AI is automating the production and drafting that marketing agencies have long billed for, while the strategic side of the work, and the demand for it, has mostly held up. The wholesale replacement of agencies that many expect has not shown up yet in the research.
This draws on peer-reviewed studies of agencies themselves and on independent, non-commercial labor economics on how AI is changing knowledge work across the economy. Every figure names its source, and where a source is a working paper or government data rather than a peer-reviewed study, it says so.
The headline findings about AI and marketing agencies
Three peer-reviewed studies of agencies independently find the same shift, from execution work toward strategic and advisory work (Wahid et al., 2025; Lanfranchi et al., 2026; Cramer & Rudeloff, 2026).
In one of these studies, managers described AI as “not, as yet, a silver bullet,” because its output almost always needs human review (Cramer & Rudeloff, 2026).
Generative AI raised worker productivity about 15% on average, with the largest gains, around 30%, going to the least-experienced workers and little or none to the most experienced (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025).
Workers with two months of experience using the AI matched the output of workers with more than six months of experience who did not have it (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025).
Early-career workers aged 22 to 25 in the most AI-exposed occupations saw a relative employment decline of about 13%, rising to 16% on more recent data, while more experienced workers showed no comparable decline (Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab, 2026, working paper).
Only 19.1% of a nationally representative US sample used deliberate prompting strategies (Xue et al., Scientific Reports, 2026).
Across 57 models tested on realistic, multi-step tasks, none scored above 15% accuracy (Yu et al., ICLR 2026).
Most dedicated tactics for improving a page’s visibility in AI answers were ineffective or lowered its ranking, and conventional ranking work performed better (Puerto et al., C-SEO Bench, NeurIPS 2025).
Gemini gave no clickable citation in 92% of its answers, and ran no search at all on 34% of queries (Strauss et al., Data & Policy, 2026).
What the agency research finds
So far, three peer-reviewed studies have looked directly at agencies, using interviews and expert panels. Broadly, they reach the same conclusions.
A study in Industrial Marketing Management, based on interviews with 22 people across content marketing agencies and their clients, describes generative AI as a democratizing force that hands the creation, technical, and economic capabilities agencies used to hold to almost anyone.
As a result, clients increasingly treat standard deliverables as a commodity and question paying agency rates for work that now takes far less time. The study documents agencies responding by moving from producing content toward providing creative and strategic direction, packaging their own AI expertise as a service, offering personalization at a scale clients cannot manage in-house, and putting named human experts at the front of their work (Wahid et al., Industrial Marketing Management, 2025).
A Delphi study of 14 experts in the European Journal of Innovation Management reaches a related conclusion. Adopting AI does not by itself create an advantage, because the tools are available to every competitor. The experts locate the advantage in combining AI with resources that are hard to copy, such as proprietary data, specialized know-how, and an adaptive culture, and they single out prompt engineering as the scarce new skill (Lanfranchi et al., European Journal of Innovation Management, 2026).
Consistent with that, a nationally representative survey of 937 US adults found only 19.1% used deliberate prompting strategies (Xue et al., Scientific Reports, 2026).
A study of 13 senior managers at communication agencies in Corporate Communications found AI already present across every value-creation activity, though unevenly. Its use is heaviest in operational content work and lightest in strategic advising, which the managers identified as the part of their work least exposed to AI and most likely to differentiate them. Their overall assessment was that AI is “not, as yet, a silver bullet,” because its output almost always needs human review and revision before it reaches a client (Cramer & Rudeloff, Corporate Communications, 2026).
Across the three studies, the change is most acute in perceptions of what agencies get paid for. Production work is becoming a commodity, and these studies place the value that remains in strategic and advisory work that the tools do not replace.
The numbers behind the thesis
The agency studies are qualitative, built on interviews and expert panels, so they describe the shift without measuring it. Two independent, non-commercial labor-economics studies measure the same dynamic across the wider economy. Neither is about marketing agencies, so they’re best read as evidence for the general mechanism the agency studies describe rather than measurements of agencies themselves.
The first, published in the Quarterly Journal of Economics, studied more than 5,000 workers given a generative AI assistant. Productivity rose about 15% on average. The gain was concentrated among the least-experienced workers, at around 30%, and was close to zero for the most experienced, who saw only small speed gains and small quality declines. The authors attribute this to AI capturing the tacit knowledge of the best performers and making it available to everyone else.
In the same study, workers with two months of experience using the tool matched the output of those with more than six months of experience without it (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025). The tasks that this affects most are the ones a junior would otherwise spend time learning, which in an agency is often the production work rather than the strategic direction.
The second, a Stanford Digital Economy Lab working paper using payroll data on millions of workers, tracks the employment side. It finds no widespread displacement across the economy, but a widening gap for younger workers. Early-career workers aged 22 to 25 in the most AI-exposed occupations saw a relative employment decline of about 13%, rising to 16% on more recent data, while more experienced workers in the same occupations showed no such decline. The decline runs through reduced hiring rather than layoffs, and it concentrates in occupations where AI substitutes for tasks rather than complementing them.
The authors are explicit that these are descriptive patterns rather than proof of cause (Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab, 2026, working paper). Both studies describe the pattern the agency research points to. The measurable effects land on entry-level and production-heavy work, and the studies find little effect on more experienced workers. For agencies that bill mainly for production, this is the pressure the interviews describe, which is visible in economy-wide data.
Why AI is not a silver bullet for marketing agencies
The same research also documents clear limits on what AI does. Agency managers reported that AI output almost always needs human review, that a fast draft is no saving if someone then has to spend time correcting it, and that data privacy and quality control constrain how far they will let it into client work (Cramer & Rudeloff, 2026). The Delphi experts added that AI lacks the empathy and register that persuasive work depends on, and that heavy reliance on it risks eroding the writing and thinking skills agencies sell (Lanfranchi et al., 2026). A benchmark study that evaluated 57 models on realistic, multi-step tasks found that none scored above 15% accuracy (Yu et al., ICLR 2026).
Where the new expertise goes
If production is commoditizing and strategy is the differentiator, agencies need new services to sell. The agency studies point to strategic advising, AI-enabled service design, and personalization.
A concrete and fast-growing example is managing how a client appears in AI search. As buyers move product research into tools like ChatGPT and Google’s AI answers, whether a brand is mentioned and recommended there is something clients cannot see in their own analytics and cannot easily manage in-house. A study of about 14,000 logged search-assistant conversations shows how opaque these systems are, finding that Gemini gave no clickable citation in 92% of its answers and ran no search at all on 34% of queries (Strauss et al., Data & Policy, 2026).
That work is measurement and strategy rather than production, which is the kind of service the research suggests that agencies now need to build. A benchmark study of methods for improving a page’s visibility in AI answers found that most tactics were ineffective or even lowered its ranking, that conventional ranking work performed better, and that gains shrank as more competitors adopted the same methods (Puerto et al., C-SEO Bench, NeurIPS 2025).
How this fits with our own research at Traqer & Grow and Convert
The academic evidence matches what we’ve found running AI visibility for clients at our agency, Grow and Convert. Over the past year we tested the on-site tactics that dominate most GEO advice, including llms.txt files, FAQ sections, key takeaways, and headings rewritten as questions, and none made a measurable difference to our clients' visibility in AI search. That matches the study. The tweaks meant to help a model read a page did little, while ranking the page and earning mentions elsewhere did more.
What worked instead was exposure. Clients appeared in AI answers when their owned content ranked in traditional search for the topics buyers ask about, and when they were mentioned on the third-party sites those models already draw from. We organized this into a framework we call Prioritized GEO, which puts owned content and off-site mentions ahead of on-site tactics.
We also measure visibility differently from the standard blended score, tracking it per model and at the topic level and separating brand mentions from citations, because a single overall percentage hides where a brand is actually gaining or losing ground. That is consistent with the attribution research above, which shows how rarely these systems cite their sources and how much their behavior differs from one model to the next. Measurement and strategy work are what the studies suggest will hold value for agencies.
The limits of this evidence for marketing agencies
The evidence we’ve featured in this article has some clear limits:
The three agency studies are qualitative, based on 13 to 22 experts each, and drawn largely from Europe, so they describe patterns rather than measuring their size, and may not transfer cleanly across different markets.
The two labor-economics studies are not about marketing agencies. One studies customer-support workers and the other covers occupations across the economy, so they evidence the general execution-versus-expertise mechanism rather than the agency case directly.
The Stanford employment study is a working paper, not yet peer-reviewed, and its authors present its findings as descriptive rather than causal.
That said, it is clear that across the agency studies and the labor data, the effect concentrates on production and entry-level work, while strategic and advisory work shows little of the same pressure. The studies also agree that AI’s output still needs human oversight.
Sources:
Wahid, R., Mero, J., & Ritala, P. (2025). Technology-enabled democratization: impact of generative AI on content marketing agencies. Industrial Marketing Management, 131, 1 to 16. Peer-reviewed. https://doi.org/10.1016/j.indmarman.2025.09.007
Lanfranchi, G., Cioli, A., Amanti, A., & Marinelli, L. (2026). Reconfiguring competitive advantage: a generative AI resource framework for dynamic content marketing adoption. European Journal of Innovation Management, 29(3), 770 to 798. Peer-reviewed. https://doi.org/10.1108/EJIM-03-2024-0317
Cramer, A., & Rudeloff, C. (2026). Not a silver bullet (yet): exploring artificial intelligence and its determinants in communication agencies. Corporate Communications: An International Journal, 31(1), 71 to 87. Peer-reviewed. https://doi.org/10.1108/CCIJ-01-2025-0012
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889 to 942. Peer-reviewed. Also NBER Working Paper 31161. https://www.nber.org/papers/w31161
Brynjolfsson, E., Chandar, B., & Chen, R. (2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab (working paper, not peer-reviewed). https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine/
Puerto, H., Gubri, M., Green, T., Oh, S. J., & Yun, S. (2025). C-SEO Bench: does conversational SEO work? Advances in Neural Information Processing Systems 39 (NeurIPS 2025), Datasets and Benchmarks Track. Peer-reviewed.
Yu, P., Liu, W., Yang, Y., Li, J., Zhang, Z., Feng, X., & Zhang, F. (2026). Benchmarking LLM tool-use in the wild. Proceedings of the International Conference on Learning Representations (ICLR 2026). Peer-reviewed.
Xue, H., Oh, Y. J., Zhou, X., Zhang, X., & Oxley, B. (2026). Users' prompting strategies and ChatGPT's contextual adaptation shape conversational information-seeking experiences. Scientific Reports, 16, 12112. Peer-reviewed. https://doi.org/10.1038/s41598-026-42465-4
Strauss, I., Yang, J., O'Reilly, T., Rosenblat, S., & Moure, I. (2026). The attribution crisis in LLM search results: estimating ecosystem exploitation. Data & Policy, 8, e15. Peer-reviewed. https://doi.org/10.1017/dap.2026.10064