A practical guide to semantic core development: keyword research, demand analysis, SERP-based clustering, and mapping queries to pages for SEO and paid search.
A semantic core is the structured set of search queries a website is built to answer. It decides which pages exist, what each page targets, and which phrases an advertising account bids on. Specialists at Digital-agency Profi Studio begin almost every SEO or paid search project with this step, because a site built on guesses about how people search tends to rank for the wrong phrases or for none. This article explains how to collect keywords, measure demand, remove noise, and group queries into clusters that map cleanly onto pages and ad groups.
What a Semantic Core Is and What It Is Not
A semantic core is not a long spreadsheet of every phrase related to a business. A raw export from a keyword tool for a mid-sized online store can easily contain 30,000 rows. Most of them are duplicates, irrelevant combinations, or queries with no measurable demand. The finished core is what remains after filtering and grouping: usually a few hundred to a few thousand queries, each assigned to one specific URL.
Three properties separate a working core from a keyword dump. Every query in it reflects real demand confirmed by data. Every query is relevant to what the business actually sells or explains. Every query belongs to exactly one cluster, and every cluster belongs to exactly one page. If two pages target the same cluster, they compete with each other in search results, a problem known as keyword cannibalization.
The core also serves two channels at once. For SEO, it defines site structure, page titles, headings, and content scope. For paid search, it becomes the basis of campaigns, ad groups, and negative keyword lists. Building it once and reusing it for both saves duplicated research.
Search Intent: The First Filter
Before collecting anything, it helps to understand that queries differ in purpose, not only in wording. "Running shoes" and "how to choose running shoes" share most of their words but call for different pages. Search engines classify intent and show results accordingly, so a page that mismatches the dominant intent rarely reaches the top positions regardless of its quality.
| Intent type | Example query | What the user wants | Suitable page type | Value for ads |
|---|---|---|---|---|
| Informational | how often to service a bicycle | An answer or explanation | Blog article, guide | Low to medium |
| Commercial investigation | best road bike under 1000 | To compare options before buying | Comparison, review, category with filters | Medium to high |
| Transactional | buy shimano 105 groupset | To purchase or order now | Product page, service page with form | High |
| Navigational | decathlon bike repair | To reach a specific brand or site | Brand or location page | High for own brand only |
| Local | bike repair near me | A nearby provider | Location page, Google Business Profile | High |
Intent is not always obvious from the words. The query "pizza oven" returns mostly product listings, while "pizza oven temperature" returns articles. The only reliable check is to look at the current search results page. If the top ten are online stores, the query is commercial; if they are blogs and forums, it is informational.
Collecting Seed Keywords
Seed keywords are short, broad phrases that describe the main categories of the business: "bike repair", "bicycle tune-up", "wheel truing". They are the starting points from which tools generate longer variations. A weak seed list is the most common reason a core misses entire segments of demand, so this stage deserves more attention than it usually gets.
Where Seed Keywords Come From
- The product catalogue or service price list, including internal category names and model lines
- Customer language from emails, chat logs, call recordings, and reviews, where people describe problems in their own words
- Google Search Console data for an existing site, which shows queries that already bring impressions
- Competitor navigation menus and category pages, which reveal how the market segments itself
- Synonyms, slang, abbreviations, and common misspellings ("bicycle", "bike", "cycle"; "tuneup", "tune up")
- Problem-based phrasing that does not mention the product at all ("chain keeps slipping", "brakes squeaking")
Customer language is the source most often skipped. A shop may call its service "drivetrain overhaul" while customers search for "replace bike chain and cassette". Only the second phrase has search volume.
Expanding the Keyword Pool
Once the seeds are ready, each one is expanded into longer phrases. A seed like "bike repair" might produce "bike repair cost", "bike repair shop open sunday", "electric bike repair", and hundreds of other variations.
Keyword Research Tools
Google Keyword Planner is free and draws on Google's own data, but it shows volume ranges such as "1K–10K" unless the account has active ad spend. Semrush, Ahrefs, and Serpstat provide exact monthly estimates, keyword difficulty scores, and competitor data, at a monthly subscription cost. For markets where Yandex has a large share, Yandex Wordstat is the primary demand source, and desktop tools such as Key Collector automate mass collection from it.
No tool is accurate in absolute terms. Third-party volumes are modelled from clickstream panels and can differ from real figures by a factor of two or more for low-volume phrases. Treat volumes as relative indicators: a phrase with 2,400 searches matters more than one with 90, even if neither number is exact.
Competitor Gap Analysis
Entering a competitor's domain into Ahrefs or Semrush shows the queries it ranks for. Comparing three to five competitors at once reveals phrases they rank for and your site does not. This method surfaces queries that would never come from seed expansion, such as brand-model combinations or niche problems a competitor happened to write about.
Choose competitors by search visibility, not by business size. A small blog that ranks well for repair guides is a more useful source than a national chain that relies on brand traffic.
Autocomplete, Related Searches, and Question Boxes
Search suggestions, the "People also ask" block, and the related searches at the bottom of the page show phrasing that real users type. Tools such as AnswerThePublic or AlsoAsked collect these at scale. They are especially valuable for informational clusters and for new trends that keyword databases have not yet indexed, since database updates can lag by several months.
Analyzing Demand
Collected phrases must be measured before they can be prioritized. Demand analysis answers three questions: how many people search for this, when do they search, and how hard is it to compete.
Broad, Phrase, and Exact Frequency
A reported search volume can mean different things. In Yandex Wordstat, the default number for "bike repair" includes every query containing those words, such as "bike repair cost moscow". Wrapping the phrase in quotes ("bike repair") restricts the count to queries made of exactly those words in any order and form. Adding the exclamation operator (!bike !repair) fixes word forms as well. The gap between the broad and exact numbers is often tenfold. A phrase showing 50,000 broad searches may have 3,000 exact ones, and only the exact figure tells you how much traffic that phrase itself can bring.
Google Keyword Planner reports closer to exact-match demand by default, but groups close variants and plurals together. Ahrefs and Semrush report per-phrase estimates. Compare numbers only within one tool and one counting method.
Seasonality and Trends
Monthly averages hide seasonal patterns. Demand for "bike repair" in the Northern Hemisphere typically peaks in April and May and falls by half or more in December. Google Trends shows the shape of the curve over five years, which separates stable seasonality from a long-term decline. A phrase with steady 1,000 monthly searches and a phrase averaging 1,000 but concentrated in two months require different content timing and ad budgets.
For SEO, seasonal content should be published two to three months before the peak so that it has time to be indexed and gain positions. For paid search, budgets are shifted toward peak months.
Difficulty and Commercial Value
Keyword difficulty scores estimate how many backlinks and how much authority are needed to reach the top ten. The scales differ between tools, so a score of 30 in Ahrefs is not comparable to 30 in Semrush. A more direct method is to open the search results and check who ranks: if the top positions belong to marketplaces and national media, a new site will struggle regardless of the score.
Cost per click (CPC) from advertising tools works as a proxy for commercial value. A query with a 6 USD CPC signals that advertisers earn money from it. An informational phrase with a CPC of 0.10 USD may still be worth targeting with content, but it should not receive the same priority as a transactional one.
Cleaning the List
A raw expanded list contains a large share of irrelevant phrases. For a bicycle repair shop, these include "bike repair simulator game", "bike repair stand diy plans", "bicycle repair jobs", and queries for cities where the shop does not operate.
Cleaning works best through negative words rather than line-by-line review. Sort the list alphabetically or by word, scan for recurring irrelevant terms such as "free", "jobs", "game", "pdf", "course", or competitor brand names, and remove every phrase containing them. Each negative word found also goes into the negative keyword list for paid campaigns, which prevents ads from showing on those searches.
After negative filtering, remove phrases with zero or near-zero exact demand. A common threshold is fewer than 10 monthly searches, though in narrow B2B niches a phrase with 20 searches a year can still justify a page if each lead is worth thousands. Implicit duplicates also need removal: "repair bike" and "bike repair" are the same query for clustering purposes, and keeping both inflates the core without adding coverage.
Edge cases require judgment. A query like "bike repair kit" is irrelevant to a repair service but highly relevant to a shop that also sells parts. The decision depends on the business model, not on the query alone.
Clustering
Clustering groups queries that can be served by one page. It is the step that turns a list into a site structure.
SERP-Based Versus Semantic Clustering
Semantic clustering groups phrases by shared words or meaning. It is fast and requires no external data, but it often errs. "Bike chain replacement" and "how to replace a bike chain" look nearly identical, yet the first may show service pages and the second video tutorials.
SERP-based clustering compares the actual search results. If two queries return many of the same URLs in the top ten, the search engine treats them as answerable by one page, so they belong together. Tools such as Serpstat, SE Ranking, KeyAssort, and Keyword Insights automate this by collecting results for every query and counting overlaps. This method is slower and costs money for large lists, but it reflects how the search engine itself groups intent.
Thresholds and Clustering Strength
The key parameter is the overlap threshold: how many shared URLs in the top ten make two queries part of one cluster. A threshold of 3 is a common starting point for commercial niches; informational niches often work better at 4 or 5, because results there are more varied.
There are also two grouping logics. In soft clustering, each query only needs to share the threshold number of URLs with the cluster's central, highest-volume query. In hard clustering, every query in the group must share them with every other query. Soft clustering produces fewer, larger clusters and suits category pages. Hard clustering produces tighter groups and suits product pages or narrow articles, where mixing intents would weaken relevance.
Results must also be collected for the right location and device. A query like "bike repair" returns completely different pages in Manchester and in Leeds, so clustering a local business on national results produces misleading groups.
Mapping Clusters to Pages
Each cluster is assigned to one URL. For an existing site, check whether a page already ranks for the cluster's main query in Google Search Console; if it does, that page becomes the target and is improved rather than replaced. For gaps, new pages are planned.
The cluster's highest-volume, most precise query becomes the main focus for the title tag and H1. Secondary queries shape subheadings, body copy, and FAQ sections. Long-tail questions within a cluster often become separate sections of the same page rather than separate pages.
Clusters that do not fit any planned page type signal either a new section of the site or a query that should be dropped. A repair shop finding a large informational cluster around "how to fix a flat tire" can decide to launch a guides section; a store finding a cluster around a product it does not sell should discard it.
Using the Semantic Core for Paid Search
The same clusters translate directly into ad groups. A tight cluster with a shared intent allows one set of ads and one landing page, which raises Quality Score in Google Ads and lowers cost per click. Mixing intents in one ad group produces generic ad copy and lower click-through rates.
Match types follow from demand data. High-volume, clearly transactional phrases can run on phrase match with careful negatives. Rare long-tail phrases are often better covered by broad match combined with Smart Bidding and a strong negative list. Informational clusters usually stay out of paid search entirely, except for remarketing audiences.
The negative keyword list built during cleaning is applied at the account level from day one. Without it, a new campaign typically spends a noticeable share of its first-month budget on irrelevant searches.
Practical Scenario: A Core for a Local Bicycle Repair Shop
Consider a shop in Bristol that repairs road, mountain, and electric bikes, sells parts, and has a five-page website with weak traffic. The process below took roughly 12 working hours.
- Seeds. From the price list and 40 customer reviews, 26 seeds were collected, including "bike service", "e-bike repair", "puncture repair", "brake bleed", and "wheel truing". Reviews added "gears skipping" and "bike won't change gear", which the price list did not contain.
- Expansion. Seeds went through Semrush (UK database), Keyword Planner, and autocomplete scraping. Three local competitors were run through a gap analysis. The raw list held 4,180 phrases.
- Cleaning. A negative list of 64 words removed 2,950 phrases: other cities, "jobs", "course", "stand", "tools", "game", and brand names of motorbikes that share terms. Phrases under 10 monthly searches were cut unless they contained "Bristol" or a district name. 1,020 phrases remained.
- Demand check. Google Trends confirmed a peak from March to June, with April demand roughly 2.3 times December levels. E-bike repair queries showed steady year-on-year growth rather than pure seasonality.
- Clustering. SERP-based clustering on Bristol results, threshold 3, soft logic for services and hard logic for guides, produced 74 clusters.
- Mapping. Clusters were assigned to existing pages, new service pages, and a guides section; 9 clusters were dropped as off-topic.
| Cluster (main query) | Queries in cluster | Combined monthly volume | Intent | Target page |
|---|---|---|---|---|
| bike repair bristol | 38 | 1,900 | Local transactional | Home page |
| e-bike repair bristol | 17 | 480 | Local transactional | New: /e-bike-repair/ |
| bike service cost | 22 | 720 | Commercial | /bike-service/ with price table |
| puncture repair | 14 | 390 | Transactional | New: /puncture-repair/ |
| how to index gears | 26 | 1,300 | Informational | New guide: /guides/gear-indexing/ |
| hydraulic brake bleed | 11 | 260 | Mixed | /brake-service/ with how-to section |
Within four months of launching the new pages, the site's monthly organic clicks rose from about 140 to about 610, with the e-bike page bringing most of the booked repairs. The gear indexing guide drew traffic but few direct bookings; it was linked to the service page, and the paid campaign excluded informational phrases entirely, using the cleaning-stage negatives from the start.
Common Mistakes and Edge Cases
Targeting broad frequency instead of exact frequency leads to inflated forecasts. A business expects 50,000 visits from a phrase that actually brings 3,000 exact searches, half of them going to marketplaces.
Building separate pages for near-identical queries is the second frequent error. "Bike servicing" and "bike service" do not need two URLs; SERP overlap usually confirms they share intent. Two pages split links and relevance signals and often trade places in rankings week by week.
Ignoring mixed intent creates a third problem. Some queries, such as "standing desk", return a mix of reviews, stores, and guides. In those cases the ranking page type with the largest share in the top ten guides the decision, and a hybrid page, such as a category with a buying guide section, often performs best.
Zero-volume queries deserve a separate note. Tools report zero for many phrases that do receive searches, especially in new or narrow niches. If a phrase appears in Search Console impressions, autocomplete, or customer conversations, it has demand regardless of what the database says.
Multilingual and multi-regional sites need a separate core for each language and market. Translating a core from English to German produces phrases Germans do not type, because search habits differ beyond vocabulary.
Maintaining the Core
A semantic core is not finished after launch. New products, new competitors, and shifts in search behaviour change demand every year. A practical schedule is a light review each quarter, checking Search Console for new queries with impressions but no dedicated page, and a full re-collection every 12 to 18 months.
Search Console's query report is the most useful maintenance source because it reflects actual impressions on your site. Queries with high impressions and an average position between 8 and 20 are the cheapest wins: the page already has relevance, and targeted content updates often move it onto the first page. For paid search, the search terms report serves the same purpose, revealing both new keywords to add and new negatives to block.
FAQs
How many keywords should a semantic core contain?
The size depends on the business, not on a fixed target. A local service business may need 200 to 800 queries across 30 to 80 clusters. A large online store can require tens of thousands. The useful measure is coverage: every product, service, and common customer question should have a cluster and a page.
Can I build a semantic core with free tools only?
Yes, for a small site. Google Keyword Planner, Google Search Console, Google Trends, and manual autocomplete collection are enough for a core of a few hundred queries. Manual SERP comparison replaces automated clustering at that scale. Paid tools become necessary for large lists, competitor analysis, and exact volume estimates.
What overlap threshold should I use for clustering?
Start with 3 shared URLs in the top ten for commercial niches and 4 for informational ones. If clusters come out too large and mix different intents, raise the threshold. If nearly every query ends up alone, lower it. Review a sample of clusters by hand before accepting the results.
Is a semantic core for SEO the same as a keyword list for Google Ads?
They share the same foundation but differ in use. The SEO core includes informational clusters that rarely justify ad spend. The ads list emphasizes transactional queries, assigns match types, and relies heavily on negative keywords. Building one core and deriving both from it avoids duplicated work.
How long does semantic core development take?
For a small local business, 8 to 20 working hours. For a mid-sized store with several hundred categories, two to four weeks, most of which goes into cleaning and checking clusters. Collection itself is fast; judgment-heavy filtering is what takes time.
Conclusion
Semantic core development turns guesses about customer behaviour into a measured plan. The sequence is consistent across niches: collect seeds from the catalogue and from customer language, expand them with tools and competitor data, measure exact demand and seasonality, remove irrelevant phrases with negative words, cluster by search results rather than by wording, and assign each cluster to one page. The same structure then feeds paid campaigns with tight ad groups and a ready negative list. A core built this way gives every page a defined purpose and every ad a defined audience, and it stays useful as long as it is reviewed against fresh Search Console and search terms data.