Understanding intent - what a user is trying to accomplish - is critical for SEO and digital marketing. However, identifying intent across hundreds or even thousands of keywords is a complex challenge, and it’s only getting harder whilst becoming MORE critical.
Recently, Semrush analysed 80 million lines of clickstream data from AI chatbot interactions, attempting to classify intent based on user prompts. The result? A staggering 70% of classified intent was labeled as "unknown." In other words, their model couldn't determine what the user was trying to do.
Why Is Intent Classification So Difficult?
There are several potential reasons behind this high rate of "unknown" classifications. The first and most obvious explanation is that Semrush’s intent model was built from Google search data. This means it was trained on structured queries that follow search engine behavior, not the more conversational, open-ended nature of chatbot interactions.
Since chatbot prompts often deviate from traditional search queries, they introduce new types of intent that the model simply hasn't encountered before.
To their credit, Semrush has been transparent about this gap rather than attempting to conceal it. They are likely not alone in facing this challenge-classifying chatbot prompts is an emerging issue that many tool providers are grappling with right now.
The Flaws in Traditional Intent Models
Another fundamental problem is that our existing intent models are not entirely fit for purpose. I have spoken on these in the past and appreciate they have some use, but others may even debate they have always been heavily flawed or even broken.
The traditional framework for search intent, Semrush provides a good example - categorising queries into Informational, Commercial, Navigational, and Transactional - has always been a useful oversimplification. However, as search behaviour evolves, this model continues to be challenged.
The idea that users move logically from one phase to another (e.g., Information > Commercial > Transactional) in a predictable funnel has long been a convenient fiction. While this structure provided a useful guideline, it has never perfectly reflected real-world search behaviour.
Users don’t follow a neat, step-by-step path. Their journey is influenced by context, emotions, and external triggers that don’t fit neatly into a four-stage model. Models like this not need to be perfect, they just needs to provide enough structure to give useful insights and ways of working with data quickly. But 7/10 queries unable to be classified (as per the above example) strains it beyond usefulness.
The Impact of Conversational Search
One fascinating insight from Semrush’s study was the length of chatbot prompts. Users tend to write much longer, more detailed queries when interacting with chatbots compared to traditional search. This suggests a significant shift in how people articulate their needs when not confronted with a search engine. Anyone who has worked with/on Google for 10 or more years will have been "trained" in basic keyword search. Searchers learned how to query Google in a way that got better results, but was also different to how you would otherwise have done it. Compare how a 15 year old might search google today vs 10years ago, it would be totally different. Maybe LLM-based search - with it's "Chat bot" feel - just presents a less constrained search experience. "Conversational search" isn’t new - Google has been exploring it for years - but chatbots bring it to the forefront in a way that tests our existing mental models. Every individual has unique ways of learning, searching, and decision-making, which means that attempts to rigidly classify their intent will always be strained. Semrush also suggest that the sheer diversity of what a service like ChatGPT can do (vs "conventional" search results) is another huge complicating factor. Like going from 2D chess to 3D chess - there's another significant learning curve there!
A More Dynamic Approach to Intent Analysis
While our intent models might be struggling, there are still valuable tools that can help deepen our understanding. One particularly insightful tool is AlsoAsked.
AlsoAsked scrapes the "People Also Asked" (PAA) boxes in Google results, revealing the types of related questions users ask about a given topic. But the real power of AlsoAsked lies in its methodology. Instead of just collecting a list of related questions, it "clicks" through the answers, mimicking a user’s journey by triggering the next set of related questions.
The PAA boxes in Google are dynamic - new questions appear based on what has been clicked. By following different branches of questions, AlsoAsked helps simulate potential user journeys. This allows us to visualize how users might navigate through various layers of inquiry, uncovering different moments that could be pivotal in their decision-making process.
If we want to think about mapping our content (and more importantly services) against the most common user journeys, the first major hurdle everyone faces is knowing what that journey is. PPAs aren't THE answer for everyone, but they give a significant window into the minds of searchers even if they're not powered by LLMs (yet).
What This Means for Search Intent Strategies
Does this fix our search intent models in a neat, convenient way? No. But does it give us valuable insights into the way users seek information and make decisions? Absolutely.
Instead of trying to force users into outdated intent categories, we need to embrace a more fluid, exploratory approach to search behaviour.
Understanding the different knowledge gaps users experience along their journey, and identifying key moments of interest, will be far more valuable than attempting to fit them into rigid classifications.
Weening Off a Linear Four-Stage Model
If you’re still relying on the traditional linear four-stage intent model, it’s time to rethink your approach. Interest in deeper, more nuanced understanding of user behaviour will give you a competitive edge in search marketing. There is a ballooning space for tools and techniques in this area, and those looking to get ahead now may have a true advantage.
Kommentare