
- Important but not urgent → schedule it
- Urgent but not very important → delegate it
- Neither urgent nor important → eliminate it or postpone it
The goal is to avoid focusing only on what is urgent and to give more room to truly strategic activities, Ed.).
Unless a task has a truly high impact on the business, I always evaluate it in light of the operational and financial aspects, because they carry real costs and can have concrete consequences for the company if they are tackled without considering the bigger picture.
In other words, before pushing an initiative, I always consider the operational and financial implications, so that a seemingly sound decision doesn’t end up having unintended effects on the business.
In your view, what are the risks, especially for less experienced marketers, of delegating the bulk of the work to LLMs without first getting their hands dirty analyzing real SERPs and real user intent?
An LLM’s job is to provide the most probable, most “average” answer possible based on existing information. That’s why it tends to explain what has already been done or what works on average, rather than suggest genuinely new or unconventional ideas. In a sense, it can make thinking more uniform and less creative.
Humans, on the other hand, are capable of sudden intuitions, of connecting seemingly distant elements and taking interpretive risks in their analyses: those “leaps of thought” that often lead to original ideas or innovative strategies, and that an LLM still struggles to truly replicate.
Amanda, today many marketers use ChatGPT (or other AIs) to analyze large amounts of customer feedback, reviews, or data collected in Excel files, hoping to find useful insights. The problem is that AI can sometimes misread the data or even make up conclusions. What’s your method for using artificial intelligence on large amounts of qualitative data reliably, avoiding errors and hallucinations, and getting genuinely useful results?
Hmm… honestly, I’m not sure I have a definitive answer to that. I know people who use NotebookLM, but I’ll be spending the Australian winter building a lot of custom automations precisely to tackle problems like this, so maybe you should ask me again in a few months.
There is one fundamental principle I believe you should follow when using an LLM, though: never trust the first answer. The best method is to push the model to continuously verify what it produces, asking it to check the data, question its own conclusions, and look for errors or contradictions. In other words, AI should be treated as a tool to supervise and validate constantly, not as an automatically reliable source.
There’s a big debate in the industry about whether or not to block the bots that train language models (such as OpenAI’s or Anthropic’s). Where do you stand on this? What are the risks and benefits of letting them into your site?
Today, the benefits seem to outweigh the risks, without much doubt. The point is that the “Pandora’s box” has already been opened: websites and brands have already ended up in the data used to train many AI models, often without companies having any say when it all started.
At this point, shutting yourself off completely would make little sense. If a company wants artificial intelligence to keep knowing its brand, its products, and its most up-to-date information, it needs to keep publishing content and making useful data available. In other words, maintaining a certain openness toward these systems is by now almost necessary to stay visible and up to date in the AI ecosystem as well.
Amanda, when we talk about AI we often focus on prompts or on the quality of the answers. But behind these models there are people who help train them, often under opaque working conditions and for very low pay, as you rightly wrote here. In your opinion, how much does the way Big Tech builds and manages these AI models affect the quality of the answers, the errors, the systems’ biases, and the reliability of the tools marketers and companies use every day?
Thanks for reading the article. As for RLHF, and the way it seems to be managed today, the biggest problem is that, very often, the training data is validated in psychologically unsafe work environments, and this ends up compromising the quality of the information.
(RLHF, Reinforcement Learning from Human Feedback, is a method used to improve the answers of artificial intelligence. In practice, people rate the answers generated by the AI, indicating which ones are correct, useful, or appropriate and which ones are wrong or unconvincing. The model “learns” from these human judgments to give better answers in the future. Put very simply: it’s human beings teaching the AI what a good answer is and what it isn’t, Ed.).
It can happen, for example, that a piece of content gets approved as correct simply because the person evaluating it has to hit a work quota to get paid at the end of the week. Or because, instead of hiring a specialist, the job is given to a generalist who can’t catch the nuances or the more complex errors. In many cases, working conditions push data labelers to move quickly on to the next task, instead of stopping to really analyze what’s in front of them and check it carefully.
The problem is that, if this data foundation is compromised, the model’s subsequent layers also end up being built on distorted information. We know very well that, in the end, LLMs work by recognizing and replicating patterns. If the initial pattern is wrong, the risk of errors increases enormously.
In effect, we are using models trained on huge amounts of information validated by people who often have few real incentives to care about data accuracy, partly because of working conditions that can be highly questionable from an ethical standpoint.
Then there’s another important aspect: the data these people have to evaluate isn’t neutral. It’s already shaped by the viewpoints, values, and biases of the big tech companies they work for, which are largely American and often led predominantly by white men.
Looking at the future of AI, you warned about the looming risk of ‘Model Collapse’: the moment when AI, having run out of human-written content, will start training on its own synthetic outputs and degrading. Estimates suggest this breaking point could arrive as soon as 2026 to 2028. How do you suggest companies prepare for this scenario? How can we build a solid digital presence if the very foundations of language models risk crumbling within a few years?
Well, in my view there are two things to consider.
The first is automation. If you want to use AI to lighten your daily workload and free up more time for strategy, ideas, and important decisions, the right time to do it is now. In the next six months, better yet in the next three. Don’t assume this phase of explosive AI growth will last forever. It’s worth taking advantage of it while the tools work well enough to produce verifiable, genuinely useful results. This window of opportunity may last less than expected.
The second aspect is more important and concerns a brand’s digital presence.
How do you build a strong online identity today?
Essentially, the same way a solid reputation has always been built: by creating a recognizable brand and doing something genuinely useful for your audience. The goal shouldn’t be pleasing ChatGPT, Google, or the algorithms, but offering better content, products, and experiences for people.
In other words, you need to build your business with users first and platforms second. The real challenge is bringing online what makes a brand strong in the real world too: trust, authority, and perceived value.
Behind every AI answer, the value of human intuition endures
What I take away from this conversation with Amanda King is a simple but, today, contrarian idea: AI remains a tool, while strategy comes from the business, from people, from the intuitions only a human being can have.
In a market where everyone risks looking more and more alike, being authentic, recognizable, and deeply rooted in your own business stops being a motivational slogan and becomes a real competitive advantage.
If everyone uses the same prompts, the same tools, and relies on the same “average” answers generated by LLMs, the risk is producing increasingly identical content, increasingly predictable strategies, and increasingly interchangeable brands. Homogenization may be the most underrated danger in AI-driven marketing today.
Then there’s an even more uncomfortable aspect that gets far too little attention: the invisible labor behind these models. Every time we receive a seemingly perfect answer in a matter of seconds, we rarely ask ourselves who actually helped train those systems. And yet, as Amanda told us, behind many models there is an ecosystem of underpaid workers, subjected to intense workloads and often asked to validate enormous amounts of data on extremely tight deadlines: the so-called digital sweatshop.
The problem is not just ethical but technical, and it directly concerns marketers, entrepreneurs, and companies.
If data is validated in a hurry, by people without specialist skills or incentivized to favor speed over accuracy, the risk of bias, errors, and distortions inevitably grows. And if the foundation is fragile, the answers generated by the AI risk being fragile too.
That’s why, perhaps, the right question today isn’t “how do I get ChatGPT to love me?” but rather: how well do I really understand the tools I’m entrusting with a growing share of my work and my decisions?
A heartfelt thank you to Amanda for her time, her generosity, and the valuable insights she shared with us in this interview.
And to you, reader: we’ll see you next week, right here on SEO Confidential, with another exceptional guest from the world of SEO (and beyond), ready to give you practical advice to grow your business.
Stay tuned: it promises to be another episode worth marking in your calendar!
