SEO Content Fundamentals: Keywords, Intent and Prompt Research for AI Search in 2026

What Is SEO Content? People-First Pages Built on Research

SEO content is content made for people that search engines and AI answers can find, understand and use, with what to publish chosen by research into what people actually ask. AI answers changed the shape of those questions: longer, more conversational, often split into several searches. They didn't change the fundamentals.

That definition is this guide's stance, and it follows Google's own line. Google's guidance on helpful, reliable, people-first content asks for "people-first content" rather than "search engine-first content made primarily to gain search engine rankings."

One of its warning signs describes the opposite habit exactly: "Are you producing lots of content on many different topics in hopes that some of it might perform well in search results?"

The research behind SEO content is keyword research: finding the queries people use, and what they want when they use them. That second part is search intent, and it decides more about a page than any single keyword does.

Researching or writing with AI tools, or with an agent, is about who does the work, which the Agentic SEO guide covers. SEO content is about what the work is aimed at.

"SEO content strategy" usually means the plan for what to publish. The final section describes how to build one.

The Parts of Content Work, and What AI Answers Change in Each

AI answers change research and production most: questions got longer and split into several searches, and Google now counts mass-produced AI pages that add nothing for users as spam. For other markets and for maintenance, the engines have published nothing specific to AI answers yet.

Part of the work	Question it answers	What AI answers change	Goes deeper
Research and intent	What to publish, for which intent, on which page?	Queries are longer and often split into several searches	Below
Production	How is it written and reviewed?	Google's spam policies count using generative AI "to generate many pages without adding value for users" as scaled content abuse	AI-Assisted SEO Content Production
On-page formatting	How is the page structured and labeled?	Google's guide to generative AI features says to ignore "chunking" content for them	On-Page SEO & AI Answers Optimization
Formats beyond text	Which formats besides text?	People search with voice or images more: over one in six US searches, per Google's AI Mode data (May 2026). Google's AI features "can bring in relevant images and video"	Multimedia SEO & Non-Text Formats
Other languages and markets	Which markets, and how is content adapted?	No AI-specific engine guidance yet	Multilingual & International SEO
Local search	How do people nearby find the business?	Google says AI responses can include local business information, and a Google Business Profile helps a business appear in them	Local SEO for AI Search
Maintenance	When to update, merge or remove a page?	No AI-specific engine guidance yet	Content Freshness & AI-Assisted Maintenance

Our reading: research is the part every other part depends on, because a page can't be written, formatted, localized or refreshed well until it's clear which intent it serves.

What Is Search Intent? The Industry's Four Types vs. Google's Own Model

Search intent is what someone is trying to do when they type or say a query. The SEO industry sorts it into four types: informational, navigational, commercial and transactional.

Those four types are a convention with an origin. Andrei Broder's 2002 paper "A taxonomy of web search" proposed three: navigational, informational and transactional. Commercial, for comparing options before buying, doesn't appear in the paper; it's an addition from SEO practice.

Google uses a different model in its Search Quality Rater Guidelines (September 11, 2025 version). They're guidance for the human raters who evaluate Google's results, not a description of the ranking algorithm, but they are Google's own published model of intent.

Section 12.7 sorts queries into four kinds:

• Know: to "find information or explore a topic". A query is Know Simple "if most people would agree on a correct answer, and it would fit in 1-2 sentences or a short list of items."
• Do: "when the user is trying to accomplish a goal or engage in an activity."
• Website: "looking for a specific website or webpage."
• Visit-in-person: looking for a place to go.

The guidelines also split a query's meanings: the dominant interpretation is "what most users mean", a common one "what many or some users mean", and minor ones what "fewer users have in mind." That's an inference, not Google's statement, but it follows: a keyword carries a spread of intents, not one label.

AI assistants add a third view. OpenAI's researchers, in "How People Use ChatGPT" (a National Bureau of Economic Research working paper, September 2025), estimate that "about 49% of messages are Asking, 40% are Doing, and 11% are Expressing", from ChatGPT messages through July 2025.

Asking is "seeking information or clarification to inform a decision." Doing is when someone wants "to produce some output or perform a particular task", such as a draft, code or a plan. Expressing is "expressing views or feelings but not seeking any information or action."

Four Types, Three Models: Which Search Intents a Content Page Can Win

Side by side, the three models line up only roughly. The alignment below, and the last column, are this guide's reading: none of the sources maps one model onto another.

Industry type	Broder (2002)	Google rater guidelines	ChatGPT use (OpenAI)	Can a content page win it?
Informational	Informational	Know (Know Simple for short answers)	Asking	Yes for Know; Know Simple is often answered in the result itself
Commercial	Not in the paper	Usually Know, sometimes Do	Asking	Yes: comparisons and reviews
Transactional	Transactional	Do	Doing, loosely: Doing means producing output	Product and service pages, not guides
Navigational	Navigational	Website	No match	Only the site being looked for
None (visiting a place)	Not in the paper	Visit-in-person	No match	Local pages and listings, a separate topic
None (a task done in the chat)	Not in the paper	No match	Doing (write, code, plan), Expressing	Rarely: the assistant does the task itself

We'd argue the demand worth mapping is mostly Know and Asking. Doing and Expressing prompts are handled inside the chat, so they rarely belong on a keyword map, and a Know Simple query earns a page only when the page adds what a short answer can't.

Search intent is what someone wants to get done with a query. The industry sorts it into four types, but a single query can carry more than one meaning. Google's guidelines for its human raters separate what most people mean from what many or only a few mean, so research records the main meaning and any common ones.

Keyword Research vs. Prompt Research: Longer Queries, the Same Search Intents

Prompt research looks at the longer, conversational form of the same questions keyword research looks at. Google's guidance points to treating both as one set of intents, not two lists of pages to build.

The terms, used the same way throughout: a query is what someone types or says to a search engine or AI assistant; a keyword is a query as tracked in research; a prompt is a query written as a full request to an AI assistant, usually longer and with context.

Prompt research, then, is finding the prompts people bring to AI assistants on a topic. Vendors that sell prompt tools define it more narrowly: Semrush calls it "identifying and tracking the questions that cause AI systems to compare options and recommend specific brands."

Tracking which prompts cite a site over time is a separate job, which the AI Visibility, Mentions & Prompt Monitoring guide goes deeper on.

What changed is the shape of the question. Google's May 2026 AI Mode data found that "the average AI Mode search is triple the length of a traditional Search query", based on US searches.

One question can also become several searches. Google's AI features use query fan-out, defined in the AI SEO guide. In Google's example, "how to fix a lawn that's full of weeds" fans out to "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn".

That's one question and several keyword-like searches, but arguably still one intent: someone with a weedy lawn who wants it fixed.

What didn't change is the set of intents behind them. Google's guide to generative AI features says "AI systems can understand synonyms and general meanings of what someone is seeking", so "you don't have to worry that you don't have enough 'long-tail' keywords or haven't captured every variation of how someone might seek content like yours."

A long-tail keyword is a longer, more specific query with little search volume on its own. Google's line doesn't make long-tail queries worthless: they still show what people want. It means they're evidence for an intent, not a reason for a page each.

The same guide warns that creating separate content for every variation or fan-out query "primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy", and that "a high quantity of pages doesn't make a website higher quality or more relevant to users."

Scaled content abuse, as Google's spam policies define it, is "when many pages are generated for the primary purpose of manipulating search rankings and not helping users."

Google's own data makes the same point. Query groups, a Search Console Insights feature added in October 2025, lists eight ways to search for guacamole dip, from "guac dip recipe" to "how to make guacamole dip easy", and says: "While these are distinct queries, they reflect a similar user intent."

AI answers changed how people ask, not what they want. Keywords and prompts are the short and long forms of the same questions, so research groups every variant under the intent it serves. A site then needs one strong page per intent it can credibly answer, not one per keyword, variant or prompt.

Is Keyword Research Still Relevant?

Keyword research is still relevant because queries didn't disappear, they got longer. Google and Bing both still show site owners the queries they appear for, as the next section covers.

Where Keyword Data Comes From: First-Party Engine Reports Before Modeled Estimates

Keyword data comes from three places: first-party data, which an engine reports about a site's own appearances; modeled estimates of demand from keyword and prompt tools; and the live results page anyone sees by running a query. Where a site has first-party data, it comes first.

Checked against each source's documentation, September 2026:

Source	What it shows	What it doesn't show	Type
Search Console Performance report	Queries a site already appears for in Google, with clicks, impressions and position. AI Overview and AI Mode impressions are counted within the "Web" search type	Queries where the site never appears; which of a query's impressions came from AI features	First-party
The Performance report's branded queries filter (all eligible sites since March 2026)	Branded queries (containing the site's brand) apart from non-branded	Demand beyond the site's own queries	First-party
Query groups in Search Console Insights (October 2025)	Similar queries grouped by intent	Anything for smaller sites: only for sites with many queries	First-party
Search Console API (application programming interface)	The Performance data, available programmatically	Anything the report itself doesn't	First-party
Search Console generative AI performance report (all sites since August 31, 2026)	A site's appearances in AI Overviews and AI Mode, by page, country, device and date	Which queries triggered them: it has no query dimension	First-party
Bing Webmaster Tools AI Performance (public preview, February 2026), covering Microsoft Copilot, AI summaries in Bing and select partners	Citations, cited pages and grounding queries: "The key phrases the AI used when retrieving content that was referenced in AI-generated answers"	The prompts people actually typed: grounding queries are what the AI searched for	First-party
Google Keyword Planner	"Estimates on the number of searches a keyword gets each month", built for advertisers	Exact counts	Modeled
Google Trends	Interest in a term over time and place, from a sample of Google searches, scaled 0 to 100	Search volume: its numbers are relative	Sampled and relative, closest to modeled
Third-party keyword tools (Semrush, Ahrefs or Moz, for example)	Volumes, related queries, who ranks, and keyword difficulty: each tool's own score for how hard a query is to rank for	Exact counts; each tool models differently	Modeled
Prompt research tools	Prompts and topics people bring to AI assistants	Precise counts: Semrush says prompt research "focuses on direction and pattern recognition, not fixed positions or precise counts"	Modeled
The live results page	Autocomplete, People Also Ask, related searches, ranking page types, the AI answer and its citations, and forum threads such as Reddit's	A stable dataset: it varies by location and account	Live results page

As of September 2026, of the major engines only Bing gives site owners query-level data on AI answers, and even that shows what the AI searched for rather than what people asked.

Search volumes are estimates, not counts. Keyword Planner's own word is "estimates", so volume ranks candidates roughly rather than measuring demand.

The same keyword can swing wildly between database snapshots. One major tool's US database showed "keyword research" at 9,900 monthly searches in June 2025 and 2,740,000 in September 2026, inflated by a two-month spike earlier in the year.

It follows that a query a tool shows at zero volume can still have real demand, especially a long, specific question.

Why first-party data first: it shows what the engines already connect the site to. A new site has none yet, so it starts from the live results page and modeled estimates, and shifts as its own data builds up.

Watching the same reports over time to measure results is a separate job, which SEO Measurement goes deeper on.

From Research to an SEO Content Strategy: Keyword Mapping, One Page per Intent

An SEO content strategy decides what to publish, for which intent, on which page. Its working form is a keyword map: a table that gives each group of queries sharing an intent exactly one page, existing or planned, or a recorded reason to skip it.

When two of a site's pages target the same intent they compete with each other, which is called keyword cannibalization. Preventing it is one of the map's main jobs.

Some of the steps below are data work, with documented inputs and outputs (report exports, grouping, checking live results) that a tool-connected agent can carry out the same way every time. Others are judgment calls that depend on things no report contains, such as what the site can credibly answer and what matters to the business.

Judgment calls need the site's context written down where whoever does the work can read it; Context Engineering for SEO Agents goes deeper on that. Both the split and the step order are a proposed method, not a published one.

1. Collect the Queries the Site Already Gets

• In: Search Console Performance data (from the report, an export or the API), Bing's grounding queries and, where available, the query groups in Search Console Insights.
• How: split branded from non-branded queries with the branded filter.
• Out: a query list with clicks and impressions, branded and non-branded apart, marked first-party. A new site skips to step 2.
• Type: data work.

2. Add Demand the Site Doesn't Get Yet

• In: the topics the site can credibly cover, keyword tools, the live results page, and prompt tools as a sense of direction only.
• How: expand from those topics, not from whatever is trending: the people-first test, applied to research.
• Out: a candidate list, each query marked first-party or modeled.
• Type: both. Choosing the topics is a judgment call; expanding them is data work.

3. Group Variants by Intent

• In: the combined list.
• How: put together the queries one page would fully answer, the way Google's query groups do. A common method treats two queries as one intent when the same pages rank for both; that's common practice, not a documented rule. The industry calls this step keyword clustering.
• Out: intent groups, each with a head query (its main phrasing), variants and prompts.
• Type: data work, with edge cases that need judgment.

4. Check Each Group's Live Results and AI Answer

• In: each group's head query.
• How: run it and note which page types rank, whether an AI Overview or AI Mode answer shows, what it says and which sources it cites.
• Out: per group, the dominant and any common interpretation, the page type that fits (a guide, a comparison, a product page, a video), and whether the result already answers the query in full.
• Type: both. Recording the results is data work; reading the dominant interpretation and the page type that fits is a judgment call.

5. Classify Each Group and Decide What to Publish

• In: the groups and their notes.
• How: label each group's intent (industry type and rater category) and choose one action: target an existing page, plan a new one, or skip it with a reason.
• Reasons to skip: Doing or Expressing demand, a Know Simple answer the result already gives, another brand's navigational query, or a topic outside the site's own expertise.
• Priority: fit with the site's expertise, then first-party evidence, then value to the business, then difficulty, as one vendor's estimate. That order is a judgment, not a documented rule.
• Out: every group has exactly one target page or a reason to skip.
• Type: judgment call.

6. Record the Keyword Map and Check It

• In: the decisions from step 5.
• How: write one row per intent group.
• Out: a map where no two pages target the same group, and every target page is indexable and meets the floor that the Technical SEO guide sets out.
• Gaps and weak pages: groups with no page are the content gaps; groups whose page exists but underperforms go to maintenance.
• Type: data work.

In the example row below, the head query and variants come from Google's guacamole example; everything else is illustrative.

Intent group (head query)	Main variants and prompts	Intent	Dominant interpretation	Target page	Action and reason	Sources, date checked
how to make guacamole dip	guacamole dip recipe, guac dip recipe, easy guacamole dip recipe	Informational / Know	A recipe to follow (example)	A recipe page (example)	New page: none answers it yet (example)	Search Console query groups, live results, September 2026 (example)

The map says which page targets which intent, not how many times a word should appear. Google's SEO Starter Guide notes a page can match many queries "even if you don't explicitly use the exact terms", and says keyword stuffing is against its spam policies.

How mapped pages group and link to each other is the subject of the Site Architecture & Topical Authority for AI guide. Writing each page is production, the second part of content work above.

Running this research on a real site, with its own data and tools, is separate work.

Sources

Sourced claims:

• Creating helpful, reliable, people-first content, Google Search Central. Cited for people-first vs search engine-first content and the warning question about producing lots of content on many topics.
• Spam policies for Google web search, Google Search Central. Cited for the definition of scaled content abuse and its generative AI example.
• Optimizing your website for generative AI features on Google Search, Google Search Central (updated July 10, 2026). Cited for ignoring "chunking", images and video in AI features, local business information and Business Profiles, the lawn fan-out example, the long-tail line and the fan-out warning.
• How AI Mode is changing and expanding the way people search, Google, May 19, 2026. US data, cited for AI Mode searches averaging triple the length and more than one in six US searches using voice or images.
• "A taxonomy of web search", Andrei Broder, SIGIR Forum, 2002. Cited for the three original types: navigational, informational and transactional.
• Search Quality Rater Guidelines, Google, September 11, 2025. Guidance for human raters, cited for Know, Know Simple, Do, Website and Visit-in-person, and dominant, common and minor interpretations.
• "How People Use ChatGPT", NBER Working Paper 34255, September 2025. Cited for the Asking, Doing and Expressing split of ChatGPT messages through July 2025, and the definition of each.
• How to do prompt research for AI SEO, Semrush, August 17, 2026. A vendor that sells a prompt tool, cited for its definition of prompt research and its line that prompt research tracks direction, not precise counts.
• Introducing Query groups in Search Console Insights, Google Search Central Blog, October 27, 2025. Cited for the guacamole example, "distinct queries" reflecting "a similar user intent", and availability only to sites with many queries.
• Performance report, Search Console Help. Cited for the query, click, impression and position data it shows.
• AI features and your website, Google Search Central. Cited for AI Overview and AI Mode traffic being reported within the "Web" search type.
• Branded queries filter, Google Search Central Blog, November 20, 2025. Cited for the branded vs non-branded split, available to all eligible sites from March 11, 2026.
• Search Console API, Google for Developers. Cited for Performance data being available programmatically.
• Generative AI performance report, Search Console Help. Cited for its dimensions (pages, countries, dates, devices) and the absence of a query dimension.
• Introducing AI Performance in Bing Webmaster Tools, Bing Webmaster Blog, February 10, 2026. Cited for the surfaces covered, citations, cited pages and the grounding queries definition.
• Keyword Planner, Google Ads Help. Cited for its volumes being "estimates".
• FAQ about Google Trends data, Google Trends Help. Cited for Trends using a sample and scaling relative interest from 0 to 100.
• SEO Starter Guide, Google Search Central. Cited for pages matching queries without the exact terms, and keyword stuffing being against the spam policies.

Argued, not sourced (this guide's reasoning from the sources above):

• The definition of SEO content as people-first content chosen by research, as this guide's stance built on Google's guidance.
• The core claim: AI answers changed the shape of questions, not the set of intents, so one page per intent rather than per keyword, variant or prompt.
• Research being the part of content work every other part depends on.
• The alignment of Broder's, Google's and OpenAI's models, and which intents a content page can win.
• Commercial intent coming from SEO practice rather than Broder's paper; the paper's absence of it is checkable, its later origin is not documented here.
• A keyword carrying a spread of intents, as an inference from the rater guidelines' interpretations.
• First-party data before modeled estimates, the new-site exception, and zero-volume queries still having demand.
• That only Bing, of the major engines, gives site owners query-level AI answer data as of September 2026, from this guide's review of Google's and Bing's reports.
• The split between data work and judgment calls, which step is which, the step order and the priority order.
• Grouping two queries as one intent when the same pages rank for both, as common practice with no primary source.
• The "keyword research" volume swing, from this guide's check of one major tool's US database.