Local SEO and Research
Keyword Research for AI Search: From Keywords to Questions and Prompts
Keywords haven't disappeared; they've grown longer and more conversational. Good research now captures three layers at once: the short searches, the questions behind them, and the prompts people type into AI assistants.

Keyword research for AI search means finding the searches, questions and conversational prompts your buyers use, tagging each by intent and difficulty, grouping them into clusters that reflect how people actually search, and mapping each cluster to the one page that should win it. The keyword layer still drives rankings; the question and prompt layers shape the answer-first content that AI Overviews and assistants quote.
Keyword research is where every SEO engagement I run begins. I've built plans ranging from 100 keywords in five clusters for a New York localization studio (case study) to 278 keywords across five product families for a European lab supply store (case study) and more than 200 buyer-intent keywords over two rounds for a UK car dealer (case study). This guide is the process behind them.
What changed with AI search
Traditional keyword research optimized for short, typed queries. Three things have changed:
- Queries are longer. People type and speak full questions, especially into AI assistants.
- One question becomes many. AI systems break questions into sub-queries behind the scenes, so coverage of related questions matters.
- Some demand never shows in tools. Many prompts are unique, so they don't appear in keyword databases with volumes.
None of this makes keyword volume useless. It means volume is one input among several, and question research matters as much as head terms.
The three layers
| Layer | Example | Where it comes from | What it feeds |
|---|---|---|---|
| Keyword | "technical SEO audit" | Keyword tools, Search Console | Page targeting, titles, rankings |
| Question | "how long does a technical SEO audit take" | People Also Ask, Search Console, tools' question reports | Answer blocks, FAQs, snippets |
| Prompt | "I run a small e-commerce store and my traffic dropped after a redesign, what should I check first?" | Sales conversations, AI follow-ups, customer emails | Content depth, sub-topic coverage, AI visibility tracking |
Good research captures all three and connects them to the same pages.
Step 1: understand the business first
Before opening a keyword tool, I answer:
- What exactly do you sell, and which offers matter most commercially?
- Who buys, and what triggers them to search?
- Where are your buyers: which countries, cities or languages?
- What does a good lead look like?
This prevents the most common failure: a long list of popular keywords that don't match what the business actually sells or where it sells.
Step 2: build the seed list
Seeds are the core terms for each product or service. Sources:
- Your own service and product names, and the words customers use for them.
- Competitor pages and their titles.
- Search Console queries you already get impressions for.
- Industry terms, including technical and specification language.
For the lab supply store, seeds were product families: petri dishes, rack towers, tips in rack, micropipette tips and cuvettes. For the UK dealer, they started with Land Rover Defender models, prestige cars and commercial vans.
Step 3: expand with tools and competitors
Feed the seeds into a keyword tool such as Semrush, using the database for your target market. Volumes and difficulty differ widely between the US, UK, European countries and Bangladesh.
Then look at competitors:
- Keyword gap: terms competitors rank for that you don't.
- Missing keywords: terms where you rank poorly and they rank well.
- Paid keywords: what competitors bid on, a strong sign of commercial value.
A competitor analysis often reveals openings. In the UK dealer project, only 2 of 5 competitors ran Google Ads, meaning most of the market competed on organic search alone.
Step 4: collect questions and prompts
For each seed, gather:
- Question reports from your keyword tool.
- People Also Ask questions for the main terms.
- AI Mode follow-ups and assistant suggestions.
- Your own conversations: sales emails, WhatsApp chats, call notes.
- Forum and community questions in your niche.
For the NYC studio, this became a dedicated AEO cluster of 20 question-style queries, written later as FAQ and AI-answer content. My AEO guide shows how those questions become answer blocks.
Step 5: tag intent and difficulty
Every keyword gets two tags.
Intent
| Intent | Meaning | Example |
|---|---|---|
| Informational | Wants to learn | "what is a cuvette used for" |
| Commercial | Comparing options | "best petri dishes for bacterial culture" |
| Transactional | Ready to buy or contact | "petri dishes for sale" |
| Navigational | Looking for a specific brand | "[brand] login" |
Difficulty
Keyword difficulty scores estimate how hard it is to rank on page one, usually on a 0 to 100 scale. For newer or smaller sites, I focus on lower difficulty first. In the UK dealer project, 79 of the 111 round-one keywords had difficulty 20 or below, realistic targets for a site starting out.
Step 6: cluster the way buyers search
Clustering is where research becomes strategy. Group keywords by the dimensions buyers use, not by alphabet or volume.
For the lab supply store, each product family was split into the same buying dimensions:
| Product family | Clusters |
|---|---|
| Petri dishes | General, size, vent type, use case, long-tail |
| Rack towers | General, brand, tube size, use case, long-tail |
| Tips in rack | General, brand compatibility, use case |
| Micropipette tips | General, brand compatibility, use case |
| Cuvettes | General, material and type, use case |
That produced 19 clusters that map directly to category pages, filters, use-case landing pages and compatibility pages.
Step 7: map every cluster to one page
A keyword list becomes useful only when each cluster has one owner page. The map has columns for:

- Cluster name
- Primary keyword and secondary keywords
- Questions to answer on the page
- Intent
- Owner URL (existing or to be created)
- Priority
Rules I follow:
- One cluster, one page. If two pages target the same cluster, merge them or differentiate them clearly.
- Transactional clusters go to service, product or category pages.
- Informational clusters go to guides and articles, linking to the relevant commercial page.
- Questions go on the page that owns the topic, as answer blocks or FAQs.
Step 8: prioritize
I score clusters on three factors:
- Business value: how close the searcher is to buying, and how valuable the sale is.
- Winnability: difficulty and the strength of current results.
- Effort: whether the page exists and needs improving, or must be created.
Quick wins are high-value, low-difficulty clusters where a page already exists. Those come first.
Why low-volume keywords often win
Volume tempts everyone toward broad head terms. In niche and B2B markets, that's usually a mistake.
For the lab supply store, 97% of 278 keywords had difficulty 10 or below. Searches like "petri dishes for sale" (200 a month, difficulty 5) or "best petri dishes for bacterial culture" (100 a month, difficulty 0) bring buyers ready to order, and almost no competition means focused pages can rank quickly.
For the UK dealer, examples like "land rover defender 90 for sale" (2,400 a month, difficulty 16) and "affordable prestige cars" (1,600 a month, difficulty 14) combined real demand with realistic difficulty. Across both rounds, the research mapped about 230K combined monthly search volume.
Local keyword research
For businesses serving specific areas, location changes everything:
- Explicit local searches include the place: "roofer in Fort Lauderdale", "SEO consultant Dhaka".
- Implicit local searches don't: "roof repair" or "dentist" already show local results based on the searcher's location.
- Near-me searches reflect immediate intent and are mostly won through the Business Profile.
My approach: research the service terms without locations first, then add the main cities and areas you serve, and check which combinations show local map results. Service pages target the service plus main area; area pages are worth creating only where you have genuinely distinct things to say, such as local projects, regulations or conditions. Thin pages that just swap the city name rarely help. The Google Business Profile guide covers the map side.
Bangla, English and mixed searches
In Bangladesh, people search in English, Bangla and a mix of both, often typing Bangla words in Latin script. Keyword tools capture English best, so research needs extra steps:
- Collect real phrasing from customer messages, Facebook comments and calls.
- Check both scripts for key terms in Search Console and in Google's autocomplete.
- Decide which language each page serves. Mixing languages on one page usually works worse than separate, clearly targeted pages.
More on this is in SEO in Bangladesh.
What the deliverable looks like
Every keyword research project I deliver ends in one spreadsheet with these tabs:
- Summary: the clusters, their total volume, priority and owner pages.
- Keyword map: every keyword with volume, difficulty, intent, cluster and target URL.
- Questions: the question and prompt layer, grouped by owner page.
- Quick wins: high-value, low-difficulty keywords where a page exists or is easy to create.
- Competitor gaps: terms competitors win that you don't.
The goal is a document someone can act on the next morning, not a dump of every keyword a tool can produce.
From research to AI visibility
The question and prompt layers do double duty:
- They become answer blocks and FAQs on the owner pages.
- They become the prompt set you track monthly to measure AI visibility, as described in measuring AI search visibility.
When you can show a client the prompts you're targeting, the pages that answer them and whether assistants cite those pages, keyword research stops being a spreadsheet and becomes a strategy.
Common mistakes
- Researching the wrong market database, such as US volumes for a business that sells in Bangladesh.
- Chasing volume instead of intent and winnability.
- No map, so several pages compete for the same terms.
- Ignoring questions, leaving snippets and AI answers to competitors.
- Treating research as one-off. Markets change; I revisit maps quarterly, and some clients come back for another round, as the UK dealer did.
A quick way to start if you have no tools
If you don't have access to a paid keyword tool, you can still build a useful first map:
- List every service or product you sell, using the words customers use.
- Type each into Google and note the autocomplete suggestions and People Also Ask questions.
- Check Search Console for queries you already appear for.
- Collect the questions from your last 50 customer messages.
- Group everything by the page that should answer it.
It won't give you volumes or difficulty, but it will give you real demand and a clear page plan. Add tool data later to prioritize.
How research connects to everything else
Keyword research isn't a separate project. It feeds the content plan, the page structure, the internal links between pages, the questions in your FAQs and the prompts you track for AI visibility. When I start a new client engagement, the keyword map becomes the reference document that every later decision points back to.
Seasonality and trends
Volumes in keyword tools are usually averages. Many markets are seasonal: roofing searches rise after storms, education searches peak before admission seasons, and travel searches follow holidays. Check the monthly trend for your main clusters and plan content to be published and indexed a few weeks before demand rises. For fast-moving topics, such as AI search itself, revisit the research every quarter, because the questions people ask change faster than the tools' databases update.
Frequently asked questions
How many keywords do I need for my website?
Enough to cover every page you want to rank, not a fixed number. Past projects ranged from 100 keywords in five clusters for a service business to 278 across five product families for an e-commerce store. Clusters and pages matter more than the total count.
Are low-volume keywords worth targeting?
Often yes, especially in B2B and niche markets. Low-volume searches with clear buying intent and low difficulty can rank quickly and bring customers ready to buy. Many prompts typed into AI assistants also have no measurable volume but still represent real demand.
How do I find keywords for AI search?
Combine keyword tools with question research: People Also Ask, AI Mode follow-ups, Search Console question queries and your own customer conversations. Write the questions as headings with direct answers, and track the most important ones as prompts across AI assistants.
What is keyword mapping?
Keyword mapping assigns each group of related keywords to the single page on your site that should rank for them. It prevents pages from competing with each other and shows which pages need improving and which need to be created.
Which country database should I use?
Use the database for the market you sell in. Search volume and difficulty differ widely between countries, so a business selling in the UK should use UK data, and a business in Bangladesh should check Bangladesh data alongside any international markets it targets.