Google Gets the Thank You

We often rely on referral reports to tell us where our customers come from. But there is a growing gap between who sends the visit and who actually shaped the decision to visit in the first place. If a user asks an AI assistant for a recommendation, finds your business, and then goes to Google to search for your brand name, Google gets the credit in your analytics. The AI assistant, which did the heavy lifting of the introduction, remains invisible.

This creates a dangerous blind spot. If we only measure the final click, we ignore the influence happening upstream. We risk dismissing the impact of AI discovery simply because the attribution doesn't show up as a direct referral. This connects with We Earned 1 when the same signal needs a clearer operating decision.

The Invisible Journey from Question to Visit

I have noticed a pattern in how people are interacting with the web: they ask an assistant first and use Search second. While this is an observation based on the people I encounter rather than a global data set, it raises a critical question about what a branded search actually represents. A useful companion note is It Comes from Other People, because it looks at a nearby part of the same system. The same pattern also shows up in Not the Outputs), where the practical question is how the signal becomes visible.

A visit to your site via Google isn't always the start of an investigation. It could be the final step of a shortlist created during a conversation with an AI. A buyer might use an assistant to understand a complex problem before they even know which product category they need. In that interaction, the AI might introduce your company or, conversely, describe a competitor so poorly that the buyer rules them out. In both scenarios, the AI influenced the outcome without ever triggering a referral link.

The risk here is treating attribution gaps as evidence that AI isn't driving value. Just because we can't see the revenue in a GA4 report doesn't mean the influence isn't there. However, we must also be careful not to invent "invisible value" to justify spending without evidence. The goal is to investigate the gap, not to fill it with assumptions.

Expert Interpretation: The tradeoff here is between precision and visibility. Referral data is precise but narrow. If you only optimize for the "click," you are optimizing for the end of the journey. The decision you need to inspect is whether your current KPIs are measuring the arrival or the influence. If you only track arrivals, you are blind to the "dark funnel" of AI recommendations.

Parametric Memory vs. Retrieval

There is a lot of noise around new acronyms like GEO (Generative Engine Optimization). I am generally skeptical of new terms for work that SEO already covers. However, the distinction that makes GEO useful is the concept of parametric memory.

Parametric memory is the knowledge and associations a model has learned during its training phase, baked into its numerical weights. This is fundamentally different from retrieval. In Retrieval Augmented Generation (RAG), the model looks up external information in real time to answer a prompt. It is like a colleague who is naturally smart (parametric memory) but also has a reference folder they can open to check a specific fact (retrieval).

SEO traditionally handles the "reference folder" part: making sure your content is discoverable, accessible, and structured so that a system can retrieve it. GEO, in the sense I use it, is about influencing what the model learns about a brand during training. It is about the associations the model holds before it even begins to search the web.

Expert Interpretation: This is the most critical technical distinction in modern search. Retrieval is a tactical game of content structure and indexing. Parametric memory is a strategic game of brand authority and presence across the training corpus. The tradeoff is control: you can update a webpage today (retrieval), but you cannot easily update a model's weights (parametric memory). You must decide if you are fighting for the "folder" or the "brain."

How Memory Directs the Search

The associations a model has learned don't just affect the answer it gives; they affect the searches it performs. Google uses "query fan out" to issue related searches across various subtopics. If a model already has a strong association between a brand and a specific problem, that brand is more likely to appear in the generated queries used to refine an answer.

Research into AI citations suggests a stark difference here. In certain OpenAI API configurations, brands recommended without a search were found in subsequent queries in 67% of cases, compared to only 11% for brands that weren't initially recommended. This suggests that parametric memory gives a brand a head start. If the model already "knows" you are a solution to a problem, it will use its search capabilities to verify or expand on that knowledge, rather than discovering you from scratch.

This means that the "invisible" work of being present in training data directly impacts the "visible" work of appearing in AI generated search results.

Expert Interpretation: This reveals a compounding effect. Brand authority in the training set leads to preferential treatment in the retrieval phase. The decision for a business is to stop viewing "brand awareness" as a vague marketing goal and start viewing it as a technical requirement for AI visibility. If the model doesn't "know" you, the retrieval process has to work much harder to find you.

Studying the Learning Process

We can actually observe how training changes these associations. Research on "spillover priming" shows that when specific texts are inserted into controlled training runs, the probability of a keyword appearing in other relevant contexts increases. Interestingly, simply putting that text in a prompt (the "folder" approach) produced far less of this effect than actually training the model on it.

However, more mentions do not always equal better results. If a keyword appears in an irrelevant answer, it can actually degrade the quality of the response. there is a difference between "recitation" and "understanding." A model might be able to repeat a passage from its training data perfectly, but that doesn't mean it can apply those facts to a new, different question.

For those with the resources, open weight models allow for controlled experiments. By using resources like Pythia, which provides models at different stages of training, researchers can test how changing training material alters the model's behavior. This moves the conversation from guesswork to empirical testing.

Expert Interpretation: The danger here is the "mention trap." Many tools track how often a brand is mentioned by an AI, but they don't track the sentiment or context of that mention. A mention in a warning or a correction is a loss, not a win. You should inspect your "AI visibility" metrics to ensure they distinguish between a recommendation and a mere citation.

The Economic Risk of "Zero Click" Content

The shift toward AI answers creates a fundamental economic problem for certain types of content. Consider a recipe site. If a user asks an AI for a recipe and the AI provides the ingredients and instructions directly in the chat, the user has everything they need. They have no reason to visit the original website.

For a service based business, this is less risky because an AI recommendation still requires the user to hire a human. But for publishers whose content is the product, the AI is essentially stripping the value from the source. The publisher still pays the cost of production, but the AI captures the utility for the user.

The central challenge for these creators is finding a reason for the user to depend on the original source after the AI has already provided the answer. It is no longer enough to provide a correct answer; the source must provide something the AI cannot synthesize, such as unique authority, community, or a proprietary tool.

Expert Interpretation: This is a pivot from "information provision" to "value provision." The tradeoff is that high quality, easily synthesizable content is now a liability if it doesn't lead to a deeper relationship. The decision here is to evaluate your content: is it a "commodity answer" that an AI can steal, or is it part of a larger ecosystem that requires a visit to the site to fully experience?

Measuring Meaning Over Mentions

Ultimately, for any business seeking qualified demand, a mention in an AI response is a vanity metric unless it leads to a choice. We need to distinguish between different types of AI interactions:

• Named Queries: When a user asks about a specific company, the AI simply confirms the company exists.
• Problem Queries: When a user asks for a solution to a problem, and the AI identifies the company as a provider.

The latter is where the real value lies. However, even then, a mention could be a recommendation, a warning, or a correction. A dashboard that marks every mention as "green" is misleading. A warning about your business is not a win.

To truly understand the journey, we have to look beyond referral reports. Buyer research and self reported discovery are the only ways to uncover the paths that GA4 misses. While a rise in branded searches is a positive signal, it is not a complete picture. The goal is to ensure that when a buyer encounters your business through an AI, they understand the value proposition and are more likely to choose you.

Expert Interpretation: The final takeaway is that the "thank you" goes to Google, but the "influence" belongs to the ecosystem. To manage this, you must move away from purely quantitative metrics (traffic, clicks) and toward qualitative signals (customer surveys, "how did you hear about us" fields). The decision is to stop trusting the dashboard as the sole source of truth for your acquisition strategy.