AI Keyword Research: The Model Drafts the List, Search Data Decides
AI keyword research is a method where a language model drafts the vocabulary and real search data decides which terms stay. The guide from Portland Peak SEO shows that split, prompts that help, such as asking for 30 ways a homeowner might describe a leaking water heater, and the three Google sources that settle every term, including Keyword Planner for monthly volume.
Written and maintained by Andreas Benavente, Founder, Portland Peak SEO · Updated October 24, 2026
GOOGLE HELP CENTER · SEARCH CENTRAL · PUBLISHED PRICESWhat is AI keyword research, and where does the model stop?
AI keyword research uses a language model to brainstorm phrasing, questions and groupings, and 3 Google sources then confirm which terms real people type. Google explains that generative models predict likely words rather than retrieve facts. Monthly volume from a model alone is guesswork, so Keyword Planner settles that number.
Google's guidance on generative AI content explains the limit: generative models "don't retrieve facts"; instead, they predict a likely sequence of words.
| Job | A model alone | What settles it |
|---|---|---|
| Synonyms and buyer phrasing | Strong | Your own Search Console queries |
| Question lists for a topic | Strong | Questions on the live results page |
| Grouping terms by intent | Good first pass | A person reading each results page |
| Monthly search volume | Guesswork | Keyword Planner, with an account |
| Your current positions | Blind | The Search Console Performance report |
So the model writes the draft, and data votes on every line of it.
Which prompts make a model useful for keyword research?
Prompts work when they name the buyer, the job and the page, like asking for 30 ways a homeowner might describe a leaking water heater. Vague keyword requests, however, return head terms you already know. It also helps to paste your own queries or page titles wherever you can. Never request volumes or difficulty scores from a model, since unsourced figures belong in the bin.
| Goal | Prompt to try |
|---|---|
| Buyer phrasing | List 30 ways a homeowner might describe a leaking water heater to a search engine. |
| Questions | What would a first-time buyer ask before hiring a roofer? Give 20 questions. |
| Objections | List the doubts a clinic owner has about paying for SEO, in their words. |
| Grouping | Here are my queries (paste). Group them by intent, then name one page per group. |
| Gaps | Here are my page titles (paste). Which buyer questions do none of them answer? |
| Local terms | Which neighbourhood or landmark names might people add when searching for a dentist in (town)? |
Paste your own data whenever you can. A model that reads your real queries drafts far better groupings than one guessing.
Never ask for volumes or difficulty scores, because any figure a model prints without a source belongs in the bin.
How do you check AI keyword ideas against real Google data?
Run each drafted term through 3 Google sources, namely Keyword Planner, Search Console and the live results page. Keep a term only if it survives all three. According to Google, Planner needs billing details and hides very low-volume and sensitive terms. Search Console, on the other hand, covers only your own site.
| Source | What it tells you | Its limit, per Google |
|---|---|---|
| Keyword Planner | Keyword ideas as well as estimated volumes and costs | Needs billing details and also hides very low-volume and sensitive terms |
| Search Console Performance | Queries you already appear for, together with clicks and average position | Covers only your own site |
| The live results page | Whether guides, shops or maps own the term | One search, one moment |
Google built Keyword Planner for ad campaigns, so its cost figures describe ad prices. Treat its volumes as estimates rather than counts.
The Performance report defines average position as the topmost result from your site. For example, a query near position 11 often deserves a better page.
Read the results page by eye last. If maps and directories fill it, a guide will struggle there.
How does query fan-out change a keyword list?
Query fan-out adds follow-ups to a keyword list, so water heater replacement picks up 4: tank or tankless, permit, lifespan and cost range. One AI Overview can draw on several searches that Google runs behind the scenes. Those follow-ups become sections of one replacement page rather than pages of their own.
Google calls the technique query fan-out on its AI features page. One question can therefore reach pages written for its sub-questions.
| Head term | Follow-ups a model can draft | Where each belongs |
|---|---|---|
| Water heater replacement | Tank or tankless, permit, lifespan, cost range | Sections of one replacement page |
| Emergency plumber | After-hours fee, response area, what to shut off | The emergency page as well as a short checklist |
| Kitchen remodel | Timeline, budget drivers, permits, living through it | One remodel guide, one cost page |
However, Google's guide to generative AI search warns against separate content for every variation made mainly to manipulate. It also says that can violate its policies.
For how AI Overviews then pick sources, see our AI Overview optimization guide.
How do you turn the checked list into a page plan?
Sort surviving terms into 4 groups (learn, compare, hire and price), then give each search intent 1 owner page. Close variants become sections or synonyms on that page instead of pages of their own. Before any change, note where each owner page ranks today.
- Sort surviving terms into intent groups: learn, compare, hire, price.
- After that, pick one URL for each group, preferring a page that already earns impressions.
- Then list the follow-up questions that page must answer, in the order buyers ask them.
- Mark any two pages that share a group, then merge their content into one.
- Record the start position for each owner page before you change anything.
Note that mass-producing near-identical pages risks Google's scaled content abuse policy. Its spam policies specifically name AI tools used that way.
The map then feeds headings and copy, covered in our on-page SEO guide.
Where does AI-driven keyword research usually go wrong?
When a model drafts the list, 4 errors repeat: trusting invented numbers, chasing every variant, skipping the results page and recording no baseline. The first 3 each waste a month of writing, and the last leaves nobody able to judge the plan. Check any confident volume with no source in Keyword Planner; otherwise, drop it.
Invented volumes
A confident number, but with no source. Either check it in Keyword Planner or drop it.
Variant sprawl
Ten pages for ten phrasings of one need. One page serves them better than ten.
Wrong format
A blog post aimed at a term that maps and shops own.
No baseline
No start positions recorded, so nobody can judge the plan later.
Unfamiliar terms here, such as search intent, sit in our SEO glossary.
What does AI keyword research cost with Portland Peak SEO?
It carries no separate fee, because the research shapes the SEO pages each monthly plan already includes, starting with 7+ on the $450 Local plan. On the $950 Growth plan, by comparison, the count is 10+, across up to 5 locations. Each plan bills by the month.

Andreas Benavente, founder
Andreas leads the keyword work and also works with you by phone and video from Lake Oswego, Oregon.

| Plan | Monthly fee | SEO pages | Posts a month | Targeting |
|---|---|---|---|---|
| Local | $450 | 7+ | 2 | Local search |
| Growth | $950 | 10+ | 4 | Up to 5 locations |
| Authority | $2,500 | 20+ | 8 | Regional or national |
Ready to start? Buy the Local plan ($450 a month), the Growth plan ($950 a month) or the Authority plan ($2,500 a month) online, month to month with no contract, or ask for the free audit first.
Billing runs monthly, and stopping needs no warning. The pricing page also shows every figure, with no setup charge.
A new keyword map usually needs 90 to 180 days to show, but we promise no date.
Andreas Benavente founded the company and still leads the keyword work; read more about the founder.
If you prefer research as a standalone project, compare our keyword research services page first.

What do clients say about our AI and SEO work?
Our Google Business Profile showed a 5.0 rating from 4 Google reviews when we read it on October 23, 2026. Two of those reviews appear below word for word, one naming our AI skills directly and the other our rankings, including each reviewer's name as Google shows it. Every review, and the current count, is also on our Google Business Profile.
★★★★★
“5 star service. Whenever we need changes on our website Andreas calls us back Immediately. The AI skills are top notch. They made our website stand out and get noticed.”
★★★★★
“Andreas is a wizard with rankings. His websites are amazing, his communication is superb. Cannot recommend enough.”
These are clients' own words, so we have not edited them. One client's experience, however, does not predict another's.
Which guides continue from a keyword plan?
Keyword work hands off to 3 threads, writing, links and technical fixes, and the guides here pick each one up. The writing guides cover citation-ready copy and page-level SEO, as well as expertise signals. Links and local reach sit with the growth guides, while technical crawling sits with the AI search guides on chat assistants and overviews.
AI keyword research questions
Tools and data
Can ChatGPT tell me how many people search a keyword?
Not reliably. Per Google, a generative model writes probable text but fetches no figure from a search log. Any volume a model prints needs a source you can open, such as Keyword Planner.
Which AI tool is best for keyword research?
None stands out, because the data matters more than the model. Any capable assistant drafts phrasing, while Google tools supply the numbers.
Why does Keyword Planner show nothing for my health or finance terms?
Google says its advertising and privacy policies often block results in sensitive categories. Very low-volume terms stay hidden too.
Do I need a page for every keyword variation?
No. Google says not to worry about having enough long-tail keywords. One strong page can answer a cluster of close variants.
How do I find the questions buyers type?
Filter your Search Console queries for words like how, what, why or which. Then let a model suggest more for you to check.
Price and terms
Is keyword research sold separately?
Not by us. Our pricing page lists no separate fee; research feeds the SEO pages each monthly plan includes.
How soon does a new keyword plan show results?
Usually within two quarters, 90 to 180 days. Google sets the pace, so we name no date.
Can I cancel if the plan is not working?
Yes. Stop in whichever month you choose: no warning, and no fee at either end.
Request a free keyword check of your site
List three searches you want to win, as well as your web address. A person then reads them against your Search Console data and replies in writing.
Your information is secure. We also never spam or share your data.
More on keywords, AI search and our prices
Turn a model's draft into a plan you can check
Bring the list a chatbot gave you. We will then show which terms real searchers type.