Back to Blog

AEO Keyword Research: Find the Questions AI Answers Cite

Answer engines cite pages that answer one question well. How to find those questions with Google Trends, Autocomplete, News and Wikipedia in Claude, and write the answer.

9 min readSEO & AEO

TL;DR

Paulo H. Alkmin shows an answer engine optimization (AEO) research workflow in Claude using his free trendzeist-mcp server: aeo_opportunities clusters question-shaped searches and scores them by Google Trends momentum, Google News coverage and Wikipedia presence, mine_questions expands a seed through Google Autocomplete, and the answer_brief prompt turns one question into a 40 to 60 word answer with evidence. On his site, AI assistant referrals were 42 sessions in 90 days against 550 from organic search, but engaged more (64% vs 52%) and 12 of them landed directly on service or product pages. His worked example, llms.txt, had 40 Autocomplete questions, no Wikipedia article and steady coverage from search industry publications, making it a gap a direct, sourced answer page can fill.

In the 90 days to early November 2026, AI assistants sent my site 42 visits, 40 of them from ChatGPT. Google search sent 550.

By volume, that's a rounding error. Look at where those visits landed and the story flips. Twelve of the 42 assistant visits landed straight on a service or product page: my n8n and RAG consulting pages, and the Trendzeist page. Google's 550 landed on blog posts and the homepage, and the service pages got close to none of them. The assistant visitors were also more engaged (64% against 52% for search) and stayed longer (3 min 38 s against 2 min 46 s).

My reading: by the time someone clicks a citation in an AI answer, the assistant has already handled their beginner question. What's left is "who can do this for me", and that question lands on a service page.

So the useful question for answer engine optimization is which questions an AI answer needs a source for, and whether your page can be that source.

Questions, not keywords

Classic keyword research ranks phrases by volume. AEO keyword research ranks questions by how likely an answer engine is to need a citation for them, which depends on four things:

SignalWhy it mattersTrendzeist tool
The questions people typeAn answer engine answers questions, so the question is the unitmine_questions (Google Autocomplete)
Whether interest is growingA fading topic gets fewer questions to answerinterest_over_time (Google Trends)
Who covers it nowThe publishers covering it are the sources an assistant is likely to cite alreadynews_coverage (Google News)
Whether a definition anchor existsIf Wikipedia owns the definition, compete on something elsewiki_attention (Wikipedia)

aeo_opportunities runs all four for up to five seeds, clusters the questions by angle (how-to, definition, comparison) and scores each cluster, with a score_breakdown that explains every point. It also takes site_queries, the queries you already rank for, so each cluster comes back marked covered, partial or gap.

The prompt

Find the questions people ask about llms.txt and generative engine
optimization that an AI answer engine would need a source for.

Use Trendzeist aeo_opportunities with these seeds, passing my Search
Console queries as site_queries. Drop any question that doesn't mention
the seed. For the strongest gap cluster, expand it with mine_questions,
keep only questions a person would actually type, then check
news_coverage and wiki_attention.

Finish with an answer brief for the single best question: a 40 to 60
word direct answer, the evidence it needs, and who to cite.

I picked these seeds on purpose. Both topics sit close to what I build: this site generates its own llms.txt at build time, and ParseLess, my WordPress plugin, serves AI crawlers a Markdown version of each page. And in the visibility audit, both came back as category C, with real demand and zero impressions on my site.

Step 1: score the clusters

aeo_opportunities(seeds=["llms.txt", "generative engine optimization"], site_queries=[...]) returned six clusters in early November 2026. The top four:

ClusterQuestionsTrend (3 months)NewsWikipediaScore
llms.txt, how-to11Down 63%, with a breakout query12 articles in 30 daysNo article75
GEO, how-to7Down 34%12 articles in 7 daysArticle exists65
GEO, definition6Down 34%12 articles in 7 daysArticle exists65
llms.txt, comparison3Down 63%12 articles in 30 daysNo article55

Every cluster came back as a gap, which matches Search Console. The scores are a heuristic that only makes sense within one call, so read them as an ordering, not a grade.

Two fields matter more than the score. citability_hints tells you what kind of evidence makes an answer citable for that angle: for how-to, a direct 40 to 60 word answer first, then numbered steps with a number attached to each. And brief names the publishers to quote or pitch, which for llms.txt were Search Engine Journal, Search Engine Land and Search Engine Roundtable.

Then clean the list. Some question lists included "how to bake a cake" and "what is a roth ira", which arrived through Google Trends' related queries for the seed. Trends sometimes attaches unrelated rising searches to niche terms. The prompt's "drop any question that doesn't mention the seed" handles it, and you should still read what's left.

Step 2: expand the winner

mine_questions("llms.txt") ran 15 Autocomplete prefixes plus the bare seed and kept 40 questions, ranked by a realness score. It set 10 more aside as padding ("how long llms.txt last", "when llms.txt file"), which is what Autocomplete returns when it completes a prefix nobody really uses. The 40 it kept split into two groups:

PracticalSkeptical
how to create llms txt filedoes llms txt actually work
how to add llms txt in wordpressis llms.txt worth it
why llms.txt not foundis llms txt a standard
what is llms txt in seoshould i use llms txt
llms.txt vs robots.txtis llms txt actually used

The skeptical column is the interesting one, and three of its questions were among the six that tied for the top score. People who type "does llms txt actually work" have already read the hype and want somebody to give them a straight answer. Most pages ranking for llms.txt are enthusiastic how-to guides, so an answer engine summarizing "is it worth it" has few sources that address the doubt directly.

The practical column has one row that matches something I sell: "how to add llms txt in wordpress" is exactly what ParseLess does, at /llms.txt and /llms-full.txt, with no file to maintain by hand.

Step 3: check who answer engines already trust

news_coverage and wiki_attention decide whether you're competing with a definition or filling a hole.

For llms.txt, Wikipedia has no exact article. The closest it suggested was "Large language model", which doesn't define the file. Coverage comes from three search industry publications, steady but not saturated. That's an open slot: no canonical definition, and a set of questions with short factual answers.

GEO is the opposite. Wikipedia has a "Generative engine optimization" article averaging about 260 views a day, and it drew 12 news articles in a single week, from outlets like Yahoo Finance and USA Today. Any "what is GEO" answer already has a source it trusts more than my blog. The opening there, if any, is the comparison cluster ("generative engine optimization vs seo", "vs aeo"), where an answer has to take a position. I wrote that one as a section of my SEO overview: GEO vs AEO vs SEO.

Step 4: write the answer brief

Trendzeist ships an answer_brief prompt for this step: give it one question and it walks the assistant through the brief, direct answer first, then the evidence the angle needs and the sources. For "is llms.txt a standard?", the answer block I'd put directly under the H2:

No. llms.txt is a proposal published by Jeremy Howard of Answer.AI at llmstxt.org in September 2024: a Markdown file at a site's root that points language models to its most useful pages. No standards body has adopted it and support varies by tool, so treat it as cheap insurance rather than a ranking factor.

Fifty-four words: a verdict, the origin with a date, what it is, and a recommendation. Everything after it on the page (how ParseLess generates the file, what it looks like on this site, the robots.txt comparison) is supporting evidence. An answer engine quotes the top block, and a human reader keeps going for the proof.

The structure is the same on every page:

  1. The question as the H2, in the words people type.
  2. The answer in the first one or two sentences, with no warm-up.
  3. Evidence directly after it: a number, a date, a named source, your own data.
  4. The same question and answer in the page's FAQ schema.
  5. A Markdown version of the page for agents that ask for one. This site serves it to any client that sends Accept: text/markdown, and ParseLess does the same for WordPress.

How to tell whether it worked

Search Console tells you half of it. Since 31 August 2026 every site has a generative AI performance report that counts how often your pages appeared in AI Overviews and AI Mode, by page, country, device and date. It shows impressions only, with no queries, and assistant traffic from ChatGPT or Claude never touches it. Suganthan Mohanadasan has a good piece on AI conversations leaking into Search Console as strange query strings, which is about as close as Search Console gets to the questions themselves.

GA4 does better. Its default channel grouping has an AI Assistant channel that collects referrals from chatgpt.com, claude.ai, gemini.google.com and similar sources. Google's own AI Overviews and AI Mode aren't in it: GA4 counts those as Organic Search. Watch three numbers there:

  • Sessions per month, which will stay small for a while.
  • Landing pages, which show which answers are getting cited.
  • Engagement compared with Organic Search, which tells you whether those visitors are worth the effort.

On my site, the third number is what made AEO a priority: small counts, but engaged visitors who land on service pages.

Run it on your topic

Pick two seeds close to what you sell, pass your Search Console queries as site_queries, and let aeo_opportunities order the clusters. Expand the winner with mine_questions, keep the skeptical questions, check Wikipedia before you write a definition, and put a 40 to 60 word answer directly under every question heading. The setup is three commands, and Trendzeist itself is a one-line install. The last post in this series turns this into a calendar.

If your site runs on WordPress and you want this built in, from ParseLess configuration to an llms.txt strategy to measuring whether AI platforms actually use your content, that's the GEO/AEO implementation I offer. For the research on its own, there's the Search Opportunity Audit. Either way, it starts with a free 30-minute Intro Call.

Which question about your work do you keep answering on sales calls that no page on your site answers in fifty words?

Tags

  • Trendzeist
  • AEO
  • GEO
  • MCP
  • Google Analytics

Frequently Asked Questions

Read More Posts

Explore other articles and insights

Back to Blog