If you already have a keyword list, you have most of what you need for keyword research for AEO. The work is taking those short terms and turning them into the full questions people ask AI engines, then grouping those questions so you know which pages to write, update, or leave alone. This guide is for marketing leads who want a documented list of question opportunities built from the keywords they already track. By the end you will have clusters of questions, a note on where each phrase came from, and a decision for each cluster.
You will need your current keyword list, access to Google Search Console if your site has it, and a place to keep a working spreadsheet.
One thing to settle early. Nobody outside the AI companies can see how often a given question is typed into ChatGPT, Gemini, or Perplexity, and no source I found establishes a reliable volume number for it. So this process uses search data as evidence of interest, treats every invented question as a guess, and keeps the two labeled separately. That small habit is what makes the final list something you can defend.
Step 1: Set the research brief before you expand anything
Start by writing down who the questions are for, what decision the research should support, and which market and language you are covering. For a marketing lead, the decision is usually one of three things: create a new page, improve an existing page, or put the topic aside for now. Also note any product names or claims that must stay accurate, so nobody makes up a feature while brainstorming.
Keep the scope narrow. This research finds and organizes questions, and it is not about building the list of prompts you track in a monitoring tool, or how often you run that list. Those are separate jobs, and mixing them in at this stage makes the research harder to finish.
How to tell it is done: another person could read your brief and say whom the questions are for, what decision they support, and which keyword list and market you started from.
Common mistake: jumping straight to generic question templates before you know which audience or buying situation matters. You end up with a lot of plausible questions and no clear reason to write any of them.
Step 2: Clean and preserve your seed list
Your seed keywords are the terms you start from. Copy them into a spreadsheet exactly as they appear in the source, and record where each one came from. You can fix obvious formatting differences for analysis, but keep any qualifiers, variants, and separate intents. If you delete a variant because it looks like a duplicate, you may lose a real question.
A simple first-pass record has these columns.
| Field | What to record |
|---|---|
| Seed keyword | The original term, unchanged |
| Source | Existing keyword research, Search Console, customer research, or another documented source |
| Market/context | Geography, language, product line, or audience if relevant |
| Existing page | URL or page title if you know it, and do not assume one exists |
| Initial topic | A provisional subject label, not a final cluster |
| Notes | Ambiguity, variants, brand terms, or other context |
How to tell it is done: you can trace every seed back to its source, and you can read each one without guessing what it meant in the original list.
Common mistake: merging phrases that only look alike. "Email automation pricing" and "how to automate email" share a topic, but they need different answers, so they should stay as separate rows for now.
Step 3: Gather real search and customer-language evidence
Good AEO question research starts with the questions that already exist, before you invent any. Each source below tells you something different, and none of them tells you the whole story, so use them together and save the exact phrase and where it came from.
Google Search Console. The query data shows what people searched to reach your site. Google documents a way to group similar queries with a regex filter. In the Performance report for Search results, you add a filter, choose Queries, select Custom (regex), and enter a pattern. A pattern with alternatives such as how|what|why|when|where pulls out the queries that start with question words. Read the strings it returns instead of assuming the pattern caught everything relevant. Search Console also says it may leave out very rare or anonymized queries and that it stores top rows, not every row, so a phrase missing from the table does not prove nobody searches for it. If you want Search Console and AI visibility data side by side, there is a comparison of AEO tools that connect to Search Console. To see how snippet and overview queries show up in that data, read how to analyze featured snippet and AI Overview performance.
Google Keyword Planner. The average monthly searches figure is a rough signal of Google demand. It depends on the date range, location, and network settings you pick, it includes close variants, and it is rounded. It is not a count of AI prompts. Its competition column describes advertiser competition, not how hard it is to rank in organic results.
Google Trends. Trends shows relative interest across terms, time, and places, scaled from 0 to 100. It is not absolute volume, and a low-volume term can show as zero. A zero does not mean nobody asks the question.
Autocomplete and People Also Ask. Google describes autocomplete as predictions based on real searches that help people finish a query. They are useful for spotting phrasing and related angles, though not every prediction is helpful. People Also Ask only appears for some searches, so a missing box is not evidence of anything. Treat both as clues to follow up on, not as measured AI prompts.
Customer language. Support tickets, sales questions, and interviews are the closest thing you have to how your audience really talks. Keep the source and context with each phrase, and do not turn one customer's wording into a claim about how common it is.
Add two columns to your working file, one for the exact phrase and one for the evidence type. Label each row as an observed search query, a Google suggestion, a People Also Ask question, customer wording, or a question you wrote yourself.
How to tell it is done: every phrase on the sheet can be traced to a source, and you can tell at a glance which ones are observed and which ones are ideas.
Pro tip: if you want to go beyond your own site's data, DeepSmith's Content Map turns your site and your competitors' sites into one map of topics, so you can see what each of you covers. It works from the pages that exist, which makes it a good input to this step. It does not tell you how many people ask a given question of an AI engine, and it should not be described that way.
Step 4: Expand each seed across intent and context
Now you can build out the questions, which is the part of answer engine keyword research that differs most from a normal keyword pass. For each seed, ask what the person is trying to get done, then write variants across the different jobs they might have. These are the dimensions that tend to matter.
- Learn or define: What is it, and what does it do?
- Do or implement: How do I do it, and what are the steps?
- Choose or compare: Which option fits a specific use case, and how do the alternatives differ?
- Troubleshoot: Why is it failing, and how do I fix this specific problem?
- Evaluate: What does it cost, and what are the requirements, risks, or trade-offs?
- Context: Who is asking, and which product, market, company size, tool, or stage of the decision are they in?
Keep the wording natural. A short keyword gets expanded by adding a real goal or constraint, not by putting a question word in front of every term. AI engines sometimes break one request into several smaller searches. Google describes this for AI Mode as query fan-out across related searches and subtopics, and OpenAI says ChatGPT Search may rewrite a request into one or more targeted searches. That is a good reason to cover the separate sub-needs a complete answer would have. It is not a reason to say those sub-searches are what users actually typed, because you can't see that.
Here is an illustration of the method, and it is made up for the example, not research data. Say the seed is "email marketing automation." Variants might be "What is email marketing automation?", "How do I automate a welcome email sequence?", "How do I choose an email automation tool for a small online store?", and "Why are my automated emails not converting?"
The Pew Research Center looked at Google searches from March 2025 and found that longer and question-style searches were more likely to produce an AI summary. In that study, 8% of searches with one or two words produced one, compared with 53% of searches with ten or more words, and 60% of searches that began with a question word such as who, what, when, or why did. It is a reasonable argument for including full questions in your research. It is one study of Google results at one point in time, so it does not show that question wording causes citations, and it says nothing about what share of AI prompts are questions.
You can also use an AI assistant to speed up the brainstorming. Ask it for the likely audience and decision for each seed, and for question variants across learning, implementation, comparison, evaluation, and troubleshooting. Tell it to label every question as researcher-created, and not to invent search volume or claim a question was sent to an AI engine. What comes back is a brainstorm, not evidence, so keep your evidence table from Step 3 next to it and never mix the two.
How to tell it is done: each seed has a manageable set of questions that cover different needs, and the label on each one says whether the wording came from a source or from you.
Common mistake: calling every expanded question a real prompt. A useful guess is not the same thing as a query seen in Search Console, a question a customer asked, or a prompt someone sent to an AI engine.
Step 5: Group questions by the answer they need
Once you have a long list, group it. The test is simple. Put questions together when one page could answer them all well, and split them when the intent, the decision, the audience, or the evidence you need changes enough that a single page would feel unfocused. Keep a link from every question back to its seed keyword so you can trace it.
Matching on shared words is the usual trap. Two questions about the same product can need separate pages if one asks how to set it up and the other asks which vendor to pick. A practical table for this step looks like this.
| Field | Purpose |
|---|---|
| Topic/cluster | The shared subject and user need |
| Main question | The clearest representative question |
| Related questions | Close variants and useful sub-questions |
| Intent | Informational, implementation, comparison, troubleshooting, or another plain label |
| Audience/context | Who is asking and under what constraint |
| Funnel stage | Awareness, consideration, decision, or not yet clear |
| Phrase provenance | Search query, autocomplete, People Also Ask, customer wording, or constructed |
| Existing coverage | A relevant page, no match found, or not checked |
| Proposed action | Improve, consolidate, create, investigate, or defer |
| Confidence/notes | Strength of the evidence and open questions |
Check whether each cluster already maps to a page. Search Console shows query and page performance, and its Pages view groups data by page.
How to tell it is done: every question belongs to a cluster or is marked unresolved, and you can say in a sentence why the grouped questions would get the same answer.
Common mistake: grouping by word overlap alone. If you can't explain why two questions would get the same answer, they probably belong in different clusters.
Step 6: Check topic coverage and gaps
Now compare each cluster with what you have already published. For each one, you are looking for a missing answer, a page that only partly covers the need, a few overlapping pages that should be merged, or a question that does not deserve content at all. Your own site search and Search Console are the simplest tools. Competitor pages can give context, but copying them is not evidence that it suits your readers. An AEO audit tool can also score the pages you already have, which saves some of the manual checking.
This is where DeepSmith's Content Map earns its place. It crawls your site and your competitors' sites onto one shared set of topics and funnel stages, then shows coverage gaps, where you publish but a competitor publishes more, and untapped topics, where a competitor has pages and you have none. It also shows page counts and funnel distribution per topic, so you can tell whether a topic is thin or missing its decision-stage pages.
For help with the checking itself, a sibling how-to on running a GEO check on a page covers what to look at on a page you already have.
How to tell it is done: each candidate has an explicit coverage decision. An existing page can be improved, several pages need merging, a new page is justified, you need more evidence, or no content action is warranted.
Common mistake: turning every keyword variant into its own page. You end up with repetitive pages instead of one useful answer. Google's guidance on helpful content says there is no ideal page length, so size the page to what the reader needs, not to a word-count formula.
Step 7: Prioritize opportunities without inventing prompt volume
You will probably have more clusters than you can write, so you need a way to rank them. A simple qualitative review works well. For each cluster, look at these six things.
- Audience and business fit. Does the question reflect a real audience need and a relevant product or decision?
- Evidence strength. Did the wording come from a search query or a customer, from a platform suggestion, or from you?
- Search demand signal. Are there Search Console or Keyword Planner numbers? Note the source, location, period, and limits, and never relabel Google volume as AI prompt volume.
- Coverage gap. Is the answer missing, thin, out of date, or duplicated on your site?
- Ability to add value. Can you offer specifics, first-hand experience, original data, or reliable evidence that other pages lack?
- Production feasibility. Can your team write an accurate answer and keep it current?
Score each one high, medium, low, or unknown, with a short reason. If your team prefers a numeric score, call it an internal rubric, define the inputs, and don't present it as an industry benchmark.
On the volume question, the honest answer is that it is unknown. If you come across a third-party "AI prompt volume" number, look at where it came from, what sample it used, which engines it covers, and how it was calculated. If you can't find those, call the figure an estimate or mark it unknown. There is a longer look at this in whether anyone can know AI prompt volume. Google also says the usual SEO basics still apply to its AI features and that no special optimizations are required to appear in AI Overviews or AI Mode. So don't promise anyone that a question-shaped page, a bit of markup, or this process guarantees a citation.
DeepSmith's Opportunity Agents can help move from the ranked list to a plan. They read your AI visibility data or your Content Map and return content ideas, and each idea carries the data point that justifies it, so you can defend the backlog instead of guessing. They are a way to attach evidence to ideas, not a replacement for checking the source and intent of each question yourself. After you act on a cluster, an AEO reporting dashboard for marketing leads is one way to see whether it moved anything.
How to tell it is done: the opportunities you picked each have a stated reason, an evidence label, and a note on how sure you are. Nothing in the ranking depends on a made-up AI prompt count.
Common mistake: treating a vendor's prompt volume number as verified because it is a clean number. A precise figure with an unknown method is less useful than an honest "unknown."
Step 8: Turn the chosen clusters into content briefs
For each cluster you decided to act on, write a short brief so the writer doesn't have to redo the research. A good brief names the main reader question and the decision behind it, the seed keyword and secondary phrases with where each came from, the related questions the page should answer, and the ones that need their own pages. It also gives the audience and buyer stage, the action (update, consolidate, create, or defer), the evidence and its limits, and a coverage check against what you already publish so the new page adds something.
Put the direct answer high on the page and follow it with detail. Clear headings and a plain answer at the start of each section also make a page easier for people and for AI engines to use. Tools built for this are compared in best answer and snippet optimization tools. Some teams also run drafts through a content optimizer, and it helps to know whether content optimizers help with AEO before you lean on one.
Once the research decisions are made, you can hand the brief to production. DeepSmith's Content Studio takes a planned idea through the Writer, which produces a researched, linked, brand-grounded article with a cover image and publish-ready metadata, and you review it in Produced Content before it goes out. Content Studio works from the idea and brief you give it. It does not find private AI prompts from a keyword list, and the research above is still what makes the brief worth writing.
How to tell it is done: a writer can explain the reader, the answer, the evidence, the page action, and the limits without doing more keyword research or guessing which phrases were really observed.
Common mistake: handing over a bare list of questions. Without the intent, the evidence labels, and the page decision, the writer has to guess, and the guesses end up on the page.
What to do next
Pick one seed keyword and run it through all eight steps this week. A single seed, taken through this keyword research for AEO routine, will show you where your sources are thin and what your columns need before you scale up. Once you have a few clusters with decisions attached, you can move into setting up the prompts you will track, which is its own task and belongs to a different guide. If you are comparing tools for that work, a roundup of AI visibility tools is a good place to start.
If you want one place to see your AI visibility, map your topic coverage, and turn gaps into articles, you can start a free DeepSmith trial and try it on your own keywords.
Related guides
- What Your CMS Needs to Support for AEO: Evaluate whether your CMS supports AEO by checking structured data and schema control, clean semantic markup, fast updates, redirect and metadata handling, and what to ask vendors before.
- Self-Serve vs Managed AEO: Which Fits Your Team?: Choosing between self serve AEO tools and a managed AEO service?
- What Are Agentic Marketing Platforms for SEO and AEO?: Tell an agentic marketing platform from a writing tool by how it plans, acts, checks and adapts.
- White-Label AEO Reporting: What Agencies Should Include in a Client Report: Build a white label AEO report your clients can read: which sections to include, how to define the metrics, and how to word each finding for non-specialists.
- AI Brand Monitoring vs Social Listening: What Each One Covers: See what AI brand monitoring and social listening each cover, such as AI answers versus public posts, where their data and goals overlap, and when a team needs.



