Keyword and AI Prompt Research

Keyword and AI prompt research turns observed demand into page decisions. It ends when each retained opportunity has an intent, post type, target page, priority, and supporting evidence.

Phase: P6, Stage B — Decide. Timebox: three to five working days for one site, one language, and up to two countries; add time for each different language, catalog, or audience. Owner: the SEO strategist, with input from content, product or sales, and analytics owners.

Why this phase comes here

People express demand on two different surfaces. A keyword is a phrase typed into a search engine, where demand can be observed through impressions, clicks, and positions. An AI prompt is the fuller instruction or question given to an assistant, often including audience, constraints, location, or comparison criteria. They can share a topic without being interchangeable.

For example, expense management software is a keyword. What expense management software works for a 60-person company with European subsidiaries and approval workflows? is a prompt. The prompt exposes company size, geography, requirements, and the decision to support. A modest-volume term can represent a valuable purchase, shortlist, or implementation prompt; a high-volume informational term may have little business value.

This phase comes after access, tracking, and data sources and baseline measurement because they establish what the client’s properties already earn. Starting before Google Search Console, Bing Webmaster Tools, analytics, and internal search are usable can miss customer language, split one page across targets, or recommend new content where an existing page is close to winning.

Discovery supplies vocabulary and business boundaries. The baseline supplies query, page, country, and performance evidence. This phase combines both with controlled competitor and prompt discovery.

Inputs and outputs

Inputs and outputs form a contract. Inputs need a known scope, date range, market, and owner. Outputs must let the next operator make a page decision without repeating the research.

Ready to Monitor Your AI Visibility?

Track how AI chatbots mention your brand across ChatGPT, Perplexity, and other platforms.

Get Started Contact

The checklist

Follow the sequence. Skipping directly to scoring gives precise-looking numbers to an incomplete or duplicated set.

1. Establish seeds from discovery

• What to do: Turn the discovery document into seed concepts: problems, solutions, categories, products, features, use cases, audiences, industries, locations, objections, alternatives, and branded terms.
• Why it matters: Seeds keep expansion inside the business scope. Without them, tools return adjacent demand the business cannot serve or does not want.
• How to do it: Preserve the customer’s phrasing, add a normalized label, and mark every seed included, excluded, or uncertain. Do not silently treat internal product language as customer language.
• Tool: Discovery document and a working research table.
• Done when: Every priority offer, audience, and country has at least one seed; every exclusion has a reason; the product or sales owner has reviewed the set.

2. Expand from first-party query data

• What to do: Pull actual queries from Google Search Console, Bing Webmaster Tools, and internal site search before using market estimates.
• Why it matters: These sources show how the site is already being found and which page search engines associate with a need. They reveal language a brainstorm often misses.
• How to do it: Use at least the latest 90 complete days and compare with the preceding period when seasonality or a migration could distort the view. Export query-to-page relationships, separate brand from non-brand, and segment material countries. Keep zero-click impressions as demand evidence.
• Tool: AmICited Google and Bing query reports, plus the client’s internal-search export.
• Done when: Each imported row has source, window, market, landing page where available, and core metrics; the totals reconcile with the selected source filters; missing sources are documented.

3. Add a controlled competitor expansion

• What to do: Add terms and answer patterns competitors win that the first-party set does not contain.
• Why it matters: First-party data cannot show demand for topics on which the site has no visibility. A controlled gap pass prevents the current footprint from defining the entire future plan.
• How to do it: Sample at least three relevant domains when the market supports them. Separate commercial rivals from publishers, directories, forums, or vendors that merely compete for attention. Capture the query or prompt, winning page, result type, and relevance. Leave exhaustive analysis to competitor and gap analysis .
• Tool: Search results, AI answers, and the AmICited Semantic Map.
• Done when: Every added opportunity names the competitor evidence and business fit; irrelevant adjacency is excluded; the detailed gap backlog is ready for the next phase.

4. Discover prompts as prompts

• What to do: Build natural questions and instructions that could trigger an AI answer about the category.
• Why it matters: Prompts expose context that keyword rows flatten. Treating them as long-tail keywords loses audience, constraints, comparison frame, and requested outcome—the parts that often determine which brands and sources appear.
• How to do it: Start from sales and support questions, internal search, query clusters, and real AI answers. Expand material topics across definitions, recommendations, comparisons, alternatives, troubleshooting, implementation, risk, cost, and suitability. Use URL suggestions for discovery, not as a substitute for editorial review. Keep variants only when they change the expected answer.
• Tool: AmICited Prompts and Semantic Map. Follow the guide to generate prompts automatically from a URL where an existing page is a useful seed.
• Done when: Each priority topic has prompts spanning its material intents, audiences, and decision stages; each prompt reads like a real request; near-duplicates with the same expected answer are merged.

5. Select engines, countries, and a trackable sample

• What to do: Define where each prompt will be run and which prompts deserve ongoing tracking.
• Why it matters: AI answers vary by engine, market, and language. Mixing them without labels makes changes impossible to interpret, while tracking every wording variant wastes runs and creates noisy reporting.
• How to do it: Start with 40–100 core prompts for each distinct country-language combination. Cover every high-value intent and journey stage before adding variants. Track three to five engines when they matter to the audience; use a smaller set only when discovery evidence shows the audience is concentrated. Do not copy one country’s list and translate it mechanically—validate products, currency, regulation, and local phrasing. Use the guide to choose which AI engines to track .
• Tool: AmICited Prompts, with country, provider, tag, and schedule settings.
• Done when: Every tracked prompt has country, language, engine set, intent tag, topic tag, and schedule; every priority intent has at least three non-duplicate prompts in the initial set; the total stays within the approved tracking budget.

6. Cluster by shared need and expected answer

• What to do: Group keywords and prompts that can be satisfied by the same page and separate those that need different answers.
• Why it matters: A cluster is a publishing decision, not merely lexical similarity. Poor clustering either creates multiple pages competing for one need or forces incompatible intents into one page.
• How to do it: Compare meaning, intent, result overlap, AI answer pattern, audience, and evidence. Preserve the original wording and assign a stable cluster ID. Split when the desired action, answer format, audience, or winning page type changes.
• Tool: AmICited Semantic Map plus manual result and answer review.
• Done when: Every retained row has one cluster ID; no cluster mixes incompatible primary intents; a reviewer can explain in one sentence why one page could satisfy the group.

7. Classify intent, then assign a post type

• What to do: Give every row a primary search intent —informational, navigational, commercial investigation, or transactional—plus a more specific job such as define, learn, compare, choose, buy, implement, or prove. Then select the page shape that best completes that job.
• Why it matters: Intent classification is the output that joins research to production. A keyword list without post-type assignment pushes the hardest decision downstream to the brief writer, who then has to infer whether the page should define, compare, guide, sell, or prove.
• How to do it: Inspect search results and AI answers instead of classifying from wording alone. Map definitions to glossary or what-is pages; tasks to how-to guides; broad learning to ultimate guides; two options to comparisons; shortlists to best-X-for-Y or alternatives pages; and commercial needs to product, category, use-case, or proof pages. Use the post-types library as the allowed set.
• Tool: Search and AI answer review, research table, and playbook post-type specifications.
• Done when: Every retained cluster has one primary intent, one assigned post type, and a written reason; mixed-intent clusters are split or have an explicit primary/secondary relationship.

8. Choose the target page and score the opportunity

• What to do: Decide whether each cluster belongs on an existing page or needs a new page, then prioritize it using business value and attainability rather than volume alone.
• Why it matters: The target-page decision prevents duplicate production. A transparent score makes a low-volume, high-value decision prompt capable of outranking a high-volume topic with weak commercial fit.
• How to do it: Score five dimensions from 0 to 3: business value, intent proximity to a meaningful action, evidence of demand across keyword or prompt surfaces, attainability, and strategic coverage. Total the dimensions out of 15 and retain the component scores. Review striking-distance opportunities first because an existing relevant page near the selected target band can be faster to improve than a net-new page. Never let a modelled score override a legal, reputational, technical, or offer-fit constraint.
• Tool: AmICited Unified Keywords and Striking Distance, joined with conversion and business evidence.
• Done when: Every cluster has existing/new, canonical target, component scores, total, priority tier, rationale, and owner; two clusters do not target the same page unless their relationship is intentional.

Tools in AmICited

Save each view’s date range, country, and filters so another operator can reproduce the evidence.

1. Open Google Search Queries at AmICited Reports → Google Search → Queries . Sort by impressions to find observed demand, inspect weak click-through rates, and preserve the query-to-page relationship.

2. Open Bing Queries at AmICited Reports → Bing Webmasters → Queries . Capture Bing-specific demand and the pages Bing associates with each query.

3. Open Unified Keywords at AmICited Reports → Keywords . Reconcile the keyword set across connected organic and paid sources instead of treating each source as a separate plan.

4. Open Striking Distance at AmICited Reports → Striking Distance . Set the position band, minimum impressions, and target position explicitly, then review opportunities grouped by the page that already owns them.

5. Open the Semantic Map at AmICited Semantic Map . Review prompts, fan-out queries, cited pages, and competitor coverage by meaning to challenge clusters and expose missing questions.

6. Open Prompt Tracking at AmICited Prompts . Add the approved representative set, apply country and tags, choose providers and schedule, then confirm the prompts are active.

Stay Updated on AI Visibility Trends

Get the latest insights on AI mentions, brand monitoring, and optimization strategies.

Email address

Subscribe

Decision rules

Thresholds exist to force consistent decisions, not to imply universal market facts. Record any deliberate exception beside the row.

Search volume remains evidence, but it is neither a business-value score nor a prompt-frequency measure. Discuss the opportunity score dimension by dimension; an unexplained total is not actionable.

Deliverable

Hand over one classified opportunity list as a spreadsheet or database export, plus a one-page decision summary. Use one row per keyword or prompt and a stable cluster ID to preserve many-to-one relationships.

The minimum columns are:

The summary names the date range, markets, connected and missing sources, prompt sample size, engines, scoring model, priority totals, unresolved conflicts, and decisions the next phase must validate.

What goes wrong

Ranking by volume. Volume rewards broad awareness even when the offer cannot satisfy the searcher. It also undervalues specialized prompts that express a serious decision. Keep volume as one demand signal and score business value separately.

Treating prompts as long-tail keywords. A long-tail keyword is simply a more specific search phrase. A prompt can contain role, history, constraints, requested output, and follow-up context. Removing those details to deduplicate the list can change the answer being measured.

Leaving post type for the brief writer. This defers the central intent decision until delivery pressure is higher. Missing assignments block the handoff.

Starting with a third-party database. External tools discover absent topics but cannot reveal all language already earning impressions. First-party expansion comes first unless the site has no history.

Tracking every paraphrase. Repeated variants can dominate averages while audiences or journey stages remain uncovered. Build a coverage matrix before increasing the count.

Combining countries too early. The same English phrase can imply different products, currencies, laws, availability, and competitors by market. Keep country and language fields explicit through clustering and scoring.

Calling every near-ranking term a quick win. The page must be relevant, query ownership stable, impressions meaningful, and the proposed change plausible. Position alone is insufficient.

Handoff to competitor and gap analysis

The next phase receives the opportunity list, source register, tracked prompts, target-page decisions, scores, and conflict log. It tests whether competitor coverage, evidence, authority, formats, and citations make each opportunity defensible.

The handoff is accepted when:

• every retained cluster has an intent, post type, target page, priority, and owner;
• keyword and prompt evidence remain distinguishable but joined by cluster;
• markets, engines, windows, and missing data are explicit;
• provisional competitor findings link to the exact result, answer, domain, or page observed;
• cannibalization questions and scoring overrides are visible; and
• the next-phase owner can select a priority cluster and reproduce why it was selected.

That is the join between research and production in the wider SEO process : evidence becomes a classified page opportunity before anyone writes a brief.

FAQ

How many AI prompts should we track?

Start with 40–100 decision-relevant prompts for each distinct country and language combination. Expand only when a new prompt adds an intent, audience, use case, comparison, objection, or journey stage that the current set does not represent.

Can search volume be used to prioritize AI prompts?

Not by itself. Search volume measures keyword demand in a search engine, not how often people ask an AI assistant a prompt. Prioritize prompts using business value, decision relevance, observed answer behavior, and coverage gaps.

Should every keyword become a tracked prompt?

No. Track a representative prompt set that covers material intents and phrasings. Near-identical keywords can belong to one cluster, while prompts should preserve meaningful differences in audience, constraints, and requested outcome.

What is a striking-distance keyword?

In this process, it is a query already ranking close enough to the selected target band that improving its existing page is likely to be cheaper and faster than creating a new page. The band is set explicitly for the site rather than assumed universally.

When is keyword and prompt research complete?

It is complete when every retained term or prompt has an intent, cluster, assigned post type, target page decision, priority, evidence source, market, and owner, and unresolved conflicts are recorded for the next phase.