Google Gets the Credit

A customer may discover a business through an AI assistant, research their options and then turn to Google to find the company and make a purchase. In that situation, Google could receive the credit for the visit, even though the assistant helped influence the buyer’s decision earlier in the process.

This raises an important question for businesses measuring the impact of AI search: how much influence happens before a customer reaches a website?

In an article examining the relationship between AI recommendations and traditional search, Pedro Dias argues that website referral figures alone cannot explain the full customer journey. While referral data can show where visitors arrive from, it does not necessarily reveal what influenced their decisions beforehand.

The Customer Journey Starts Before Google

Some buyers may now turn to AI assistants to understand a problem, compare suppliers or identify potential solutions before using a search engine.

When they eventually search for a company on Google, that visit might represent the final stage of a decision rather than the beginning of their research.

An AI assistant could introduce a business, recommend a particular product or influence a buyer’s opinion without generating a direct website visit. It could also discourage a potential customer from considering a company by presenting unfavourable or inaccurate information.

These possibilities highlight why businesses should look beyond referral traffic when assessing how AI influences purchasing decisions.

However, the existence of these journeys does not establish how frequently they occur or how much revenue they generate. Businesses still need evidence before assigning a financial value to influence that cannot be directly attributed.

Understanding the Difference Between SEO and GEO

Dias questions whether Generative Engine Optimisation (GEO) should be treated as an entirely separate discipline from Search Engine Optimisation (SEO). He argues that businesses should identify what GEO adds beyond established SEO practices rather than simply giving familiar activities a new name.

The distinction he focuses on is the difference between information a model has learned during training and information it retrieves while answering a question.

A language model’s learned associations are represented in its numerical parameters, often referred to as its weights. These associations can influence how the model responds to questions about particular brands, products or problems.

Retrieval works differently. When an AI system searches for external information to answer a query, it can use that material alongside its existing learned knowledge. Retrieval-augmented generation, or RAG, is one approach that combines these sources.

SEO remains important for helping relevant web pages become discoverable and accessible to search systems. However, a page being available does not guarantee that it will be retrieved or used in an AI-generated answer.

Google has also stated that established SEO practices remain relevant to its AI search features and that businesses do not need a separate set of special optimisation requirements.

Dias describes AI visibility as a combination of SEO and GEO, while acknowledging that instructions, user context and retrieved information can also shape an assistant’s response.

How Learned Associations Could Influence Search

An AI system’s existing knowledge may affect the searches it generates when investigating a topic.

Google’s query fan-out process, for example, can involve running related searches across different subtopics and sources to build a more comprehensive response.

Research titled What Drives AI Citations? compared brands recommended by an AI system without search against those included in its generated search queries. In the particular OpenAI API configuration tested, brands recommended without search appeared in generated queries in 67% of cases, compared with 11% for the comparison group.

These findings suggest that a model’s initial associations may influence which brands it investigates further. However, they do not establish whether changing a model’s training would necessarily change its searches. That question requires further testing.

The research also observed AI systems identifying less familiar suppliers through initial search results. This suggests that learned knowledge is not the only influence: retrieved information and user instructions can introduce new brands and shape subsequent searches.

For businesses, the important question is whether a brand’s associations within a model influence the searches used to investigate a buyer’s needs.

Testing How AI Models Learn

Research into language model training provides ways to investigate whether learned associations can influence future responses.

One paper discussed by Dias found that introducing specific information during controlled training could increase the likelihood of related keywords appearing in other contexts. The researchers described this effect as spillover priming and found that including the same material in a prompt produced a substantially smaller effect than training on it.

However, more mentions do not automatically mean better answers. If a keyword appears in an irrelevant response, the result could reduce the quality of the information provided.

Other research has examined whether the way facts are presented during training affects a model’s ability to answer factual questions. These studies suggest that repeating information and successfully using it in a new context are not necessarily the same thing.

Researchers can investigate these effects using open-weight models, which make their learned parameters available for examination. When training code and data are also accessible, they can compare model versions and test how different training materials affect their responses.

Nevertheless, findings from controlled experiments cannot automatically be applied to commercial AI systems operated by companies such as OpenAI, Anthropic or Google. Those systems may use different training methods, and their behaviour requires separate evidence.

Dias argues that businesses should assess how consistently an intervention changes responses across different questions and conditions rather than relying on a handful of favourable examples.

He also stresses that publishing content, having it crawled or receiving a citation does not prove that a model has learned useful information from it. Demonstrating an actual effect requires a suitable experiment.

When AI Answers Reduce Website Traffic

AI-generated answers may also create challenges for publishers whose business models depend on website visits.

A recipe website, for example, could attract search traffic by providing ingredients and instructions. If an AI assistant supplies that information directly, users may have less reason to visit the original page.

The publisher has still invested time and money in producing the content, but may lose the traffic that supports its business.

The situation can differ for service-based companies. An AI assistant might recommend a supplier, helping a customer choose a business even if the recommendation does not lead to an immediate website visit.

Dias argues that publishers need to consider what continuing value their information provides after an AI system has used it. Information that requires regular testing, verification or updating could give users and AI systems a reason to return to the original source.

He also anticipates that some businesses may move towards smaller websites supported by application programming interfaces (APIs) and Model Context Protocol (MCP) connections. These interfaces can allow AI assistants to access current information or use services without necessarily directing customers to individual web pages.

If these interactions become more common, website visits alone may provide an incomplete picture of how customers engage with a business.

However, remaining useful to an AI system does not automatically generate revenue. Publishers still need a sustainable business model that supports the cost of producing and maintaining their information.

Why AI Mentions Need Context

Simply counting how often a brand appears in AI-generated responses can be misleading.

A mention might represent a recommendation, a warning or a correction. Treating every appearance as a positive result could therefore give businesses an inaccurate impression of their visibility.

Dias argues that the more meaningful questions concern whether potential customers encounter a business, understand its offering and become more likely to choose it.

Research into customer journeys and self-reported discovery can help uncover influences that referral analytics miss. Where possible, controlled comparisons can also test whether an intervention produces a measurable improvement.

However, increases in branded searches or improvements in AI-generated answers do not, on their own, demonstrate additional revenue. Businesses need evidence linking these changes to meaningful commercial outcomes.

Measuring the Influence of AI Search

For companies assessing their AI visibility, the challenge is to understand what happens before a customer reaches their website.

An assistant may help a buyer identify a supplier, while Google provides the route to the company’s website. Standard analytics may attribute the resulting visit to Google without revealing the role AI played earlier in the decision.

This does not mean businesses should assume that hidden AI influence is substantial. Instead, they should investigate the customer journey and establish what the available evidence can actually demonstrate.

As AI assistants become more involved in research and purchasing decisions, businesses may need to broaden their measurement beyond referral traffic and brand mentions.

Google may receive the credit for the final visit, but understanding who introduced the business to the buyer could provide a more complete picture of how the sale happened.