How to Do Keyword Research for AI Search: An AEO + LLM Guide

TL;DR

Keyword research for AI search starts with prompts, not search volumes. Collect 20–40 questions your buyers actually ask, group them by the kind of answer they need, check which pages AI cites today, then publish one page per group. Track mentions, citations and position every week. Google says no special schema or AI text files are required — ordinary indexing and a snippet-eligible page are the baseline.

Why AI keyword research is not Google keyword research

Traditional keyword research assumes one query, one page, one rank. AI search breaks all three assumptions.

Google says its AI features "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources." One prompt can pull in five pages that were never optimized for that exact string. That cuts both ways: a page can be cited for prompts you never targeted, and a page that ranks first for a head term can be ignored because it does not answer the sub-question the model needs.

Two more differences matter:

Volume does not predict demand. A prompt like "how do I track AI Overviews rankings over time?" shows nothing useful in Keyword Planner and gets asked every day in chat. The reverse is also common: high-volume head terms get answered with a definition and no citation.
A citation is not a ranking. In a blue-link result, position one is the product. In an AI answer, the model writes the response and selects a few sources. You are competing to be one of those sources, not to be the result.

Google also removes the technical shortcut. The company states there are "no additional requirements to appear in AI Overviews or AI Mode" and "no special schema.org structured data that you need to add." Eligibility is ordinary: the page must be indexed and eligible to show a snippet. The work is coverage, clarity and proof, not markup.

Step 1: Build a prompt set, not a keyword list

Start with 20–40 prompts. The best sources are not keyword tools:

Sales calls and demo notes: the question someone asks before they buy.
Support tickets and onboarding emails.
"People also ask" boxes for your head terms — already question-shaped.
Your own chats with ChatGPT and Perplexity while researching the category.
Competitor comparison searches: "X vs Y", "X alternatives", "X pricing".

Write each prompt the way a buyer would say it, not the way a keyword tool formats it. "How do I check if my brand shows up in ChatGPT?" is a prompt. "ChatGPT brand visibility" is a keyword.

Then add four columns to a sheet. This is the only artifact the process needs:

Prompt	Answer type	What AI cites today	Do we have a page?
how do i track ai overviews rankings over time	How-to	Competitor blog, no template	No
best answer engine optimization tools for agencies	Recommendation	Directory listing, vendor listicle	Yes, but no pricing
what is answer engine optimization	Definition	Explainer, no examples	Yes
your brand vs competitor	Comparison	Competitor's own page	No

Fill the third column by running the prompts, not by guessing. Fill the fourth with a URL or "no". Most teams discover that the pages they own answer the questions they wanted to answer, not the questions buyers ask.

For the commercial end of the topic, a free keyword research tool still helps — use it to size paid demand and find adjacent terms, not to decide whether a conversational prompt is worth answering.

Step 2: Group prompts by answer type

Answer type determines page format, and different types get cited for different reasons. Five groups cover most B2B categories:

Definition — "what is X". Short, direct, one citable paragraph. Easiest to win, least commercial.
How-to — "how do I X". Steps, a template, a checklist. Models reward extractable structure.
Comparison — "X vs Y". A table plus a decision rule. High commercial intent.
Recommendation — "best X for Y". Lists with prices, limits and drawbacks. Hard without brand authority, easy to lose on stale prices.
Cost — "how much does X cost". Numbers with dates and sources.

One group, one page. Do not put "what is AEO" and "best AEO tools" on the same URL. They answer different questions and collect different citations. For the writing side of this — how to make each page quotable — our GEO guide covers the structure in detail.

Which prompts to do first

Score each group on three questions:

Does a buyer ask it before a purchase? Comparison and cost prompts sit closer to revenue than definitions.
Does anyone answer it well? An unowned prompt is a cheaper win than a contested one.
Can you prove the answer? If the page needs data you do not have, it waits.

Do the highest-scoring group first and ignore the rest for a month. Twenty prompts do not need twenty pages; they usually collapse into three to five.

Step 3: See who AI cites today

Run each prompt in the engines that matter to you: ChatGPT, Perplexity, Gemini, and Google AI Overviews where available. Log the cited URLs, not your impression of the answer.

Three patterns appear quickly:

Nobody owns the prompt. The model cites generic pages that only partly answer. This is your best opening: publish the page that answers it completely.
One competitor owns it. You need a better page for a narrower version of the question, not a rewrite of theirs.
Directories and forums own it. G2, Reddit and review sites are hard to displace with a vendor page. Either earn a listing or answer a sub-question those pages ignore.

Freshness is a tiebreaker, not a strategy. Ahrefs analyzed 16.975 million cited URLs and found AI-cited pages average 1,064 days old versus 1,432 days for organic Google results — 25.7% fresher. But the average AI citation is still about 2.9 years old. Updating a page helps when the facts changed. Publishing a new page every week does not.

Engine behavior differs too. In the same study, ChatGPT skewed newest while Google AI Overviews and organic results skewed oldest. A page that wins in ChatGPT may never appear in an AI Overview, so track engines separately.

Step 4: Turn each group into one page

For every prompt group with no clear owner, publish one page built to be quoted:

Use the prompt as the H2, in the buyer's words. It is also the string a model is most likely to match.
Answer in the first 60 words. One direct paragraph, then detail. If a reader has to scroll to find the answer, so does the model.
Keep paragraphs to one to three sentences. Long paragraphs get summarized; short ones get extracted.
Add one asset a competitor cannot copy quickly. A table, checklist, template or worked example.
Link to the commercial page it supports, with descriptive anchor text. An article about tracking should point to the tool that does the tracking.

That last point is how content props up a product page. If a tool page already ranks for a phrase like keyword research tools for AEO in LLMs, the informational article should target the upstream question and send authority to it rather than compete with it.

A worked example, using the table from Step 1:

Twelve prompts about tracking AI answers collapse into three pages: how to track AI Overviews (how-to), AI visibility tools compared (recommendation), and what is share of voice in AI search (definition).
The how-to page carries the template, so it earns citations from the prompts that need a process.
The recommendation page carries prices and limits, so it earns citations from buyers.
The definition page is short and links to both. It exists to be quoted, not to convert.

If the team has time for one page this month, it is the how-to. Templates and checklists are the most extractable format in AI answers, and they are the hardest for a competitor to copy without doing the work.

Step 5: Measure mentions, citations and position

Four numbers per prompt, recorded weekly:

Metric	What it means	How to check
Mention rate	Share of runs where your brand appears at all	Manual or tool
Citation rate	Share of runs where an answer links your URL	Manual or tool
Position	Where you appear in the answer: first mention, second, footnote	Manual
Share of answer	How much of the response is about you versus competitors	Manual

Two cautions. First, scores are not comparable across vendors: a visibility score of 70 in one tool can mean something different in another, because engines, prompt sets and weighting differ. Second, if a tool hides the raw answer, you cannot verify the number. Keep the screenshot.

Start with a spreadsheet. A tool becomes worth paying for when you need daily runs across several engines, or when several people need the same baseline. Before that, a sheet and a weekly habit will show you more than a dashboard you check once.

The 30-minute weekly loop

5 minutes — rerun your 10 most important prompts.
10 minutes — log mention, citation, position, and the URLs cited instead of you.
5 minutes — note which pages entered or left the citation set.
10 minutes — pick one page to fix: add the missing sub-answer, update the price, add the table.
Monthly — retire prompts nobody asks anymore and add the questions from last month's sales calls.

That is the entire system. It needs a sheet and a habit, not a new platform.

Key Takeaways

AI search fans one prompt into many searches, so prompts — not volumes — are the unit of research.
Group prompts by answer type and publish one page per group.
Check who is cited before writing; an unowned prompt is the cheapest win.
Google requires no special schema or AI files; indexing and snippet eligibility are the baseline.
Track four numbers weekly: mention, citation, position, share of answer.
Freshness is a tiebreaker; the average AI citation is still about 2.9 years old.

Frequently Asked Questions

How is AI keyword research different from traditional keyword research?

Traditional research counts searches for a string. AI research maps questions to answer types and checks which pages models cite. Volume still helps, but it does not predict whether you will be cited.

Should I stop using Google Keyword Planner?

No. Use it to size the commercial end of a topic and to find adjacent terms. Do not use it to decide whether a conversational prompt is worth answering — most of those show zero volume there.

How many prompts do I need?

Start with 20. Run ten weekly and ten monthly. More than 40 becomes a data-entry job nobody maintains.

Do I need a paid tool?

Not to start. Twenty prompts by hand for two weeks will show you the gaps. A paid tool earns its price when you need daily runs across several engines, or when several people need the same baseline.

Does schema markup help me get cited?

Not the way tool vendors imply. Google states there is no special schema.org markup required for AI Overviews or AI Mode. Clean structure helps readers and extraction; schema is not the lever.

How long until AI answers mention my brand?

Chat engines can move within days for a specific prompt; Google AI Overviews move on Google's crawl and ranking cycle. HeyEmmett's published case studies include Assembo.ai going from a 1% to 10% ChatGPT show-up rate in seven days. That is one prompt set, not a category-wide guarantee.

Sources

Ready to see which prompts cite you?

Run a free check on AI keyword visibility. It shows how your brand appears in AI answers before you pay. If you want the full workflow — prompt sets, tracking and the pages that fix the gaps — HeyEmmett's app starts at $99/month; plans are here.

Brand Info

Sign in to HeyEmmett

Course Coming Soon!