For years, digital marketing agencies operated on a familiar playbook: identify the top-performing search results for a target keyword, analyse the language and topics those pages used, and craft new content that mimicked their structure closely enough to earn a comparable ranking from Google. It was labour-intensive, error-prone work, and it occupied a substantial share of what is now a multi-trillion-dollar global content marketing industry.

Thomas Reutterer, Professor of Marketing at the WU Vienna University of Economics and Business, and coworkers spent several years asking whether that process could be done better, faster, and at a fraction of the cost. The answer, it turned out, was yes. And the implications for the industry are more complicated than they first appear.

Prof. Reutterer and his collaborators developed what they describe as a "sandwich model": an automated content writing workflow in which a large language model (LLM) sits at the core, surrounded by an application layer that handles crawling, content scoring, and human review. The system scrapes the best-performing organic listings or sponsored ads for a given keyword, passes that material to the language model for fine-tuned content generation, then scores the output against a set of quality metrics before returning it to a human editor for final refinement.

"Even if you do it well, Gen AI outperforms humans," said Prof. Reutterer, who presented his latest research findings at the UNSW Marketing Analytics Symposium Sydney (MASS) 2026. "We demonstrated that, and of course, it's a no-brainer: enormous cost savings are involved, up to 90 per cent efficiency gains. That's why companies and agencies are using this kind of stuff."

Photo gallery: The UNSW Marketing Analytics Symposium Sydney (MASS) 2026

Size is not all that counts

One of the more counterintuitive findings from the research concerned the relationship between model scale and output quality. When the team originally developed the workflow in 2022, they used GPT-2. More recent experiments found that upgrading to larger models produced only marginal gains. The architecture around the model (not the model itself) drove the quality of results.

"The LLM is basically just used as a workhorse," Prof. Reutterer said. "The ball can be replaced by any language model. What makes the difference is the application layer, and specifically the content scoring. You need to have smart metrics for scoring the quality of the generated content."

That scoring methodology evaluates output across several dimensions, including topic relevance relative to the top-ranking pages, keyword integration, readability, and natural language quality. The research was also a catalyst for the development of tools that commercialised the underlying technology. Separately, a wave of commercial content platforms, including Jasper, Copy.ai, and CopySmith, as well as Google itself through its own client-facing tools, brought comparable approaches to market, reflecting how widely the core methodology had taken hold across the industry.

Learn more: A marketing code red: When AI broke HubSpot’s inbound engine

The research also found that efficiency gains from the workflow came primarily through increased visibility. whereby AI-optimised content earned more impressions and greater exposure in search results. "The gains are more coming from an increase in visibility, and not so much in the clicks and in the conversions," Prof. Reutterer said. 

"The trick is you get more volume, and you get more exposure by simply convincing Google to show you more often, but proportionally, we don't get more clicks. The increasing clicks are coming from the visibility and not from the interaction of the customers with the content."

Visibility up, variety down: the similarity crisis

The efficiency gains from AI-generated content are real, along with a second-order consequence that Prof. Reutterer and colleagues described as a "similarity crisis". If every organisation deploys the same class of tool, trained on the same top-10 search results, producing content designed to mirror those pages as closely as possible, the endpoint is a web in which all content risks converge toward a single undifferentiated template.

Research from colleagues at MIT, who conducted a large-scale analysis of AI-based search, found results consistent with this pattern, documenting the loss of information at the edges of the distribution: the niche, the specialist, and the semantically distinct content that gives the broader information environment its depth.

"If the humans are not involved anymore, what does this mean for the industry and for the digital marketing space in the future?"

THOMAS REUTTERER

Prof. Reutterer's own data reinforced the concern. Using a scrape of 388 keywords across four industry sectors collected in 2019 (a period before the widespread availability of large language models) and comparing it with a parallel scrape from 2024, his team found that content similarity scores had increased. A fine-tuned AI classifier designed to distinguish human from AI-generated content found that by 2024, roughly one quarter of the sample web pages contained at least some AI-generated content. The higher-ranked pages were predominantly the AI-generated ones.

"The tools are working: the better ranks are dominated by the AI-generated websites," Prof. Reutterer noted. "But it comes at the price of increased similarity. Everybody tries to mimic the top 10 performers, which might essentially result in a situation where the content is becoming just more and more similar over time, and we are losing variety in what is out there on the websites."

The downstream consequences extended beyond marketing. If future generations of large language models were trained predominantly on AI-generated content that was itself homogeneous, the models began to degrade.

Research cited by Prof. Reutterer demonstrated that after several generations of retraining on their own output, language models experienced what researchers called "model collapse": rather than hallucinating, they produced nonsense. "If the humans are not involved anymore, what does this mean for the industry and for the digital marketing space in the future?" he asked. "I think that's a critical question that will keep us engaged over the next couple of years."

What GEO means for search: early signals

The question of content visibility became even more complex with the emergence of what practitioners began to call generative engine optimisation (GEO): the discipline of getting content to surface in AI-powered search summaries from tools such as Perplexity, SearchGPT, and Google's AI Overviews. The shift from a results page of ten links to a single AI-synthesised answer carried commercial implications for publishers and brands.

If users received a response without clicking through to source pages, the traffic and advertising revenue flowing through conventional search could contract. Prof. Reutterer explained: "The zero-click game is a problem for the traditional online advertising industry, because advertising dollars that are currently in Google will sooner or later go to the LLMs, or at least a substantial part."

Whether the content optimisation principles that governed conventional SEO translated to the GEO environment was an open question. To test it, Prof. Reutterer's team adapted their sandwich model, replacing the step that crawled Google's top-ranking pages with one that scraped responses generated by large language models for the same keywords, then applied the same fine-tuning, scoring, and content-generation process.

Learn more: Marketing in the AI agentic enterprise: Lessons from Salesforce

An initial experiment tracked 24 keywords over a 13-week period, and Prof. Reutterer was careful to frame the findings as preliminary. What it showed, however, was enough to shift his thinking. "Apparently, you can use the same procedure to perform well both in Google and in an AI-assisted search," he said. "The top-10 similarity is still apparently important, also in the LLM space. That was the key message, and it surprised me enormously. But we have also early evidence that this might be contingent on the specific foundation model used by the AI, meaning that we would need to adopt for each LLM separately."

Brand distinctiveness in a world of converging content

For marketers, the practical implication of this finding is twofold. On one hand, the tools already used for conventional SEO could potentially be directed toward GEO with relatively little adaptation. On the other hand, the similarity dynamic already visible in organic search carried a risk of replicating itself in AI-generated overviews. 

The team is currently running a series of experiments to understand whether injecting brand-specific language, proprietary terminology, and tonal variation into AI-generated content could preserve distinctiveness without sacrificing search performance. The hypothesis is that a degree of divergence from the top-10 template was both commercially necessary and technically achievable.

Learn more: Playing the long game: how Wix turned trust into a growth strategy

Whether that balance can be struck and how tools to support it might be designed remains the subject of ongoing research. What Prof. Reutterer's work had already established was that the gains from AI-assisted content were real, measurable, and accessible to organisations willing to invest in the right workflow architecture rather than simply the most powerful model.

The challenge that remains, both for the industry and for the researchers working alongside it, is ensuring that the pursuit of visibility does not erode the information diversity that gave search its value in the first place, Prof. Reutterer concluded.