How to Measure the Geography of AI Adoption and Diffusion? Reflections from GEOINNO 2026

AI activities and their spatial patterns

In my view, discussions about the geography of AI need distinguish between different AI activities and technologies, as these follow different location patterns and play different roles in adoption and diffusion. Companies often combine multiple activities and technologies, but analytically it is useful to separate these activities to better understand observed spatial patterns.

First, there is the development of generative AI foundation models and the associated infrastructure that enables this activity, including data centres and specialised hardware production. This type of AI activity is highly concentrated and requires large amounts of capital, computing capacity, energy, and specialised talent. At present, the development of foundation models is concentrated in a small number of locations, in particular in the United States and China (see our foundation model ranking). In response to this concentration and the strategic importance of large-scale compute, Europe has started to invest in AI factories. One paper presented at GEOINNO 2026 entitled 'Varieties of AI factories in the European Union' and presented by Nicole Lemke examined the characteristics and location of these AI factories in Europe and provided an empirical basis to analyse how this emerging infrastructure may shape the geography of advanced AI development in the European context. For me, this raises an important question for future research: to what extent will investments in AI factories act as catalysts for new AI-related activity, and whether they will lead to clustering of AI development around these facilities. This is relevant for understanding whether compute investments translate into new firm formation, spatial clustering, and measurable changes in regional AI activity.

Second, there is the development of generative AI tools, applications, and services built on top of existing foundation models, often by small teams or individual developers. This activity focuses on applying available generative AI models to specific use cases and contexts rather than on building new models from scratch. Compared to generative AI foundation model development, this activity is more footloose and can, in principle, take place virtually everywhere. Based on my experience with RankmyAI data, this type of activity plays a key role in the diffusion of AI across sectors and regions. It also creates opportunities for developing countries, which may lack the infrastructure and capabilities required for compute-intensive AI development and hardware production. At the same time, the economic contribution of this group of AI-first initiatives varies widely. Many tools add real economic value by lowering costs, improving productivity, or enabling new services. Others appear to generate limited value beyond providing an additional source of income for their owners. This becomes particularly visible in the RankmyAI Nano Banana AI tools ranking, which showed that many generative AI tools function as wrappers around the original Nano Banana image model, generating attention and income without clear additional functionality. Although this type of AI activity represents the largest group of AI-first initiatives globally, it was only marginally discussed at the conference.

Third, there is AI consultancy and AI solution development aimed at supporting the integration of AI into existing organisations. This activity focuses on translating AI technologies into operational use within internal processes of incumbent firms and their products and services. From a geographical perspective, this activity either takes place close to concentrations of economic activity and potential users, or is delivered offshore from locations with lower labour costs to international clients. In my view, this type of AI activity plays a central role in the broader adoption and diffusion of AI across incumbent firms by providing AI strategy development, solution design and implementation, and support with governance and compliance. During the conference, this activity received limited attention in the paper presentations.

Fourth, there is the development of AI solutions based on more ‘traditional forms of AI’, such as robotics and computer vision. This activity often focuses on domain-specific applications (such as AI fish monitoring) and is often closely linked to university research, industrial partners, and existing sectoral clusters. Its location patterns may reflect longer-term processes of industrial specialisation and knowledge spillovers. At the same time, our analysis of the Norwegian AI ecosystem suggests that the degree to which firms engaged in this type of AI solution development remain locally embedded varies, indicating that not all AI technologies exhibit the same level of ‘geographical stickiness’.

From this perspective, it is difficult to speak about “the geography of AI” as if AI were a single, uniform technology developed and used in similar ways across contexts. AI consists of different technologies and activities, each with its own location patterns and measurement challenges.

Which population of firms is being studied when we measure AI adoption?

When studying the geography of AI adoption and diffusion, it is essential to be explicit about which population is being analysed. In particular, a distinction needs to be made between AI-first companies and incumbent firms that are adopting AI within existing internal processes and products and services. This distinction became highly visible at GEOINNO 2026, where different papers relied on different populations, data sources, and units of analysis to study AI adoption and diffusion.

At the conference, Bram Timmermans presented our paper on the AI ecosystem of Norway using data from RankmyAI. For this paper, our unit of analysis is the company rather than the individual AI tool (which is the unit of analysis in the RankmyAI database). This choice enables linkage with Norwegian company registry data and allows the analysis of firm dynamics over time, including relocation across regions.

This approach differs from that presented by Nils Grashof in the paper “Knowledge Integration Patterns in AI: A Firm-level Taxonomy and Performance Implications”, which uses data from ISTARI.AI. That study analyses AI adoption and AI intensity across a broad population of companies in Germany, including both AI-first firms and incumbent companies that are integrating AI into existing internal processes and products and services. AI adoption is operationalised using company website content, resulting in an intensity measure rather than a binary classification.

Taken together, the comparison between the RankmyAI-based analysis of Norway and the ISTARI.AI-based study of Germany highlights a clear methodological distinction. RankmyAI focuses on AI-first companies and captures the supply side of AI-related activity, while the ISTARI.AI approach examines adoption and integration of AI across a broad population of firms. As a result, the two approaches address different research questions, operate at different units of analysis, and rely on different data sources. In our Norway paper, linking AI-first companies to AI-related patents and scientific publications provides insight into how AI-first activity relates to regional knowledge production and sectoral specialisation. The ISTARI.AI approach, by contrast, uses website-based AI intensity measures that are better suited to observing adoption and integration within incumbent firms. These differences do not indicate superiority of one approach over the other, but rather underline that different data sources capture different aspects of AI activity. This, in turn, shows that choices regarding population, data source, and unit of analysis directly shape empirical outcomes and conclusions about AI adoption and diffusion, making careful reflection on what is actually being measured essential when analysing the geography of AI technologies and uses.

Defining and measuring AI adoption

One of my main takeaways from the conference is how strongly empirical insights on AI adoption depend on the data sources and indicators that are used. While earlier sections discussed differences in AI activities and in the populations under study, this section turns explicitly to the empirical side of the question: how AI adoption and diffusion are observed in data. Across the papers presented, a wide range of indicators was used, each capturing a different aspect of AI-related activity and each coming with specific strengths and limitations.

Several papers at the conference explored alternative indicators of AI adoption that illustrate these differences. One paper I found particularly insightful was presented by Johannes Wachs, based on joint work by Daniotti et al., entitled “Who Is Using AI to Code? Global Diffusion and Impact of Generative AI”. The study analyses the global use of AI for coding and provides a detailed empirical perspective on a specific form of generative AI adoption. The paper has been recently published in Science and is freely accessible for those who wish to read it in full. The authors used GitHub commits to estimate the share of code generated by AI versus humans. This approach provides detailed insight into a specific form of AI use, but it also raises questions about representativeness. Not all developers use GitHub, and GitHub users are unlikely to represent the full population of developers. New practices such as vibe coding further complicate this, because the code is no longer directly observable: applications are generated by AI and evaluated at the level of the end product rather than through inspection of the underlying code. As a result, a growing share of AI-driven application development (e.g. websites and apps) cannot be identified by analysing code repositories.

Another paper, “Riding the AI Wave? The Role of Knowledge Spillovers for AI Innovations”, presented by Fulvio Castellacci, used AI-related trademarks as an indicator of AI activity. The number of firms with AI-related trademarks is small, but higher than the number of firms with AI-related patents. For me, this suggests that trademarks are a relevant additional data source to explore, but that they cover too few firms to provide a representative picture of AI adoption at the regional level on their own. Instead, trademark data appear most useful as a complementary source of information alongside broader data collection approaches such as those used by ISTARI.AI and RankmyAI.

Several other contributions focused on skills as a lens to study AI adoption and related ecosystems. One paper, “Skills and AI Innovative Ecosystems: Evidence from Brazilian Microregions”, presented by Cecilia Seri, constructed skill profiles for Brazilian microregions and related these to the presence of AI-first firms. The scope of AI activity analysed is broadly comparable to that used in RankmyAI, but the study relies on a single large commercial data provider. While this enables consistent measurement, it also raises questions about completeness and bias, for example due to a stronger focus on B2B-oriented firms and more limited coverage of generative AI–focused activity.

Two other papers examined skills in relation to a broader set of emerging digital technologies, including AI. “The Employment Impact of Emerging Digital Technologies” used patent data to identify around 40 emerging technologies and analysed their relationship with regional employment outcomes in Europe. “Advertised Technologies: Identifying Adoption of Emerging Technologies in Online Job Postings” focused on detecting emerging technologies through job vacancy data. For me, these papers raised questions about how emerging technologies are defined at the level of measurement. Some of the technologies included appear to be domain-specific applications of AI rather than technologies in their own right. In addition, the use of patent and job posting data raises broader issues related to keyword-based identification, potential biases in which firms patent or advertise, and the representativeness of the underlying firm populations captured by these data sources.

Another paper focused on AI adoption within existing firms and its regional economic impact. “How AI Adoption Shapes Regional Development through Export Performance”, presented by Virág Bittó, examined the relationship between AI adoption and firm export behaviour. The scope of AI adoption in this study is clearly defined and primarily relates to earlier and more established forms of AI, in particular the use of industrial robots. The study combines multiple data sources, including business survey with questions about adoption of AI technologies, firm-level characteristics and revenue statistics, and information on export activities. Within this clearly delineated scope, the results show that AI adoption is associated with a higher likelihood that firms engage in export activities. A key limitation of the analysis mentioned by the authors is that the exact timing of adoption at the firm level cannot be observed.

Another paper that stood out was presented by Andrea Conte, entitled “Effect of digital technologies and artificial intelligence funded investment on regional productivity and employment growth.” The study analysed the geographical distribution of AI-related funded investments across the European Union and examined their association with regional employment and productivity growth. By taking regions as the unit of analysis, the paper adopted a different perspective than the firm-level studies discussed earlier, highlighting how the economic effects of AI-related investment vary across regions.

A final paper worth highlighting is “Knowledge Integration Patterns in AI: A Firm-level Taxonomy and Performance Implications”, presented by Nils Grashof. Building on joint work with Dahlke and Schiller, the paper shows how AI intensity measures based on company website content can be used to compare AI adoption across regions. This approach captures how AI is communicated and positioned at the firm level and reflects a different stage of adoption than indicators based on patents or scientific publications. In that sense, it provides a novel and useful perspective that links directly to the earlier comparison between the RankmyAI approach and the ISTARI.AI methodology. At the same time, I had some reservations about the underlying unit of analysis. The analysis starts from company registry data at the level of establishments or offices, while AI intensity is determined using firm-level website content. This raises questions about whether observed AI adoption can be meaningfully attributed to specific offices or regions, particularly in cases where local establishments may not have adopted AI at all.

Taken together, the papers discussed in this section made clear to me that there is no single indicator that provides a comprehensive or representative picture of AI adoption and diffusion. Different data sources capture different processes, stages, and expressions of AI-related activity, ranging from knowledge creation and skills to adoption within firms and broader economic effects. As a result, empirical outcomes and regional patterns are highly sensitive to choices regarding data source, population, and unit of analysis. Being explicit about these choices, and about what a given indicator can and cannot capture, is therefore essential for meaningful interpretation and for drawing conclusions that are relevant for policy, firm strategy, and regional development.

Data collection and processing using AI

A recurring theme at GEOINNO 2026 was the increasing role of AI in data collection and data processing for research on AI adoption and diffusion. Several contributions showed how techniques such as natural language processing and machine learning are now routinely used to extract, structure, and analyse large volumes of unstructured data, including web content, job vacancies, patents, and other text-based sources. These approaches are becoming increasingly important given the speed and diversity of AI-related activity and the limitations of more traditional data sources.

The conference opened with a workshop on WebAI, an approach to collecting and analysing structured and unstructured web data using AI-based methods. The workshop was organised by Johannes Dahlke, Milad Abbasiharofteh, and Sebastian Smidt, together with ISTARI.AI, a data provider specialising in AI-based company data collection and regional ecosystem mapping. WebAI combines large-scale web data collection with machine learning and natural language processing techniques to identify technology adoption patterns based on online content. The methodological foundations of this approach are discussed in more detail in Dahlke et al. (2024), “Epidemic Effects in the Diffusion of Emerging Digital Technologies: Evidence from Artificial Intelligence Adoption”.

Beyond the workshop, many papers at the conference relied on similar techniques to process and interpret large text-based datasets. What became clear to me is that AI is increasingly part of the research infrastructure used to study AI adoption and diffusion. In some cases, AI is used to classify and interpret existing data, while in other cases it is also used to generate reference outputs that enable new forms of measurement. For example, in the paper on AI use for coding presented by Johannes Wachs, language models were used to generate code in order to enable systematic comparison with human-written code. This illustrates how AI-based methods expand the range of phenomena that can be observed empirically.

At the same time, the use of AI in data collection and processing does not remove familiar methodological challenges. Questions related to data coverage, transparency, reproducibility, and bias remain central. Which firms have a strong online presence, how technologies are described in text, and how models classify and interpret content all influence the resulting indicators. As such, AI-based approaches should be seen as complementary to, rather than substitutes for, other data sources such as surveys, administrative data, patents, or registries.

Overall, the discussions at GEOINNO 2026 reinforced for me that advances in AI-based data collection and processing significantly expand the empirical toolbox for studying the geography of AI adoption and diffusion. At the same time, they make careful choices about scope, unit of analysis, and interpretation even more important, as methodological decisions are increasingly embedded directly in data generation and processing pipelines.

Concluding remarks

Looking back at the discussions and papers presented at GEOINNO 2026, what stands out to me most is the methodological diversity in how AI adoption and diffusion are currently studied. There is no single way to observe or measure AI adoption, and the conference showed that different studies often focus on different phenomena while using similar terminology.

Across the contributions, AI appeared as a set of distinct activities and technologies with different spatial patterns, adopted by different populations of firms, and captured through a wide range of data sources. As a result, empirical findings on the geography of AI adoption and diffusion are highly sensitive to choices regarding definitions, populations, data sources, and units of analysis. These choices shape not only the observed patterns, but also the conclusions drawn about economic impact, regional specialisation, and diffusion dynamics.

Rather than viewing this diversity as a problem, I see it as a call for greater precision. Patents, job vacancies, website content, trademarks, skills data, surveys, and web-based indicators each capture specific aspects of AI-related activity, but none provides a complete picture on its own. Meaningful analysis therefore depends on being explicit about what is being measured, what is not, and how a given indicator relates to the underlying research or policy question.

This perspective also underpins the mission of RankmyAI. RankmyAI focuses on mapping AI-first activity in a transparent and systematic way, with clear choices about scope and unit of analysis. By tracking AI tools and companies globally and linking them to regional and sectoral contexts, RankmyAI aims to provide an independent and data-driven view on where AI activity emerges and how it diffuses. Our partnerships with Norway and Denmark, among others, illustrate how this approach can be applied in practice to study regional patterns of AI adoption and diffusion, while remaining clear about the limits of the data.

Finally, I would like to thank the conference organisers, and in particular Carlolina Castaldi, Fulvio Castellacci, and Johannes Dahlke, for organising the special sessions on the Geography of AI, which helped put this topic more firmly on the research agenda and provided a strong basis for future work in this area.