The paradigm shift: from keyword lists to intent clusters
Step-by-step keyword research is no longer about opening a tool, exporting hundreds of terms, and grouping them by lexical match. In 2025-2026, search engines and AI assistants prioritize intent and context, and that forces a rethink of the process. According to Google’s official documentation on quality updates, the focus is on usefulness and experience, not on repeating terms. This means that a rigid step-by-step approach can lead to outdated decisions if layers of semantic and behavioral analysis are not incorporated.
In this article we will look at the developments that are making a difference: the rise of intent clusters, the influence of generative summaries in the SERP, the measurement of real demand with Search Console data, and the practical limitations that few mention. This is not a universal guide; it is an update for those who already have a foundation and need to adapt their workflow.
Development 1: intent clusters with generative AI support
Until recently, grouping keywords was a manual or semi-automated process based on string similarity. Now, language models make it possible to group by underlying intent. For example, “how to change my Instagram password” and “recover Instagram account without email” may seem like different queries, but they share the intent of regaining access. Tools such as generative model APIs (with human oversight) can classify hundreds of queries into coherent clusters in minutes.
However, this does not work for every niche. In sectors with highly specific jargon or high ambiguity (for example, medical or legal terms), generative AI tends to over-group or invent relationships that do not exist. In my experience with a client in the insurance sector, an automatically generated cluster mixed “car insurance for young people” with “motorcycle insurance for young people” because they shared the demographic modifier; the result would have been a page with diluted content. The lesson: AI speeds things up, but human review is still mandatory to validate the real intent.
How to integrate this step into the classic workflow
The traditional step-by-step process (seed → expansion → grouping → prioritization) remains, but in the grouping phase it is advisable to use a double filter: first an automatic grouping by embeddings and then a manual review of the clusters with the highest volume. According to industry studies, teams that combine both approaches reduce grouping time by up to 40%, although there is no conclusive public data. The important thing is not to delegate the final decision.
Development 2: the impact of generative summaries on the SERP
Generative summaries (AI Overviews) are changing the way the success of a keyword is measured. An informational query may show a summary that answers directly, which reduces the click. This does not mean the keyword is dead, but rather that its value shifts toward brand and citability. In practice, when researching keywords for 2026, it is advisable to classify each term not only by intent (informational, transactional, etc.) but also by “citation potential”: is this a topic where the generative summary can cite sources? Topics with original data, comparisons, or first-hand experience are more likely to be cited.
A common mistake is to keep measuring only organic CTR in Search Console for informational keywords. If the summary appears, CTR drops and it seems that the keyword is bad. But if the page is cited in the summary, brand traffic may increase. In a technical content project, we found that a keyword with low CTR generated more assisted conversions than another with high CTR, precisely because it appeared in the generative summary. The recommendation is to cross-reference Search Console data with brand mention analysis and with the performance of landing pages.
Development 3: real demand data with Search Console and APIs
Traditional keyword tools offer volume estimates that may be biased. In 2025, Google Search Console allows you to export queries with impressions and clicks, and its API facilitates automation. The new development is to use this data as a primary source of demand for sites with established traffic. The step-by-step process is enriched: instead of starting only from third-party tools, you start from the real queries the site is already receiving, filter those that have impressions but low CTR, and prioritize them as content opportunities.
This does not work for new sites with few impressions. In those cases, research must rely on external sources and competitor analysis. In addition, Search Console has a row limit on export (1,000 rows in the interface, although the API allows more), and the data may be sampled. It is not a perfect source, but it is the closest to the site’s reality.
Development 4: conversational keywords and generative voice search
AI assistants and voice search are normalizing natural-language queries. This is not new, but in 2026 the nuance is that assistants not only respond, they also execute tasks. For example, “book me a table for two at an Italian place near here” combines transactional and local intent. Keyword research must include conversational variants, but without falling into the trap of creating a page for every long-tail phrase. The key is to identify the intent pattern and create content that solves the task, not the exact phrase.
In a local booking project, we saw that queries with “near me” and “open now” had a much higher conversion rate than generic ones, but they required structured data and integration with Google Business Profile. The limitation: if the business does not have updated data, conversational content is useless. It is an example of how keyword research must go hand in hand with technical and operational feasibility.
The updated step-by-step process (no magic formulas)
With the developments above, the step-by-step keyword research workflow looks like this:
- Define the business objective and the type of content: not all keywords deserve a page; some only need to be covered in a paragraph.
- Gather seeds from diverse sources: Search Console, competitors, forums, sales questions, and keyword tools. Include conversational variants.
- Expand with tools and APIs: use both estimated volume and real impression data.
- Group by intent with AI support and human review: validate that each cluster has semantic coherence and content feasibility.
- Classify by citation potential and SERP format: identify whether the generative summary may appear and whether the page can be cited.
- Prioritize according to effort, impact, and technical feasibility: discard those that cannot be covered with current resources.
- Measure and readjust: use Search Console to see which real queries are coming in and adjust the clusters.
This step-by-step process is not linear; it is iterative. In practice, prioritization often forces you to return to the grouping phase.
Limitations and mistakes to avoid
First, there is no tool that guarantees results. Any promise of “keywords that rank on their own” is false. Second, keyword research does not replace understanding the business: without knowing the margin, the sales cycle, or production capacity, you can prioritize terms that generate traffic but not revenue. Third, generative AI can hallucinate relationships between terms; you must always validate with real SERP data. Finally, the volatility of SERPs with generative summaries means that volume and CTR data change frequently; it is advisable to review the clusters every quarter, not every year.
In short, step-by-step keyword research in 2026 demands more judgment and less blind automation. Technological developments help, but the value still lies in human interpretation and in the ability to connect the keyword with content that truly resolves the intent.