TL;DR

Spam, spam, spam, spam, spam, spam, spam, spam, spammity spam! Google’s August Spam Update (we miss the days when they had fun names – we might call this one Roland) took 2 days and 16 hours to roll out and was built (and was somewhat successful in trying) to tackle the re-emergence of what has come to be known as ‘scaled content abuse’. To sum things up:

• It was pretty bad for those impacted.
• It is still a mistake to use low effort AI content even if it isn’t at scale.
• This is only the latest in what will be an ongoing campaign.
• It is possible to recover, but it’s going to take time.
• Yeah, you can use AI to write your content if it’s heavily edited, but that time and money would be better spent on a good copywriter.

It was pretty bad for those impacted

We’ve mentioned the Mt.AI phenomenon (credit to Glen Gabe for the term) where brands have rapidly scaled content, seen huge gains, released a case study, then been heavily and broadly penalised (manually or algorithmically) to the extent that their hard won previous visibility was lost along with the less hard won visibility they’d gained from the scaled content.

This is documented throughout SEO publications (including by the aforementioned Glenn Gabe), so we won’t go over individual cases – suffice to say, there are plenty of colourful screenshots on blogs all throughout the SEO industry to satisfy even the strongest hunger for schadenfreude. 

It’s still a mistake to use low effort content, even if not at scale

Instead of all the lovely screenshots, which are probably what most people want and are looking for, we’re going to get into the dry technical information that underpins the reason that AI content remains a silly idea if you’re trying to build a sustainable brand and rely in any way on revenue generated by traffic from the organic search channel. Because we’re fun like that.

A few months ago, we wrote a few blog posts on why EEAT was set to be one of the most important aspects of any strong SEO strategy (both in and out of YMYL verticals), and in that article we mentioned something called stylometry and adversarial stylometry (essentially the same thing but with an actively oppositional author) which include methods of algorithmically determining author providence. 

As it happens, one month after we published that blog, Google published a paper titled Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse: A LoRA-Enabled Multimodal Defense System which details Google’s efforts to tackle scaled AI content abuse algorithmically – and it does so in a very similar way. What we overlooked (naively in retrospect) was that it isn’t necessary to begin identifying every author online – it just needs to be able to identify the authorial style and combat adversarial masking of the main LLMs. Does that render our previous points about EEAT obsolete? No. In fact, creating genuine expert entities will be extremely valuable as Google continues to combat AI slop, but what it does do is bring the sloppocalypse much closer than we would have expected. In fact, the published paper states the following impressive benchmark:

The S-CTS system represents a scalable and highly accurate classification and triage defense against coordinated actors exploiting Generative AI to flood platforms. The success hinges on the robust integration of multi-account coordination detection (Bot-Nets) and the utilization of LoRA-adaptedLLMs for nuanced synthetic content classification. Operational evaluations confirm the system’s precision in identifying co-ordinated clusters, and the exceptionally low <1% secondary review discrepancy rate validates the approach’s engineering resilience.

But it doesn’t even end there, their plans for ongoing work are set to incorporate watermarking, track adversarial adaptation and “ensure rapid adaptation of detection models”:

Ongoing research efforts focus on continuous engineering improvements:

• Integration with C2PA and Watermarking: Futurework will ingest cryptographic signals (C2PA) and digital watermarks (e.g., SynthID) as ground-truth features forthe ΨCclassifier, moving from “detection” to “prove-nance verification.”
• Expanding to Deepfake Impersonation: The LLM-driven detection framework will be extended to specif-ically target high-harm deepfakes involving political fig-ures or non-consensual imagery.
• Tracking Adversarial Adaptation: The team will con-tinue to leverage LLMs for daily monitoring to track emerging “Slop” trends, ensuring rapid adaptation of detection models to new foundational models released by the open-source community.

This is only the latest in what will be an ongoing campaign

It’s far from over – anyone thinking that their particular prompt has managed to evade detection is just as short sighted as anyone believing that another workaround isn’t already in the discovery phase. Spam has been an ongoing arms race since the Matt Cutts days and while the technology and techniques have become more sophisticated, the war (to misquote the age old adage of the Fallout games), the war never changes.

Whether it’s SpamBrain or the addition of the Multimodal Defense System, or use of SentenceBERT as part of synthetic copy detection, Google is beavering away behind the scenes to combat not just the production of AI content at scale, but the use of AI content as a whole.

The reasons for this are many and varied, but my guess is that there are two primary driving factors behind Google’s tenacity in this area – and they are:

1. Google search is broken, but remains Alphabet’s main source of income. AI content is only making search less useful, and they’re worried that the limited and decreasing goodwill that they earned in the early days of the internet, and the growing discontent with the performance of search will accelerate the (albeit slow and so far negligible) loss of market share it has experienced in the last few years.
2. LLMs are expensive to feed and the initial training data is becoming increasingly more expensive to obtain. Google’s access to a live and gargantuan data-set, both past and present, has provided them with an almost insurmountable headstart in training LLMs. However, synthetic content leads seemingly inevitably (there is some work which suggests that this can be mitigated by maintaining a significantly large ratio of ‘real’ content to synthetic content, but combatting model collapse is still a science in its infancy) to model collapse – something which is gradually poisoning the well from which Google drinks and from which almost all of its revenue is generated. 

Both of these reasons make the proper discovery and deprioritisation of synthetic content a huge priority for Alphabet. In fact, it’s one of the very few potentially existential problems that Google faces over the next five or so years. Anyone believing that they can ignore such motivations on Google’s side and that their particular prompt produces content which won’t be discovered, may well be in for a surprise.

Final thoughts

There will be a large number of sites which have used mildly edited, or extensively prompted synthetic content which survived this update. That is not an indication that this style of synthetic content use is SEO safe. Instead it means that Google has work to do in developing its detection methods. 

However, that’s not work that they need to do from first principles. The methods of identifying and discounting synthetic content already exist, what is left to do is to find a way to incorporate them into an increasingly complex mix of algorithms that are, themselves, black boxes. When you can only judge the success of your work by live outputs from initial weightings, but have no real way to determine exactly what weighting has produced what outcome or, perhaps more importantly, why that has happened, incorporating new technology becomes an unwieldy and complicated process of trial and improvement. 

That’s where we are at the moment – Google has everything it needs to penalise synthetic content; what it’s lacking is only the implementation step, made difficult by the nature of its own AI powered algorithms. While we may not trust Google as far as we could throw a thirsty data centre, we wouldn’t underestimate their ability and need to overcome the problem of AI slop.