Most keyword research produces a spreadsheet nobody uses. Here's how to build one that decides your site structure, your content plan and your priorities — in that order.
Every SEO project starts the same way: someone exports five thousand keywords, sorts them by volume, highlights the big ones, and files the spreadsheet away forever. Three months later, content is being written on whatever topics felt right that week.
That’s not a semantic core. That’s a keyword graveyard.
A real semantic core is something else entirely: a map of everything your customers search on their way to paying you, organized into groups, with each group assigned to a specific page — existing or planned. Done properly, it stops being a research document and becomes the operating system of your site: it decides your structure, your content calendar, and where every euro of SEO budget goes first.
This guide walks through how we build them — and where the revenue connection usually breaks.
Why "sort by volume" quietly loses money
Search volume is the most seductive and least useful column in your spreadsheet. Three reasons:
Volume ignores intent. “Wooden pallets” gets vastly more searches than “buy EPAL pallets bulk price” — but the second query is a purchase order looking for a supplier, and the first is mostly students, competitors and the idly curious. If you rank #1 for the giant keyword and nobody buys, you’ve won a trophy, not a market.
Volume ignores winnability. The biggest keywords in your niche are owned by marketplaces, Wikipedia and companies with six-figure budgets. A keyword you can actually reach in six months is worth more than one you’ll never touch.
Volume ignores the money path. Some queries sit next to a credit card; others sit three research sessions away. Both matter — but treating them identically is how blogs end up with traffic charts going up and revenue staying flat.
The fix isn't abandoning data. It's adding two dimensions volume doesn't capture: intent and page mapping.
The five-stage build
Stage 1 — Collect wide, from more than one well
Start by gathering raw material from every source that reveals how people actually phrase things:
- Your seed terms — products, services, problems you solve, in every wording you’d use
- Competitor rankings — export what the top 3–5 competitors rank for; their keyword sets are years of research you get for free
- Search Console — queries you already get impressions for are the warmest opportunities you own
- Autocomplete, People Also Ask, related searches — the questions real people type
- Your sales team and support inbox — the phrases customers use before they know your industry’s vocabulary
A serious core for a mid-sized site starts at 1,000–3,000 raw keywords. Don’t filter yet — premature filtering is where good long-tails die.
Stage 2 — Clean ruthlessly
Now remove what will never pay: other brands’ navigational queries, jobs and salary searches, DIY intent if you sell done-for-you, free-seekers if you sell premium, and regions you don’t serve. Expect to cut 30–50% of the raw list. Every keyword that survives should pass one test: could a person typing this ever become our customer? Not “will they” — could they.
Stage 3 — Cluster by SERP, not by words
This is the stage most DIY cores get wrong, and it’s the one that determines your entire site structure.
Two keywords belong on the same page if — and only if — Google shows substantially the same results for both. Not if the words look similar. “Pallet delivery” and “pallet shipping” might share a page; “pallet delivery” and “pallet delivery cost” might need two, because one SERP shows service pages and the other shows pricing guides.
Word-similarity clustering tools guess. SERP-overlap clustering checks. The difference sounds academic until you realize what’s at stake: cluster too aggressively and one page tries to rank for incompatible intents, satisfying none. Cluster too finely and you create ten thin pages cannibalizing each other. Both mistakes cost exactly the same thing — rankings you should have had.
The output of this stage: your keywords collapsed into clusters of 3–30 queries, each cluster = one future page with one primary keyword and a family of variations.
Stage 4 — Score clusters by revenue distance
Here’s where the “maps to revenue” part becomes literal. For each cluster, assign two scores:
Intent stage — how close is this searcher to money?
| Stage | What they’re searching | Example | Page type |
|---|---|---|---|
| Transactional | Ready to buy or order | “buy X”, “X price”, “X supplier” | Product / service page |
| Commercial | Comparing options | “best X”, “X vs Y”, “X reviews” | Comparison / category |
| Informational | Solving a problem | “how to choose X”, “what is X” | Guide / blog |
| Navigational | Finding someone specific | Brand names | Usually skip |
Winnability — realistically, can you reach the top 10 in 6–12 months? Look at who ranks now: if it’s all DR80 giants and marketplaces, the honest answer is no, park it. If mid-sized sites and thin pages rank, that’s your green light.
Now the priority order writes itself: transactional + winnable first, commercial second, informational to support them. This ordering is the single biggest difference between cores that generate revenue and cores that generate blog traffic.
Stage 5 — Map every cluster to a URL
The final act: every cluster gets an address.
- Cluster matches an existing page → mark it for optimization
- Cluster has no page → it just defined your content plan
- Two clusters point at one page → split it
- Two pages chase one cluster → merge them (that’s cannibalization, and it’s costing you positions right now)
What emerges is your site’s target architecture: which categories exist, which service pages you’re missing, which guides support which money pages, and how they interlink. The semantic core stops being research. It is the plan.
The multilingual trap
One warning for anyone expanding across markets, because we see this mistake weekly: translated keywords are not a semantic core.
Germans don’t search a translation of what Latvians search. Different phrasings dominate, volumes redistribute, and competitors are entirely different. A word-for-word translated core routinely misses 40–60% of real local demand — and targets phrases nobody types. Every language market needs its own Stage 1–5, built from local data. It costs more than translation. It’s also the difference between entering a market and merely existing in it.
- Dynamic Adaptability: Unlike traditional software that breaks when an unexpected variable appears, Agentic AI can "reason" its way through anomalies.
- Tool Interoperability: Agents act as the ultimate glue. They can use APIs, browse the web, and interact with legacy systems just as a human would, but at machine scale.
- Self-Correction: Modern workflows now include feedback loops. If an agent fails at a task, it can analyze the error and try a different path without human intervention.
What "done" looks like
A finished semantic core is one spreadsheet where every row answers four questions: What’s the keyword? What cluster does it belong to? What intent stage is that cluster? What URL owns it?
From that single artifact you can read off your content calendar for the next six months, your on-page optimization queue, your internal linking map, and — this is the part that makes it a business document rather than an SEO document — a defensible answer to “why are we writing this page now?”
The answer is never “because it has 5,000 searches.” It’s “because it’s transactional, winnable, and two links away from checkout.”
That’s a semantic core that maps to revenue. Everything else is a spreadsheet.
I look forward to seeing how these developments will improve service levels and customer satisfaction in the freight industry!