Every SEO programme I have seen fail at the planning stage failed in the same way: someone exported a few thousand keywords from a tool, sorted by volume, and started writing from the top. The result was a set of pages chasing numbers that were either wrong, unreachable, or pointed at an intent the business could not serve. Keyword research is supposed to prevent that. This guide is how I actually do it, with the parts that tools get wrong called out explicitly.
I will cover the difference between demand and intent, the seed, expansion and clustering workflow, why search volume and difficulty scores need to be treated with suspicion (especially for Indian markets), how to map keywords to pages, how to prioritise, and a worked example you can copy.
Demand is not the same as intent
A keyword tool measures demand: roughly how many times a phrase was searched. It does not tell you what those people wanted, and that is the thing you are actually trying to serve. "CRM software" has large demand, but the people typing it include students writing assignments, founders comparing vendors, employees looking for a login page and journalists researching a story. One page cannot satisfy all of them, and Google will choose which intent to favour based on what most searchers seem to want.
So I treat every keyword as two things: a demand number, which is approximate, and an intent, which I read from the live search results. If the top results for a phrase are category pages, the intent is commercial and I need a category page. If they are long guides, the intent is informational. If they are mixed, Google is hedging, and I may be able to win with either type or with a page that does both. The search intent guide covers classification in more depth; the short version is that the SERP is the ground truth and the tool is a hint.
The workflow: seed, expand, cluster, map
I run keyword research as a pipeline. Each stage has an input, a method and an output, and I do not skip stages because the later ones depend on the earlier ones being clean.
- Seeds
- Expansion
- Clustering
- Mapping
- Prioritisation
Seeds
Seeds are the handful of terms that define what the business does, written the way customers say them rather than the way the company does. A payments company might call itself a "fintech infrastructure platform"; its customers search for "payment gateway for website" and "UPI integration". I get seeds from the sales team, support tickets, the existing Search Console queries, competitor navigation and plain conversation with customers. Ten to thirty good seeds is usually enough.
Expansion
Expansion turns seeds into a long list of candidate queries. Sources I use: keyword tools (any of the major ones; they overlap heavily), Google autocomplete and the People Also Ask boxes, related searches at the bottom of the SERP, Search Console queries that already send impressions, competitor pages that rank, and community sources such as forums and Q&A sites where people phrase problems in their own words. The output is messy by design. A few thousand rows is normal for a mid-sized site.
Clustering
Clustering groups queries that share an intent and would be satisfied by the same page. This is the stage most people get wrong, because they group by words instead of by intent. "payment gateway charges" and "payment gateway fees" are the same cluster. "payment gateway charges" and "payment gateway integration" are different clusters, even though they share two words. The reliable test is SERP overlap: if two queries return mostly the same pages, they belong together. Several tools automate this; a manual check on a sample keeps the automation honest.
Mapping
Mapping assigns each cluster to a page, either one that exists or one that needs to be built. One cluster, one page. If two clusters map to the same page, either they are really one cluster or the page is trying to do too much. If one cluster maps to two pages, you have cannibalisation, and Google will pick one of them, often not the one you want. The keyword mapping guide goes into the mechanics; it is the stage where keyword research turns into site architecture.
The limits of search volume
Search volume numbers are estimates derived from samples, clickstream panels and advertising data, then smoothed over a year. They are reasonable for broad, high-demand English queries in large markets. They are poor for almost everything else, and the "else" is where most useful queries live.
In India the problem is sharper. Volumes for specific, commercially valuable queries are often reported as zero or "10–100" when Search Console shows real impressions. Hindi and regional-language queries are badly covered. Transliterated queries ("paise kaise bheje") are barely covered at all. And the Indian search market is growing in ways that a twelve-month average will understate. If you filter out everything under a volume threshold, you throw away a large share of the demand that actually converts.
I use volume for relative ranking within a cluster and for sanity checks, not as a gate. The better demand signals are impressions in Search Console, the number of distinct phrasings in a cluster (many phrasings usually means real demand even if each shows as zero), and whether competitors have bothered to build a page for it.
Keyword difficulty: what the number actually measures
Most difficulty scores are a function of the link profiles of the pages currently ranking. That is one input into whether you can rank, but it is not the whole picture. A query with a high difficulty score might be winnable if the ranking pages are off-intent, outdated or thin. A query with a low score might be unwinnable if the SERP is dominated by a feature, such as a local pack or a product grid, that your page type cannot enter.
My practical read of difficulty combines four checks: the type of pages ranking (can I build that type?), the authority of the domains (are they all national publishers and marketplaces, or are there smaller sites in the top ten?), the freshness of the results (is there an obvious gap?), and the SERP layout (how many organic results are even visible above the fold on mobile?). A spreadsheet score is a starting filter. The judgement comes from looking.
Long-tail queries and why they add up
Long-tail queries are the specific, lower-demand phrasings that make up most of what people actually type. "best payment gateway" is head; "payment gateway for small business in India with low fees" is long-tail. Individually they are small. Collectively they are often larger than the head terms, and they convert better because they are specific.
The mistake is to build a page for each one. Long-tail queries belong inside the page that serves their cluster, as sections, FAQs or examples. A single thorough page on payment gateway fees can rank for dozens of fee-related phrasings. A dozen thin pages on each phrasing will cannibalise each other and look like spam. This is the practical link between keyword research and the way I approach topic clusters.
Prioritisation: what to build first
After mapping, a mid-sized site often has fifty to several hundred clusters. Nobody can build them all at once, and building them in volume order is usually wrong. I score clusters on four dimensions and build in score order, with the caveat that technical blockers on existing pages always come first.
Business value
How close is this intent to revenue? A query about pricing or comparison outweighs a definition query at the same volume.
Winnability
Based on the SERP read above, not the tool score. Existing pages that already get impressions at positions 8–20 are usually the fastest wins.
Effort
Does this need a new template, a developer, a data source or a subject-matter expert, or is it a writer and a brief?
Cluster leverage
Does building this page strengthen a topic the site already covers, or start a new one from zero? Completing a cluster compounds; scattering does not.
Tools: what they are good and bad at
I use keyword tools from the major vendors for expansion and for rough demand, Search Console for real demand, a crawler for seeing which pages currently rank for what, and a spreadsheet for everything else. The specific brand matters less than people think; the databases overlap, and all of them share the limitations above. The one tool I would not skip is Search Console, because it is the only source that reports what actually happened on your site rather than a modelled estimate. The SEO tools guide covers what each category of tool is for.
A worked example
Suppose an Indian D2C brand sells air purifiers, mainly in Delhi NCR and other north Indian cities. The seed terms are "air purifier", "air purifier for home", "HEPA filter", and "AQI". After expansion and clustering, a handful of the resulting clusters look like this. The structure demonstrates how intent groups map to distinct page types.
| Cluster (sample queries) | Intent and format | Mapped page | Demand signal | Priority |
|---|---|---|---|---|
| best air purifier for home india, best air purifier under 10000, top air purifiers 2026 | Commercial, comparison list | Buying guide (new) | [DATA], many phrasings, strong PAA | High: closest to purchase, winnable against listicles |
| air purifier price, air purifier for bedroom, buy air purifier online | Transactional, product grid | Category page (exists, thin) | [DATA], Search Console impressions at positions 11–18 | High: existing page, fast to improve |
| what is aqi, aqi delhi today, aqi levels explained | Informational, quick answer with table | AQI explainer (new) | Seasonal spike each winter; local pack and live data widgets in SERP | Medium: huge but hard to monetise and dominated by weather sites |
| hepa filter vs activated carbon, do air purifiers work for dust, air purifier for allergies | Informational, explainer | Cluster of three articles under the buying guide | [DATA], low tool volume, real impressions | Medium: builds topic depth that supports the buying guide |
| air purifier filter replacement, when to change hepa filter | Informational, existing customer | Support article + product page section | Small, but retention value | Low for acquisition, high for support |
| air purifier ka price, air purifier kaise kaam karta hai | Transliterated Hindi, mixed | Fold into category page and explainer; test a Hindi version later | Tool reports zero; Search Console shows impressions | Test: cheap to add, under-served |
Notice what the table does. It turns a keyword list into a page list with reasons. Nobody has to argue about volume. The decision about what to build first is visible, and the SERP evidence behind each row can be checked. This is the artefact I hand to writers, developers and clients, and it is what the SEO content framework is built around.
What goes wrong
The failures I see most often: filtering by volume and losing the long tail; grouping by word overlap and creating cannibalising pages; researching once and never revisiting; ignoring the SERP and building the wrong page type; and treating keyword research as a content task when half the output is really architecture. The site architecture guide picks up where mapping ends, because where a page sits in the site affects whether it ranks at all.
Go deeper
- Keyword mapping. How to assign clusters to pages, detect cannibalisation, and turn the map into a build plan.
Keyword research feeds directly into content SEO, and the clustering logic here is the foundation of topical authority. For how search engines model the meaning behind those clusters, read semantic SEO.
Frequently asked questions
How many keywords should a page target?
One cluster, which may contain anything from a few to a few hundred phrasings. The page targets the intent, and the phrasings follow naturally. If you find yourself listing variants to insert, you are back to keyword density thinking.
Should I ignore keywords with zero search volume?
No, especially in India. Zero in a tool usually means the sample did not capture it, not that nobody searches it. Check Search Console impressions and the number of related phrasings before dismissing a cluster.
How do I handle Hindi or regional-language keywords?
Treat them as their own clusters with their own SERPs. Tool coverage is weak, so rely on autocomplete, Search Console and talking to users. Decide deliberately whether to serve them with a section on an English page, a translated page or a separate language version; each has different technical implications covered in international SEO.
How often should keyword research be redone?
The full exercise once a year or at major business changes. The Search Console review, where you look for new queries that are gaining impressions, every quarter. Demand shifts faster than the tool databases.
Is keyword difficulty reliable?
It is a rough proxy for the link strength of the current results. It ignores intent mismatch, freshness and SERP layout. Use it to sort, then look at the actual results before deciding.
What is the difference between keyword research and topic research?
Keyword research starts from phrasings and groups them upward into intents. Topic research starts from a subject and asks what a complete treatment needs to cover. I do both; the keyword side tells you where the demand is, and the topic side tells you what a page needs to include to be the best answer.
Where this connects
The clustering and mapping approach here is formalised in the SEO content framework, and its central claim, that completing a cluster beats scattering pages, is tested in the topical authority cluster test. The SaaS organic growth case study shows the method applied to a real product site. If you want this done for your business, it is the first phase of my SEO strategy work.