We did it! We finally managed to optimise the entire internet into a coma. It’s 2026, and half of all online content is AI-generated, nobody has read 90% of the human 50%, and there’s a constant deluge of synthetic, average slop gushing out into the internet like it was some kind of British waterway. We, the marketing industry, made this inevitable with a decade of soul-less, keyword-stuffed ‘ultimate guides’. 

However, by the same ‘token’ (I won’t ask you to excuse the pun, nor will I apologise), in the Google leaks, all the way back in 2024, ‘NavBoost’, an algorithm that, among other things and contrary to historic Google denials, uses human clicks to validate trust, appeared 84 times. The Google (and other search engine) algorithm(s) aren’t looking for more content, they’re looking for ways to bypass the noise we created. 

This is the same noise that led people to append ‘Reddit’ to searches and, in turn, to that weird period when Reddit answers ranked for everything. If you’re still chasing AEO/GEO or even SEO hacks instead of looking into Information Retrieval fundamentals, you’re just preparing your site to lose rankings in future.

The systems that have been powering search since the introduction of RankBrain (about a million years ago in SEO terms) are now desperate to find novel, human knowledge. As such, as an industry, we need to stop talking about ‘content’ and start talking about how these systems actually work, because the secret to beating the ‘AI Apocalypse’ (remember mobilegeddon? …simpler times) isn’t a new or better prompt, it’s a return to the science of Information Retrieval, and to human beings who actually know what they’re talking about.

Spam only works until it doesn’t

Don’t get me wrong – the spam works. It always works… for a while. From content spinners to link farms and comment spam, the latest hack will always leap a little ahead of the algorithm. Humans are pretty good at thinking around corners and finding exploits – it’s the reason our ancient ancestors weren’t eaten by sabretoothed tigers. 

It’s also the reason for Penguin, and Panda, and the Helpful Content Updates. Eventually, Google gets tired of black hats shitting in the metaphorical paddling pool, and starts trying to clean up – and that’s why we’re beginning to see study conclusions like the following from SE Ranking:

The takeaway is clear: fully AI-generated content may deliver some initial results, but it’s unlikely to be a good long-term strategy. If you do use AI to create content, it should always be followed by thorough editing, optimization, and other refinements.

But this isn’t why we’re really here – it’s just a symptom of a larger misunderstanding about what LLM influence on search is likely to be. At the moment, businesses are running around like the proverbial headless chickens, all trying to work out a way to game their inclusion in AI answers, when what we’re actually seeing is a transitional stage of search – as we begin to move to a multi-surface world that SEOs at the top of the game have been predicting for more than a decade.

Will AI bring about voice search?

We have all just lived through the death of the highstreet at 3x speed. Only, this time, online. A huge number of bootstrapped sites and (eventually) stores flourished on the early internet, these were then hoovered up (or outcompeted) by an increasingly small number of larger brands, these – in turn – were driven out of business by multinational behemoths with their fingers on the regulatory scales, these platforms gradually then became enshittified (thanks Cory Doctorow), and we’ve started to see a rise in artisanal products and small scale vendors again as consumers grow tired of, and have begun to turn away from, mega-corporations.

What LLMs represent to business-as-usual on the internet is the same as ecommerce meant to IRL stores – they represent a whole new way to do stuff. A way that is currently much shitter than the old way, just as early ecommerce was than IRL shopping (and still is in certain respects). The mistake we’re currently making is trying to map two different maps onto the same landmass, completely missing that they’re from different sections of an atlas.

No, I’m not going to start ranting to you about getting ready for the agentic future. 

You’re safe here.

The chances are, agentic AI will be a thing at some point – but it’s going to need a properly packaged and sanitised, productised agent before there will be any major adoption. Similarly to voice search, it’s going to take an incredible consumer device before anyone is really bothered by it – and that’s probably at least five to ten years off. 

Ironically (is it ironic? Probably not, actually, I’ve gone all Alanis) LLMs might bring about the tech needed to finally make voice search a thing, but it’s not going to be enough for Agents – Amazon can’t even get people to buy through Alexa, and while they explicitly refuted a story that only 2% of owners use it to shop, it has been far less forthcoming on what the number actually is. Nevertheless, a more conversational interface would likely increase trust and, therefore, purchase likelihood – though it is probably still going to take a subvocalisation device for voice to become a major interface.

However, this is an active research area – and the following is the conclusion of research published by Google back in October 2025:

The move to S2R-powered voice search isn’t a theoretical exercise; it’s a live reality. In a close collaboration between Google Research and Search, these advanced models are now serving users in multiple languages, delivering a significant leap in accuracy beyond conventional cascade systems.

To help propel the entire field forward, we are also open-sourcing the SVQ dataset as part of the Massive Sound Embedding Benchmark (MSEB). We believe shared resources and transparent evaluation accelerates progress. In that spirit, we invite the global research community to use this data, test new approaches on public benchmarks, and join the effort to build the next generation of truly intelligent voice interfaces.

Rather than an agentic future, or a future where we’re solely focused on LLMs for all of our needs (I’ll never be able to think that thought again after a recent section on UK daytime TV), what we’re moving toward is multi-surface, multi-platform search – where we’re able to constantly query data available in a whole range of areas. Some of that will lead to websites, some of that will lead to people, some of it will lead to gradually (over years) less frustrating conversations with bots, but all of it requires an understanding of Information Retrieval to give brands the best chances of success.

Data, Information Retrieval and Orangutans

Arguably, one of the ways the early internet failed and the 2.0 era also failed, was in data labelling – we piled absolutely everything onto the shelves of a library with absolutely no attention to genres, authors, topics or, frankly, good taste, and what we have now is (as Pratchett himself noted in its creation) the library of Discworld’s Unseen University, but without a certain ill-tempered orangutan to help us find what we need. 

For a long time, Google has been our ape of choice, but it is becoming overwhelmed by bad data – and the only real solution to that is to return to a fundamental precept of Information Retrieval – data labelling. IBM puts the importance of data-labelling to ML (machine learning) as follows:

Data labeling is a critical step in developing a high-performance ML model. Though labeling appears simple, it’s not necessarily easy to implement. As a result, companies must consider multiple factors and methods to determine the best approach to labeling. As each data labeling method has its pros and cons, a detailed assessment of task complexity, as well as the size, scope and duration of the project is advised.

The internet is a vast sea of unstructured data, and while that has proven to be a boon for LLM companies in the past – giving them access to an almost unlimited amount of mineable content, their own products have begun to poison the well. This leads to an increasing likelihood of model collapse – as this paper from the University of Oxford found:

Our evaluation suggests a “first mover advantage” when it comes to training models such as LLMs. In our work we demonstrate that training on samples from another generative model can induce a distribution shift, which over time causes Model Collapse. This in turn causes the model to mis-perceive the underlying learning task. To make sure that learning is sustained over a long time period, one needs to make sure that access to the original data source is preserved and that additional data not generated by LLMs remain available over time. The need to distinguish data generated by LLMs from other data raises questions around the provenance of content that is crawled from the Internet: it is unclear how content generated by LLMs can be tracked at scale. One option is community-wide coordination to ensure that different parties involved in LLM creation and deployment share the information needed to resolve questions of provenance. Otherwise, it may become increasingly difficult to train newer versions of LLMs without access to data that was crawled from the Internet prior to the mass adoption of the technology, or direct access to data generated by humans at scale.

In reality, it’s unlikely that we’ll ever be able to reverse the damage done to the internet by AI – but what we may be able to do is something similar to the work Google has been attempting for YMYL industries – the great EEATification of the web.

EEATification

In the past, SEOs have attempted to encourage the adoption of Schema markup – and while that is still an option, there needs to be some broad adoption of some kind of data-structuring for the creation of ontologies and taxonomies. Seeing as some progress has been made with schema, it would make sense for it to stem from there, but something needs to happen.

Bad data – in the form of slop content, misinformation, disinformation and more – is at the root of a dozen societal problems, and it doesn’t matter what the future looks like (whether it’s traditional search, AI search, Agentic AI or something between or none of it) bad data is toxic to all possible future environments, and the only way to solve that is through a kind of data environmentalism.

This is a problem that Google has been facing for decades, but which the majority of LLMs are facing for the first time – and so Gemini likely has that ‘first mover’ advantage. While it’s never advisable to assume good intentions in a mega-corporation (Robocop and Terminator both taught us this valuable lesson in the 80s), their revenue depends on a useful index and valuable search results. For that reason, we can assume that Google and Bing will be far more determined than OpenAI and Anthropic to tackle misinformation and the dangers inherent in generative systems – because, unlike the latter companies, this is where their financial incentive lies, not in the promise of godlike superintelligence to relieve venture capital funds of their money.

There is plenty of work going on in academia around the issue of determining text providence, and things like text stylometry, historically aimed at detecting pseudonymous writing, is already at a point where it could be integrated into search algorithms, and would make the detection of AI content far easier than a reliance on em-dashes and emojis (they’re also used to overcoming adversarial stylometry – the process of author obfuscation).

Establishing EEAT may seem like last year’s SEO sales technique, but it is a problem of genuinely generational importance, and as we proceed to the next phase of search, the great EEATification needs to happen and needs to happen soon.

Search Everywhere

Search is just inherently useful – that’s why we have indexes and the Dewey Decimal System, that’s why Google’s Alphabet parent company is one of the most valuable companies in the world. We need to know stuff, and we want to get to know it quickly, without having to search through a whole area of human knowledge. The search engine was one of the most important inventions of the last 100 years, making information instantly available to everyone, but there have, of course, been areas which have been off limits for search – where search engines simply could not reach. This is an issue that ‘agentic AI’ has sought to solve – allowing people to make their entire lives machine readable.

In reality, few people want to give something controlled by somebody else (especially not Cyberdyne Systems) access to all of their information, but – in reality – once we have moved away from the idea that all we need to create god is to give an LLM more stuff to eat, we can begin to move into the operation of pre-trained local models which are less energy hungry, less storage hungry and less likely to publish your personal data online because some dude thought he could build the next killer app without knowing anything about code security.

MCP, for example, will go some way to allowing ‘agents’ to interact with established data sources, meaning that the seemingly abandoned idea of the ‘digital assistant’ could safely and securely act as a ‘search everywhere’ device. With Google’s advancements in visual, audio, video and image search, AI assistants could once again open up a world of knowledge to everyone and allow at least a partial return to the optimism of the early egalitarian dream of the internet. Instant translation of voice and text to help language learners, AR tutorials for DIY tasks generated for your specific problem, a personal assistant to help arrange your day – something that could do what AI was always supposed to do: automating the lengthy processes of sorting, attribution, arranging and categorising information, looking for stuff, while allowing the human to do what we do best – be creative problem solvers and interact with other humans.

This is the next phase of search – search everywhere, allowing humans to be more human as computers compute. In general, we will need something to change more broadly at a societal level to push corporations away from the creation of expensive slop machines, and Google pushing for the EEATification of the web would be a sensible step both for them and towards the better use of ‘AI’ – we need to help our future Librarians filter out the noise constantly emanating from L-Space and, by doing that, we can take a step back towards that early dream of the internet. 

How SEO can (and should) help make this happen

AI content is a thing, and – to greater or lesser degrees – agencies will be pushing for it to become a part of all things content. The promise of increased output and reduced overheads is just too tempting for any self-respecting capitalist organisation to refuse (even if it has consistently been proven to be an empty promise – as this from The Harvard Business Review shows), but what is not inevitable is AI slop. SEOs and content writers have to be making the case at all levels for The Great EAATification project – ensuring that any content produced is well referenced, fact checked by experts and carries the required signals to demonstrate EEAT that we have all become used to over the last five or so years.

While it has been demonstrated that generative search engines do not use schema to understand content, this is very much a yet rather than an absolute, and SEOs need to continue pushing for proper markup and data labelling.

Since SEO became a thing, SEOs have helped to make the internet faster, more mobile friendly, more accessible and more besides. There has even been a feeling that Google has spent 30 years using SEOs to improve the web while simultaneously treating them as an adversary by feeding them incorrect or only partial information (not wholly true, but not wholly false either, as the clickstream stuff demonstrates). What we can say for definite is that SEO has had an outsized impact on the internet – and largely for the better (largely, again, not wholly) – we can use that experience to unlock the next stage of search and, more importantly, help to make sure it’s a future that benefits everyone, not just OpenAI, Anthropic, Palantir, Google and Microsoft.

Conclusion (the task ahead)

The SEO industry is, if not in the driving seat, then at least a very vocal passenger in the development of the internet, and by committing to three simple things, we can help tackle some civilisational problems with mis- and disinformation, all while continuing to help our clients succeed in search.

• Comprehensive Structured Data: SEOs need to look to move beyond basic schema to build rich ontologies that define the relationships between entities in the content of the brands they work with, not just what the entities are. This means using more than just standard Org and Location schema, but trying to make a genuine effort to construct valuable, data rich entities.
• Attribution Best Practices: SEOs need to explicitly cite expert authors with verifiable credentials to maximize EEAT signals, there should be an effort to destigmatise ‘linking out’, with content taking greater pains to be well referenced.
• Focus on Novel Knowledge: SEOs should dedicate resources to generating unique data, original research, and first-hand case studies that AI models cannot replicate – meaning that we have to move away from just trying to ‘do the same but better’. This is also likely to be the best way to feature in any generative search result.

If you’re ready for an SEO service determined to deliver real results, contact MMC Digital today. Let us take the hassle out of SEO, so you can focus on what you do best: growing your business.