Keyword research: architecture of demand, not a list

By Miguel Maio · Published 16 October 2026

The architecture of demand, layer by layer

1. 1Semantic model of the business: offer, entities, jobs-to-be-done
2. 2Pillars: the core topics and entities
3. 3Clusters and hubs: subtopics and internal-link concentration
4. 4Intent families: group by intent, not by string
5. 5Surface of applicability: search, PAA, AI prompt, YouTube, app

The tool enters at layer 5, not layer 1.

Most of the keyword research I see isn't research. It's an export from a tool, sorted by volume.

I've met SEOs who don't know how to do keyword research. You ask them for a study and they hand you the on-page of a page: title, headings, meta, the on-page factors, with no study behind it to ground the semantics. They optimise an answer without knowing what the question is.

I see people using GSC without knowing its coverage limits, and losing the best opportunities: the keywords you have no content for yet, or that aren't indexed yet, never show up there. They look for the keys under the streetlight, because that's where the light is.

I see the grave error of using Google Trends as a keyword research tool. Trends gives no volume and no keywords. It gives a relative index from 0 to 100.

I see search intent treated as a label stuck on a string.

And I see people who only look at competitors' keywords and get stuck, copying someone else's map instead of drawing their own.

I've been doing this for years. The error is almost always the same, and it's right at the start. The tool isn't the start of the work. It's the middle. Keyword research doesn't start in a tool. It starts in a model of the business's demand. Let me go in the right order.

First the business, then the demand

Before opening anything, I answer three things: what this business sells and solves, for whom, and which entities and topics it has a right to be relevant for. That's the semantic model: the offer, the jobs-to-be-done, the entities, the customer's problem space.

This sounds obvious and almost no one does it. Without this ground, the study is garbage-in. A tool returns a thousand keywords; without a model of the business, you have no way to tell which are your demand and which are noise. You end up reacting to volume instead of mapping your territory.

The semantic model is also what lets you see what's missing. A gap only exists against a map. If you have no map, you don't see the hole, and the hole is where the traffic your competitor is taking from you lives.

Give demand an architecture

Once demand is mapped onto the model, you organise it. Not alphabetically and not by volume. By architecture.

Pillars: the core topics and entities of the business, which usually align with the major sections of the site or app. Clusters: the subtopics within each pillar. Hubs: the points of internal-link concentration that build topical authority around a pillar. Intent families: grouping by intent, not by string. Two hundred keywords can be eight families. The same family asks for the same type of content and the same surface.

The difference between amateur and senior is here. The amateur has a list of two hundred rows. The senior has eight families, each with its pillar, its cluster, its hub and its surface. One is a file. The other is a plan.

Search is not the only surface

A business's demand doesn't live only in Google search. And even within Google, it doesn't show up all the same way.

The same intent appears differently depending on the surface: classic search, People Also Ask, autocomplete, a ChatGPT or Perplexity prompt, YouTube (the second largest search engine, with its own intent), Pinterest (discovery and visual search, strong in some verticals), LinkedIn (where much of the B2B demand that Keyword Planner doesn't even know lives), and the internal search of the app or marketplace itself.

That's why each cluster carries a surface tag. It's not a detail: the content that answers a search query isn't the one that answers a PAA question, which isn't the one that wins an AI prompt. Treating everything as "Google keywords" leaves half the demand on the table. The map of signals by channel is in the pillar on media channels as GEO intelligence.

Search intent, done properly

The four textbook types (informational, navigational, commercial, transactional) are the start of the conversation, not the end.

Intent isn't in the word. It's in what the surface already shows. The SERP is the truth: if Google fills the page with videos, the intent is video, whatever you write. If it fills with products, an article isn't what you want there. PAA shows you intent as a question. The AI answer shows you which entities it already considers part of the topic.

And there's what the manuals don't say: intent is mixed (the same query serves two purposes), it drifts over time (what was informational turns transactional in a season), and it's local without declaring it. Reading the SERP before writing a word isn't an optional step. It is the step.

Every source has its blind spot

Now, the data. And the first senior rule is this: no source gives you all the demand, and each one lies in its own way. Whoever doesn't know the origin of the data trusts numbers they shouldn't.

GSC is your data and it's the best you have. But it doesn't cover your whole site. First, it only reports indexed URLs, and Google doesn't index everything: in a study of 16 million pages, about 37% were indexed. Second, even among those, 46.77% of clicks come from anonymized queries (Ahrefs, 22 billion clicks, April 2025). That is, nearly half the queries bringing you traffic are hidden. GSC is three stacked filters: indexed, with impressions, and survived anonymization.

The Google Ads search terms report is gold, because it's what people actually typed. But since September 2020 Google hides low-volume terms. In real accounts, over 30% of traffic started coming from hidden terms. There's a third of your spend whose exact queries you can't see.

Keyword Planner is Google's own source, but it's built for advertisers: it rounds into buckets, groups synonyms and, by Ahrefs' own analysis, overestimated volume 54% of the time. It's a benchmark for nothing.

Third-party suites (Ahrefs, Semrush) are modeled estimates, not counts. Ahrefs itself calls its data "directionally accurate" and was roughly right for 60% of keywords against Search Console. They're for relative prioritisation. Absolute volume is an illusion of precision.

The practical conclusion: triangulate, start from your own data (GSC and the Ads search terms), and always read volume as relative.

No source covers all the demand. Knowing each one's bias is what separates research from an export.

High demand is not opportunity

Here lives the most expensive error. Keyword Planner volume is advertiser interest. For SEA, high demand is reach, and it's paid in CPC and competition. For SEO, high demand is often a trap.

High-volume head terms are dominated by big brands, have high difficulty, ambiguous intent and low conversion. And, increasingly, they generate no click at all: AI Overviews answer on top of them. Chasing the fat term because it has volume is spending months fighting for a position that no longer brings traffic.

In SEO, the value is in the specific long tail, with clear intent and a gap that's yours. Less volume, more conversion, less blood. High volume is for SEA to validate and buy. In organic, it's usually the most expensive path to the worst result.

The channels feed each other

Separating keyword research by channel wastes the best part. The sources cross over.

The Ads search terms tell you which queries actually convert. Those are priority SEO targets, and the ones that don't convert are negatives. The queries bringing organic traffic are candidates to expand in SEA. And in GEO the study stops being about keywords and becomes about questions, prompts and entities: the engines' query fan-out forces you to cover sub-questions, not just the main term. Because the engines cite different sources, that's in why each AI engine cites different sources.

And here I answer a question I hear: is Google Ads to throw away in research? Not at all. Google Ads is the only place you get real conversion data per query. You run a campaign, put money in, and in days you know which terms convert before investing months in SEO or GEO. It's the best validator there is, not the best estimator. Throwing Ads away as a research tool is throwing away the only source of truth about what converts. Keyword Planner and the search terms live inside it.

And yet I see SEMs who can't prevent internal cannibalization between pages, nor articulate Search Ads with SEO. It's a waste of everything. Sometimes because they're different departments. Sometimes because they're different product owners. Sometimes over compliance between client and agency. How did we get here, if the site is one and the market is one? How did we build structures so compartmentalised that paid and organic don't even talk to each other, when they win from exactly the same queries?

Search Ads and SEO compete for and feed the same demand. The terms Ads proves convert are SEO's brief. The queries where organic is already strong are where Ads can stop paying. And the relevance a good intent study gives the ad is what improves Quality Score and lowers cost per click. Whoever runs the two in silos pays twice for the same market.

Negatives deserve the same seriousness as targets. An n-gram analysis of the search terms builds the SEA negative list and reveals, along the way, what you shouldn't even try to target in SEO: the anti-keywords. The account structure that supports this is in a Google Ads account structure that converts.

Only the tools worth it, and by function

You don't need ten tools. You need to know what each one does best.

Google Ads (with Keyword Planner and the search terms): the real conversion validator and Google's own sense of demand. Google Search Console: your data, the most undervalued of all. One suite, Ahrefs or Semrush, not both: difficulty, SERP, and above all content gap (the keywords your competitor has and you don't, and the topical gap against your model). Google Trends: for what absolute volume can't give, seasonality, trend direction and breakouts.

GenAI for the semantic layer: expansion, clustering, intent classification, generating questions and prompts for GEO, entity and topic modeling. With a hard limit: never for volumes, which it invents. For semantics, it saves hours; for data, it lies. Question mining: AlsoAsked, PAA, autocomplete, and the Reddit and forums of the niche, which is where you catch the real questions for AEO and GEO. Screaming Frog: crawl your site and cluster by embeddings, which I cover in Screaming Frog for clustering by embeddings. And, by surface, the native tools: YouTube, Bing Webmaster, Pinterest Trends, LinkedIn.

Three or four of these cover 95% of the work. The rest is the semantic layer (GenAI) and your own data. Whoever jumps from tool to tool is running from the work of modeling the business.

Prioritise like someone who decides, not someone who lists

Sorting by volume is the lazy way. I use ETI, but with two corrections that make it honest.

First: I apply ETI to the cluster or intent family, not the keyword. Effort, time and impact are properties of covering a topic, not of a string. ETI keyword by keyword is the wrong granularity.

Second: effort and time are costs. If you add them straight to impact, the average lies, because more effort would give more priority. They're scored inverted: Effort as ease (1 heavy, 10 light: semantic gap, create from scratch vs optimise, difficulty) and Time as speed (1 slow, 10 fast: new cluster vs page to optimise, indexing lag, authority still missing).

Cluster ETI = (Impact + Ease + Speed) / 3, scale 1 to 10. Impact is business value by proximity to conversion by relative demand. Prioritise clusters by ETI, and within an approved cluster sequence the keywords. It's a number that ranks without faking the volume precision the tools don't have.

What you deliver

The deliverable of senior keyword research isn't a list of keywords. It's a demand model.

Each row carries: term, source, volume (with the source, treated as relative), trend, intent type, SERP-proven intent, journey stage, channel (SEO, SEA or both), question or prompt variant for GEO, entity and cluster, current rank and URL from GSC, cannibalization flag, difficulty, business value, cluster ETI, and the action (create, optimise, merge, negate, or validate with Ads).

On top, a measurement layer in Looker Studio that blends GSC with the Ads search terms, so the study doesn't die the day it was delivered. Keyword research isn't an event. It's a system you review.

Errors no one running SEM can make

These errors are serious. Doing on-page SEO with no study behind it. Trusting a GSC that hides half. Using Trends as if it were a keyword tool. Copying the competitor and getting stuck. None of these is a detail. They're the difference between occupying a territory and reacting to someone else's.

Whoever runs SEM can't make them. And can't delegate or skip them. Keyword research isn't a task you order and file away: it's the foundation of everything that comes after, in SEO, in SEA and in GEO. Whoever runs SEM has to be the true owner of keyword research. Not the owner of the file. The owner of the decision.

Whoever starts in the tool will always have a list. Whoever starts in the business will have a map. And it's the map that decides where the competition doesn't catch you.

Sources

• Ahrefs - Anonymized Queries Make Up Nearly Half of Google Search Console Traffic
• indexcheckr - Google Indexing Study: Insights from 16 Million Pages
• Practical Ecommerce - Google Ads Decreases Search Terms Visibility
• Ahrefs - How accurate is search volume in Ahrefs

Frequently asked questions