Keyword Research for SEO (A Practical Process, Not Just Tools)

Most keyword research advice starts with a tool and ends with a spreadsheet of monthly search volumes that doesn’t actually tell you what to write. Keyword research is the process of figuring out what your actual audience searches for, in what order, and with what intent – not just pulling a volume number and calling it done.

This post covers the process I actually use: how to anticipate real search phrasing, sort keywords by search stage instead of just volume, check what Google is already ranking before trusting any tool’s data, and map each keyword to a real page instead of a spreadsheet row. I’ll keep tool name-drops light here on purpose, since which specific tool you use matters far less than the process behind it.

I’ve run this process across iGaming, SaaS, and Fintech sites with wildly different audiences, and the failure mode is always the same: someone pulls a keyword list sorted by volume, writes content around the top ten, and wonders six months later why none of it converts.

What Keyword Research Actually Means (Beyond a Spreadsheet of Volumes)

Keyword research isn’t the list itself. It’s the process of understanding what a real person types into a search box, why they typed it, and what they need to see to consider their search answered. The list of terms and volumes is just the output of that process, not the process.

A keyword list without intent attached to each term is close to useless for content planning – you can’t decide what to write, what format to use, or what to link to next without knowing whether someone searching that phrase wants to learn, compare, or buy.

Google’s own guidance backs this up more directly than most SEOs give it credit for. Its SEO Starter Guide points out that different users search the exact same intent in completely different words – its example is someone searching “charcuterie” while another person searches “cheese board” for the same underlying need. Google’s language systems are built to understand that relationship even when your page never uses one of the two exact phrases.

That single point changes how I approach the early stage of research: I’m not hunting for one exact phrase to repeat, I’m mapping the full range of ways real people ask the same underlying question.

A seed keyword – the single broad term you start from, like “keyword research” itself – is only useful as a starting point for that mapping exercise. On its own it tells you almost nothing about what to build. The moment I have a seed term, the next question is always what topic cluster sits underneath it, not what single page could rank for it.

Step 1: Anticipate How People Actually Search, Not How You’d Phrase It

The biggest research mistake I see is starting from the phrasing an expert would use, not the phrasing an actual searcher uses. Someone new to a topic searches in plain, sometimes clumsy language. Someone who already knows the space searches in shorthand and jargon.

Both groups might land on the same page, so the research has to capture both. I start by writing down every version of the question I can think of, then check what real people actually typed using the “People also ask” box, related searches at the bottom of the results page, and autocomplete suggestions typed directly into Google.

Autocomplete and “People also ask” data reflect actual aggregate search behavior, which makes them a better starting point than a tool’s suggested-keywords feature in my experience, because a tool’s suggestions are generated from its own database of tracked queries, while the SERP itself shows you what Google has already decided real people are asking around that topic right now.

This step alone usually produces ten to twenty phrasings I wouldn’t have guessed sitting at a desk without checking. It also surfaces genuinely different intents hiding inside what looked like one topic – a pattern the next step deals with directly.

Cluster First, Sort Second

Before sorting anything by stage or volume, I group the raw phrase list into clusters that share the same underlying question, not just overlapping words. Ten phrasings might genuinely be one topic asked ten ways, or they might split into two or three real subtopics that each deserve their own page.

A cluster is defined by shared intent, not shared vocabulary – two phrases with almost no words in common can belong to the same cluster if a searcher typing either one would be satisfied by the same page, while two phrases that look nearly identical on the surface can belong to different clusters if the SERP for each one is answering a different question.

I do this manually for anything under a hundred terms, reading each phrase and asking what page it belongs next to. Past that volume, a spreadsheet with a manual “cluster” column still beats an automated clustering tool in my experience, because automated clustering groups by keyword similarity, not by the searcher’s actual underlying need, and the two only sometimes line up.

Step 2: Sort by Search Stage, Not Just Volume

This is where most keyword research stops short, and it’s the part I’d call genuinely advanced rather than beginner-level. A Search Engine Land keyword research guide lays out a three-tier structure worth adopting directly: head terms used for initial exploration and brand evaluation, body or medium-tail keywords used for broader subject exploration when someone wants more information before deciding, and long-tail searches used for deeper exploration when previous searches haven’t satisfied an unmet need.

Sorting keywords by which of these three stages they belong to matters more than sorting by volume alone, because it tells you what format and depth of content actually answers the search, not just how many people are typing it.

The same guide frames the space between initial awareness and a final decision as the “messy middle” of search behavior – the exploration and evaluation phase where most B2B and considered-purchase searches actually happen. I build two working lists from this framing on every project now: exploration keywords, for the early browsing stage, and evaluation keywords, for the comparison stage right before a decision gets made.

Where this gets genuinely useful is mapping keyword intent to something more specific than “informational vs. transactional.” The same Search Engine Land methodology borrows a six-need framework originally from Kantar’s consumer research – searches driven by wanting to be educated, reassured, helped, impressed, surprised, or thrilled by what they find. A “how does X work” search is asking to be educated. A “X vs Y” search is asking to be reassured about a decision already half-made.

Matching a keyword to one of these underlying needs, not just an intent label, changes what you actually write – an educate-me search wants mechanism and explanation, a reassure-me search wants a direct comparison and a clear recommendation, not more background.

Step 3: Check the SERP Itself Before Trusting the Data

Search volume numbers are estimates, and every tool estimates slightly differently depending on its own data sources. I never trust a volume number until I’ve actually looked at what’s ranking for that term, because the SERP tells you things no volume column can.

If the top ten results are all product pages, that’s a transactional keyword no matter what a tool’s intent label says. If they’re all long-form guides, that’s informational. If the results are a mix of forum threads and Reddit posts, that’s usually a sign the searcher wants a real, opinionated answer, not a polished corporate page.

This step also catches a mistake specific to organic keyword research: confusing a term’s paid search value with its organic opportunity. A keyword can carry heavy commercial intent and real paid search demand while being nearly impossible to rank organically because every top result is a paid listing or a dominant brand.

I’ve had this conversation more than once with a client who saw a high-volume term in a paid search report and assumed it belonged on the organic content calendar too – the actual split between what organic content and paid campaigns are each built to do matters here more than the shared keyword does.

Checking the SERP also tells you what’s missing. If every result answers the surface-level question but none address a follow-up a real searcher would obviously have next, that gap is often the actual opportunity, not the head term itself.

I’ve seen the same term come back with a 10x volume difference between two well-known tools, for no reason more sophisticated than each tool sampling a different clickstream panel. Neither number is “correct” in an absolute sense – treat any single tool’s volume figure as a rough order-of-magnitude signal, not a precise count, and use the SERP itself as the tie-breaker whenever two terms look close enough that the ranking between them actually matters for prioritization.

Common Mistakes That Waste a Research Cycle

A few patterns show up often enough across client accounts that I check for them specifically now, before a keyword list gets signed off:

Confusing a competitor’s ranking keyword with your own opportunity is the most expensive one – a term a competitor ranks for because of a decade of domain authority isn’t automatically a realistic target just because a gap-analysis tool flagged it. I still pull competitor keyword gaps, but every flagged term gets the same SERP check as anything else before it earns a place on the list.

The second is treating a keyword list as finished once it’s sorted and prioritized, without ever coming back to check whether the pages built from it actually rank for what they were built for. Search behavior shifts, and a term’s dominant intent six months after publication isn’t guaranteed to match the intent it had when the research was done.

The third is dismissing a long-tail term purely because a tool shows it at zero or near-zero volume. Plenty of genuinely valuable long-tail searches are too specific or too new for any tool’s sampled data to have caught up with yet – if the SERP for that exact phrase is thin or outdated while the phrase itself clearly reflects a real, specific need, that’s often a faster win than competing for a well-documented head term.

Step 4: Map Every Keyword to a Real Page, Not a Spreadsheet Row

A keyword with nowhere to live is a wasted research hour. Before I finalize any list, every keyword gets mapped to one of three outcomes: an existing page it should be added to, a new page it justifies on its own, or a deliberate decision to skip it because it doesn’t fit the site’s actual scope.

Skipping this mapping step is the single most common reason keyword research turns into a shelved document nobody acts on – the list gets built, gets admired once, and then nothing happens because nobody decided which page owns which term.

Mapping keywords to existing pages also catches cannibalization before it happens: two pages quietly targeting the same term, splitting authority and confusing Google about which one should rank. When a term maps cleanly onto a page I already have, I check how that keyword should actually appear in the page’s title, headings, and body rather than just adding it to a list and hoping it gets used naturally.

When a term doesn’t map cleanly anywhere, that’s usually the strongest signal I have that a genuinely new page is justified – not because the volume looked good in isolation, but because nothing on the site currently serves that specific need.

This mapping pass is also where I run a quick content audit against the site’s existing URL list, not just its sitemap – a page that technically covers a topic but was written three years ago under a completely different framing usually needs a rewrite, not a brand-new page competing against it. I’d rather find that overlap during mapping than after two pages have been quietly cannibalizing each other’s rankings for months.

Keyword Research Across Languages and Markets

Multi-market keyword research is where this process gets harder than most guides admit, and it’s a part of my day-to-day work across iGaming and Fintech clients running in more than one country.

A direct translation of a keyword list almost never matches actual local search behavior – search volume, phrasing, and even which intent dominates a term can shift completely between markets, even between two markets that share a language. A UK searcher and a US searcher typing what looks like the same phrase can be looking for meaningfully different things, and a term that’s purely informational in one Spanish-speaking market can carry real commercial intent in another.

In my experience, the fix isn’t a better translation tool – it’s rebuilding the “People also ask” and autocomplete check from Step 1 separately for each target market’s own search engine results, in the local language, rather than translating a single English-market list and assuming the intent carries over unchanged. A term that looks like a safe informational match in English can land as a completely different search stage once you check what’s actually ranking for it locally.

I’ve watched this play out directly on an iGaming account running the same core topic across three European markets: the English-language term sat firmly at the informational, exploration stage, the direct translation in one other market skewed almost entirely transactional because local regulation pushed most searchers straight to comparison and sign-up behavior, and the third market split the same topic into two genuinely separate clusters that had no equivalent split in English at all. None of that was visible from the translated list alone.

This is also where hreflang and content structure decisions get made, not just keyword lists – if two markets need genuinely different keyword sets for what looks like the same page, that’s usually a sign they need separate, properly hreflang-tagged pages rather than one page trying to serve both.

Keyword Research for Answer Engines, Not Just Search Results

Keyword research used to end at the traditional search results page. It doesn’t anymore. A growing share of the same queries now get answered directly inside an AI Overview or a chat interface, before a searcher ever reaches a list of blue links.

Question-shaped, evaluation-stage keywords are exactly the ones most likely to trigger an AI-generated answer instead of a standard results page, which changes what “ranking” for that keyword even means. I’ve covered how these answer engines actually decide which content gets featured in that generated response in more depth elsewhere.

The practical keyword-research implication is straightforward: when a term maps to an educate-me or reassure-me need from Step 2, structure the page to answer it directly and concisely near the top, not just optimize it for a ranked list position.

I now flag every question-phrased keyword during mapping specifically for this, since the content format that wins a featured snippet or an AI Overview citation is often different from the format that wins a standard organic ranking for the same underlying need.

Where Tools Actually Fit In

I use keyword tools constantly, but only after the process above, not instead of it. A tool is good at giving you a volume estimate and a list of related terms fast. It’s bad at telling you which of those terms actually deserves a new page, and it can’t check the SERP’s real intent for you.

Treat tool output as a starting list to filter through the process above, never as the finished research – a raw export sorted by volume skips every step that actually decides what gets written.

Free options (Google Search Console’s own query data, autocomplete, “People also ask”) cover Steps 1 through 3 well enough for most sites under a few hundred pages. A paid tool earns its cost mainly at scale – competitor gap analysis across thousands of terms, or tracking movement across a large keyword set over time – not by replacing the judgment calls in Steps 2 and 4.

Keyword Types at a Glance

The stage a keyword belongs to should decide the format you build, not the other way around – here’s the mapping I actually use when I sit down to plan content from a finished list.

Search stage	Typical intent	Example query pattern	Best content format
Head term	Initial exploration, brand evaluation	“keyword research”	Pillar page or category page
Body / medium-tail	Broader subject exploration, informational	“how to do keyword research for SEO”	In-depth guide or explainer post
Long-tail	Precise, unmet need after prior searches failed	“keyword research process for multi-market SaaS content”	Narrow how-to or FAQ-style answer

A Practical Keyword Research Checklist

The version of this process I actually run, in order, on a real project:

• Write down every phrasing you can think of for the core topic, then check autocomplete and “People also ask” for phrasings you missed
• Sort every term by search stage – head, body, or long-tail – not just by volume
• Attach an underlying need (educate, reassure, help, and so on) to each term, not just an intent label
• Check the actual SERP for each shortlisted term before trusting a tool’s volume or intent data
• Map every surviving keyword to an existing page, a new page, or a deliberate skip
• Flag question-phrased terms separately for direct-answer formatting
• For any second market or language, rebuild the SERP check locally rather than translating the list

The Part Worth Getting Right

If I had to cut this whole process down to one habit worth keeping, it’s Step 4: never let a keyword survive onto a final list without deciding which page owns it. Everything else here – intent sorting, SERP checks, the multi-market rebuild – exists to make that mapping decision correctly, not to produce a bigger spreadsheet.

A shorter list of keywords that each have a clear home on the site will outperform a longer list that never got past the research document, every time I’ve tested the two approaches against each other. Keyword research earns its place in an SEO process only when it ends in a page getting built or changed – not when it ends in a file nobody opens again.

The tools will keep changing, the volume estimates will keep disagreeing with each other, and a new interface for asking questions will keep showing up every couple of years. The process above – anticipate real phrasing, cluster by intent, sort by stage, check the SERP, map to a page – doesn’t depend on any of that staying still.