Keyword, Entity, Topic, and Fan-Out Research Without Duplicate Pages

Direct answer

Good SEO research starts with an audience and a task, not a spreadsheet of keywords. Use keyword language to learn how people express the task, entities to define the people, things, attributes, and relationships involved, result pages to observe current answer formats, first-party data to uncover real constraints, and fan-out hypotheses to anticipate supporting questions.

Then make one of three decisions for every sub-question: answer it as a section, give it a separate page because it completes a different task, or do not publish it. A wording variation does not deserve its own URL.

The goal is not the biggest topical map. The goal is a small set of canonical pages with distinct outcomes, enough evidence, and no hidden duplication.

What you will be able to do

By the end, you will be able to:

• distinguish a query, keyword, topic, entity, attribute, relationship, prompt, and fan-out query;
• expand a seed topic using search and first-party evidence;
• map primary and supporting entities without keyword stuffing;
• hypothesize the sub-questions needed to answer a complex task;
• decide whether a sub-question belongs in a section, a new page, or nowhere;
• check planned pages for cannibalization and missing evidence.

The terms, in plain language

Query

A query is the actual input sent to a search or answer system. It can be typed, spoken, uploaded as an image, or expressed across a conversation.

Example: how should a SaaS handle hreflang for English and French?

Keyword

A keyword is the phrase or concept an SEO team records and analyzes. It may represent one query or a family of similar queries.

Example: SaaS hreflang.

A keyword is a research abstraction. It is not the person.

Topic

A topic is a broader area of knowledge or work.

Example: international SEO.

A topic can contain many distinct tasks. “Cover the topic” is not a page purpose.

Entity

An entity is a distinct thing or concept that can be identified consistently, such as a company, product, person, country, language, protocol, or metric.

Examples: SEOryon, France, French, hreflang, Google Search Console.

Entity clarity means the page consistently explains what each thing is and how it relates to the task. Repeating an entity name twenty times does not create clarity.

Attribute

An attribute describes an entity.

Examples: a locale has a language and region; a URL has a status and canonical; a SaaS tenant has a domain and access boundary.

Relationship

A relationship connects entities.

Examples:

• an alternate URL represents a language version;
• a canonical identifies a preferred representative;
• a tenant owns content within an isolation boundary;
• a sitemap lists canonical public URLs.

Relationships often contain more information than a raw keyword list because they expose what needs explanation.

Intent

Intent is the task a person is trying to complete under conditions. You learned how to document it in the search intent lesson.

Prompt

A prompt is the input and instruction context given to a generative system. It may contain several questions, constraints, and follow-ups.

Fan-out query

A fan-out query is a related search a system may issue to research part of a complex task. Google documents query fan-out for its AI features, but it does not publish hidden query logs for you to copy.

Google's example begins with a lawn-repair question and expands into chemical options, non-chemical options, and prevention. The lesson is that a complete answer may need several evidence branches. It is not that every branch should become a separate article.

What each research source can and cannot tell you

The most valuable map combines sources. A thousand keyword exports cannot replace ten honest support conversations, and ten support conversations cannot estimate a market.

A careful note on Google Trends

Google's Trends methodology explains that public Trends data is sampled and normalized. Results are scaled from 0 to 100 relative to the highest point in the selected comparison. Low-volume and repeated searches are filtered, and the data can contain noise.

A value of 100 means peak relative interest within that request. It does not mean 100 searches, 100,000 searches, or 100 percent market share.

Google also says internal AI Mode and AI Overview searches are excluded from the public Trends dataset. That means Trends and internal product observations do not have identical universes.

Use Trends to ask questions such as:

• is interest seasonal?
• is one term relatively more common in France than the United States?
• did the relationship between two terms change?

Do not multiply a Trends score by a conversion rate as if it were volume.

The research workflow

Step 1: define the audience, task, and output

Write:

We are helping [audience] make [decision] under [conditions] by producing [output].

Example:

We are helping a SaaS technical lead choose a safe English and French URL architecture for a multi-tenant product without leaking tenants or creating duplicate locale pages.

This prevents the research from becoming “everything about international SEO.”

Step 2: collect first-party language

Gather anonymized questions from product, support, sales, onboarding, Search Console, and on-site search. Record where each phrase came from and whether it represents one person or a repeated pattern.

Never paste private tickets, customer domains, email addresses, or tenant data into a public content brief.

Step 3: expand query language

Use:

• Search Console queries;
• keyword databases;
• autocomplete and related questions;
• current results;
• community questions;
• synonyms and product terminology;
• other languages used by the target market.

Preserve the raw language before clustering. It tells writers how people describe their problem.

Step 4: map entities and relationships

For each important entity, record:

• stable name and aliases;
• type;
• attributes required for the decision;
• relationships to other entities;
• primary source;
• common ambiguity.

For international SEO, the main value may be the relationship between locale URL, canonical, hreflang alternate, tenant, and market.

Step 5: build plausible fan-out branches

Ask what evidence a good answer would need. For a complex page, branch into:

• definitions;
• eligibility or requirements;
• options;
• trade-offs;
• procedure;
• verification;
• costs;
• risks;
• failure recovery;
• examples.

Label these as SEOryon hypotheses unless an engine explicitly exposes them. Their purpose is editorial completeness, not reverse engineering.

Step 6: observe current answer formats

Record which page roles appear for representative queries. Look for tools, guides, comparisons, product pages, forums, video, local, and generated answers. Note locale, device, date, and personalization state.

Step 7: attach evidence needs

Every planned claim needs an evidence source. If the page promises a comparison but no reliable pricing or method is available, that is a content gap, not a writing task.

Step 8: decide section, page, or no page

Use this rule:

Create a section when the sub-question supports the same reader, decision, evidence set, and next action.

Create a new page when it serves a distinct outcome, audience, format, substantial evidence set, or operational stage and can stand alone.

Create no page when the variation adds no unique outcome, cannot be supported, has no durable value, or belongs in product support rather than public editorial content.

Step 9: assign a canonical owner

One route must own each primary intent. Record existing URLs, proposed URLs, and merge or redirect decisions before drafting.

Step 10: run the duplication test

For every pair of pages, ask:

• do they help the same audience make the same decision?
• would the same four sections answer both?
• do they need the same evidence and CTA?
• would combining them improve the reader's task?

If most answers are yes, the pages probably duplicate each other.

The page decision table

Worked example: international SEO for a multi-tenant SaaS

Audience and task

The audience is a technical SEO lead and engineering owner. They need to launch English and French public content while preserving tenant isolation and a reversible deployment.

Entity map

• SaaS application: owns product and public marketing surfaces.
• Tenant: owns private or tenant-specific data.
• Locale: a language or language-region experience.
• URL: represents a public resource.
• Canonical URL: preferred representative in a duplicate cluster.
• Hreflang alternate: equivalent locale version.
• Sitemap: lists intended canonical public URLs.
• Access control: protects private resources.
• Release: changes routing, rendering, and index signals.

Important relationships:

• a public lesson has one canonical URL per true language version;
• equivalent English and French pages can reference each other as alternates;
• a tenant must not appear in another tenant's sitemap, cache, page, or structured data;
• parked locales should not create indexable thin copies;
• a rollout needs monitoring and a rollback state.

Plausible fan-out branches

1. subdirectory, subdomain, or separate domain;
2. language versus region targeting;
3. canonical and hreflang alignment;
4. localized internal links and navigation;
5. translation quality and market adaptation;
6. public versus tenant-specific routes;
7. sitemap generation by locale;
8. CDN and cache keys;
9. phased release and monitoring;
10. rollback and redirect strategy.

Page decisions

The central architecture guide should explain route models, locale ownership, public and private boundaries, and a rollout decision table.

A full hreflang implementation guide deserves a separate technical lesson because it needs code, reciprocal alternate tests, canonical checks, and failure diagnostics.

Tenant isolation within programmatic SEO deserves its own advanced lesson because the security and scale decisions are broader than localization.

Release monitoring and rollback belong in a launch operations lesson. The architecture guide should link to it and summarize the dependency, not duplicate the runbook.

Variations such as SaaS hreflang, hreflang for SaaS, and how to hreflang a SaaS site do not each need a page.

Missing evidence

Before drafting, the brief still needs:

• the site's actual routing framework;
• canonical origin convention;
• real supported locales;
• public versus authenticated route inventory;
• cache and tenant-key behavior;
• sitemap generation logic;
• a synthetic test fixture;
• official Google multilingual documentation.

This is information gain in practice. The valuable page cannot be written from keywords alone.

Why this can fail

An enormous topical map can hide an undefined product decision. If the team has not decided whether French is a true translated experience or a redirect, no amount of query expansion can produce a correct canonical and hreflang plan.

Common mistakes

Copying autocomplete as demand truth

Autocomplete is a discovery source shaped by a product. Validate with other evidence.

Treating Trends 100 as volume

It is a normalized peak within the selected request.

Accepting automated clusters without review

Tools group patterns. A human must confirm that each page owns a distinct task and evidence set.

Publishing one URL per long-tail phrase

This creates duplication, maintenance cost, and weak pages. Consolidate by outcome.

Hiding weak purpose inside a giant cluster

Topical breadth cannot rescue a page that has no audience, decision, or original value.

Repeating entities instead of explaining them

Name the entity consistently, define it once, and explain its relationships. Repetition without meaning is still keyword stuffing.

SEOryon Entity and Fan-Out Mapper

Download the Entity and Fan-Out Mapper.

For every row, complete:

• the sub-question;
• the decision it supports;
• the entity and relationship;
• the best evidence source;
• the best answer format;
• the current canonical owner;
• section, new page, link, merge, or no-page action;
• duplication risk.

Before approving a new page, require this sentence:

This URL is different because it helps [audience] complete [unique outcome] using [distinct evidence or format].

If the writer cannot complete that sentence without repeating another page, do not create the route.

Exercise: consolidate 25 fragments

Group these fragments into at most six page roles:

1. international SEO SaaS
2. multilingual SaaS SEO
3. subfolder vs subdomain languages
4. language subdirectories
5. hreflang SaaS
6. hreflang examples
7. x-default meaning
8. canonical with hreflang
9. French translation SEO
10. AI translated pages
11. localization QA
12. international keyword research
13. French keywords for SaaS
14. locale sitemap
15. multilingual sitemap
16. geo redirect SEO
17. browser language redirect
18. tenant pages indexed
19. tenant sitemap leak
20. programmatic tenant pages
21. launch French site
22. locale migration checklist
23. monitor international SEO
24. rollback hreflang
25. international SEO reporting

Suggested answer

1. International architecture decision guide: 1, 2, 3, 4, 16, 17.
2. Hreflang and canonical implementation: 5, 6, 7, 8.
3. Localization research and quality: 9, 10, 11, 12, 13.
4. Locale sitemap engineering: 14, 15.
5. Multi-tenant public inventory and isolation: 18, 19, 20.
6. Launch, monitoring, and rollback: 21, 22, 23, 24, 25.

A different grouping can pass if every page has a unique outcome, required evidence, and next action. Fail the exercise if two proposed pages could share the same purpose statement.

Final checklist

• The audience, task, conditions, and output are defined.
• Raw first-party language is anonymized and source-labeled.
• Keyword tools are treated as models, not exact demand.
• Trends scores are not presented as volume.
• Primary and supporting entities have meaningful relationships.
• Fan-out branches are labeled as hypotheses where appropriate.
• Current results are observed with locale, device, and date.
• Each claim has an evidence need.
• Every sub-question has a section, page, link, merge, or no-page decision.
• Each new route has one canonical owner.
• Planned pages have distinct audience, outcome, format, or evidence.
• Unsupported and duplicate ideas are removed before writing.

Frequently asked questions

Are entities replacing keywords?

No. Query language still tells you how people express needs. Entity and relationship research helps you explain the subject consistently and completely. Use both.

What is a topical map?

A topical map is a planned set of content owners and their relationships. A useful map records audience, outcome, evidence, page role, canonical route, and internal links. A giant list of keywords is not enough.

How do I know if a keyword needs its own page?

Create a page when it completes a distinct task for a defined audience and can support that task with enough unique evidence or functionality. A wording change alone is not sufficient.

Can I see Google's real fan-out queries?

Google documents the concept and gives examples, but it does not provide a complete hidden-query log for publishers. Build plausible research branches and validate the usefulness of the resulting content.

Should I cover every question an AI might ask?

No. Cover the questions necessary for the human task. Google's 2026 guidance warns against creating separate content for every possible fan-out variation.

Sources and methodology

Current community and search-result research was used to identify practical learner questions around keyword clusters, entities, topical maps, and what to do after finding keywords. Community material informed wording and edge cases, not factual claims.

1. Google Search Central, Top ways to ensure your content performs well in Google's generative AI experiences, updated 10 July 2026. Official documentation for fan-out guidance and the warning against prompt-variation scaling.
2. Google Search Central, AI features and your website, updated 10 December 2025. Official description of AI feature eligibility and query fan-out context.
3. Google Trends, FAQ about Google Trends data, checked 27 July 2026. Official source for sampling, normalization, filtering, noise, and product-scope limitations.
4. Ahrefs, AI Overview triggers, published 10 November 2025. Observational vendor evidence about its desktop keyword database, not universal query demand.
5. Google Search Central, SEO Starter Guide, checked 27 July 2026. Official beginner guidance on useful organization, links, and search-accessible content.

Previous: Search Intent and the New Search Journey
Next: Search Console, GA4, and Generative-AI Measurement