Keyword Research in the AI Era: Do Keywords Still Matter?

This distinction matters.

In traditional SEO, we often treated the keyword as the target:
We want to rank for “FP&A software.”

In modern SEO/GEO, we would ask a bigger question:
What does a search engine or AI system need to understand about this company for it to consider the brand relevant when someone is looking for an FP&A solution?

Now things become much more interesting.

We need to understand the language of the category.

FP&A software.
Financial planning.
Budgeting.
Forecasting.
Scenario planning.
Financial reporting.
Excel.
Data consolidation.
CFO.
Finance team.

But we also need to understand the relationships between those concepts.

And that’s why keywords still matter.

A language model can understand semantic relationships between different expressions, but it still needs language from which to build that understanding.

At RockingGEO, we therefore don’t think about keywords merely as strings we need to repeat on a page.

We think of them as semantic signals that help establish what a brand, product, page and topic are about.

Then Come the Questions and Prompts

Once we understand the topic and its language, we move outward.

How might someone express this need?

A traditional Google query could be: best FP&A software.
Another person might search: FP&A software for mid-size companies.

Someone using ChatGPT could ask: “Which FP&A tools are best for mid-market finance teams?”
Someone else could say: “We’re outgrowing Excel for financial planning but don’t want our finance team to stop using Excel. What platforms should we consider?”
And another: “Compare FP&A platforms for a 500-person SaaS company.”

Different words.
Different prompts.
But potentially the same underlying intent universe.

This is why replacing a keyword spreadsheet with a prompt spreadsheet doesn’t solve the problem.
You could easily collect 5,000 prompts. Tomorrow, users could phrase the same needs in another 50,000 ways.

We don’t need to own every possible sentence. We need to own the topic and the intents behind those sentences.

Prompt research helps us understand how those intents are expressed.
It doesn’t replace the strategy.

Similarweb proposes starting prompt research with anchor queries and expanding them across areas such as definitions, comparisons, how-to questions, use cases, objections, entities and metrics.

That’s a useful way to think about prompts: not as individual keywords to optimize for, but as different expressions of an information need.

And Then AI Does Something Interesting: It Searches Again

This is where things get really fun.

You ask AI one question.
That doesn’t necessarily mean it conducts one search.

AI search systems can use a process commonly called query fan-out: breaking the original question into several related searches that help gather the information needed to construct an answer.

Imagine asking: “What’s the best FP&A software for a mid-market SaaS company that wants to keep Excel?”

The system might investigate variations around:
best FP&A software
FP&A software for SaaS
FP&A software for mid-market companies
Excel-native FP&A software
FP&A software Excel integration
FP&A platform comparison

These aren’t necessarily queries the user typed. The machine generated them.

Search Engine Land describes query fan-out as AI expanding a user’s question into related searches to better understand its context and satisfy likely follow-up needs.

Ahrefs describes the change particularly well: traditional search was largely a many-queries-to-one-result environment, while AI can turn one query into many searches.

And this has a pretty big implication for keyword research.

Welcome to Machine Search Volume

Traditional keyword tools attempt to tell us how often humans search for something.
But what happens when machines start conducting searches on behalf of humans?

Imagine a query has virtually no measurable traditional search volume.
Would you ignore it?
In the old model, probably.

But what if that query repeatedly appears as a fan-out search generated when thousands of users ask AI systems questions about your category?

Suddenly, zero human search volume doesn’t necessarily mean zero search opportunity.
We now need to think about two types of search behavior:

Human-generated queries and Machine-generated queries.

Traditional keyword tools are largely built to understand the first.
Modern GEO research increasingly needs to understand the second as well.

Fan-Out Research Changes How We Think About Content

There’s already interesting evidence here.

Surfer analyzed more than 173,000 URLs across approximately 33,000 fan-out queries. 
They found a 0.77 correlation between ranking across fan-out queries and being cited in AI Overviews.
Even more interestingly, pages ranking for both the original query and its fan-out queries were 161% more likely to be cited than pages ranking only for the original query.

This doesn’t mean:
Find every fan-out query and create a page for each one. Please don’t. It suggests something much more useful: Build content that genuinely covers the territory surrounding an intent.

A great page shouldn’t simply answer one keyword.
It should anticipate the questions, requirements, comparisons, objections, entities and related concepts someone needs to understand before their need is actually satisfied.

That’s topical coverage. And AI search makes it more important than ever.

Ranking Still Isn’t the Same as Being Cited

Here’s another place where traditional SEO logic starts to break.

If an AI system uses search, it’s tempting to assume:
Rank higher → get cited by AI.

It’s not that simple.

Ahrefs found only 6.82% overlap between ChatGPT citations for fan-out queries and Google’s top 10 results in one study.

In another study of 15,000 prompts, Ahrefs found that only around 12% of URLs cited by AI assistants appeared in Google’s top 10 for the original prompt.

Perplexity showed considerably more overlap with Google’s results than some of the other systems studied.

So ranking still matters.
Retrievability matters.
Topical relevance matters.
But none of them individually guarantees citation.

Which brings us to another major change.

There Is No Single AI Search Engine (LLM)

For 20+ years of SEO, Google gave us a fairly obvious center of gravity. GEO doesn’t have one.

ChatGPT isn’t Gemini.
Gemini isn’t Claude.
Claude isn’t Perplexity.

And their retrieval and citation behavior can differ.

Search Engine Land compared how ChatGPT, Gemini, Claude, Perplexity and DeepSeek find, synthesize and cite information, highlighting important differences between the engines.

It gets even messier.
Semrush found only 25.6% overlap in cited domains between different ChatGPT reasoning modes when given the same prompts.

Think about what that means for the traditional concept of “rank tracking.”
A keyword could have one relatively stable Google position.
An AI prompt doesn’t necessarily have one equivalent “position”:

Change the wording, the context, the model, the reasoning behavior.
Ask again tomorrow – you may get a different answer.

That’s why we don’t believe GEO research should simply take traditional rank tracking and replace keywords with prompts.

Don’t Build a Prompt List. Build an Intent Map.

This is perhaps the most important practical takeaway.

Instead of 500 keywords or its shiny new GEO equivalent – 500 prompts, we’d rather understand:

Which audiences matter?
What do they need?
What are their core intents?
Which topics satisfy those intents?
Which words define those topics?
Which questions and queries surround them?
How are those needs expressed as prompts?
Which queries might AI fan out into?
Which sources and entities currently shape the answers?
Where should our brand be present?

That’s not keyword research in the traditional sense.
It’s an Intent & Language Map.
And keywords are one of its fundamental building blocks.

Search Volume Isn’t Dead Either

We still want search volume.
We still look at ranking difficulty.
We still analyze Google.
We still look at competitors.
We still care about the SERP.

Why wouldn’t we?

All of that is useful evidence about demand, language and competition.
What changes is the weight we give those signals.

Imagine two topics:

Topic A: 10,000 searches/month, weak relevance to the company’s product.
Topic B: 500 searches/month, directly connected to a problem the company’s ideal customer needs to solve before buying.

Old-school SEO can easily lead us toward A.

Our methodology would probably make us much more interested in B.
And if B also appears repeatedly across relevant prompts, questions, comparisons and fan-out queries?

Now we’re listening.

So How Do We Do Keyword Research in the AI Era?

Our old keyword research methodology followed five steps:

Extract → Search → Delve → Sort → Prioritize

We’re keeping the idea. But we’re remixing the process.

1. Understand

Before opening Ahrefs, Semrush or any other tool, understand the human.

Who are we trying to reach?
What do they need?
What problems are they solving?
What decisions are they making?
What would make them search, ask or talk about our category?
And which of those conversations should our brand legitimately belong to?

Start with people, not volume.

2. Extract

Now we start collecting words.

Core category terms.
Products.
Problems.
Solutions.
Features.
Use cases.
Industries.
Roles.
Competitors.
Alternatives.
Entities.

This is where classic keyword research remains extremely useful.
We’re discovering the language of the topic.

3. Expand

Now take those terms beyond the keyword.

Explore:
Keywords → queries → questions → comparisons → objections → prompts → entities → fan-out queries.

Look at Google.
Look at competitors.
Look at Reddit, LinkedIn, YouTube and industry communities.
Look at what customers say.
And ask the different AI engines.

The goal here is to understand the semantic territory surrounding the audience’s needs.

4. Validate

Now bring the data in.
Search volume.
Ranking difficulty.
SERPs.
Competitor visibility.
AI answers.
AI citations.
Sources.
Fan-out queries.
Brand mentions.
Entity associations.
And differences between LLMs.

Ask the same underlying question in different ways and across different AI engines.

What stays consistent?
What changes?
Which brands keep appearing?
Which sources influence the answers?
What language surrounds them?

That’s where traditional SEO research starts becoming GEO research.

5. Prioritize

Finally, don’t simply sort your spreadsheet by volume.

Prioritize according to something closer to:
Business relevance × User intent × Topical importance × Search demand × Retrieval opportunity × Authority opportunity

The best opportunity isn’t necessarily the biggest keyword.
It’s the territory where your audience’s need, your business value and your ability to become an authoritative source intersect.

One More Thing: Being Relevant Isn’t Enough

Traditional SEO research often ended with: We found the keyword. Let’s create the page.

Today there’s another question:
Will an AI system actually want to use this content?

Semrush’s research into AI citations found positive associations with characteristics including clear summaries, E-E-A-T signals, Q&A formatting, strong section structure and structured-data elements.

So modern research needs to lead not only to content that can rank, but content that can be:
understood → retrieved → extracted → trusted → cited.

That’s a different standard.

🎵 Our Last Note: So, Are Keywords Dead?

No. keywords are not dead. Keyword-first SEO is.

People don’t wake up in the morning wanting to search for keywords. 
They have problems, questions, needs, fears, requirements, decisions to make.

AI systems are getting much better at understanding those intentions – and at breaking them into additional searches in order to find the information required to answer them.

But every layer of that process still depends on language.

Prompts are made of words.
Queries are made of words.
Fan-out queries are made of words.
Web pages are made of words.
Categories are defined by words.
Relationships between entities are communicated through words.
And brands become associated with topics partly through the language consistently used around them.

That’s why at RockingGEO, we don’t start keyword research by asking:
Which keywords should we rank for?

We start with:
Which topics and user intents should this brand own?
Then we find the keywords, questions, queries, prompts, entities and fan-outs that define that territory.

Because the keyword isn’t the destination anymore.

It’s one of the building blocks that helps humans, search engines and AI understand how everything connects.

And words still matter.