AI Keyword Research: What Assistants Do Well and What They Invent
Table of Contents
Keyword research with AI means using an assistant such as ChatGPT, Claude or Gemini for the language side of keyword work, then checking every phrase and figure against real search data. ProfileTree, a Belfast SEO and content agency, uses assistants to expand and group search terms, and never as a source of search volumes.
Most people start by asking an assistant for keywords with volumes attached. The table that comes back looks like the output of a paid tool; the phrases are often useful, and the numbers beside them are invented.
This guide sets out the tasks an assistant handles better than measurement software, what it makes up, a working method with prompts you can copy, the checks that keep a plan honest, and where a person has to take over.
Keyword research with AI and how it differs from a keyword tool
An assistant writes and sorts language, while dedicated keyword software measures demand. Treating one as the other is the source of almost every problem people run into, so the difference is worth setting out before any prompts are written.
What a language assistant brings to keyword work
A general assistant is trained on a vast amount of written text, so it is good at producing the many ways a person might phrase the same need. Give it a topic and it returns plain-English versions, trade jargon, beginner wording and the half-formed questions people type before they know the right term.
It is also quick at grouping. A list of several hundred phrases can be sorted by what the searcher wants in a minute or two, which is the slowest part of the job when done by hand. Neither task requires any knowledge of how often anything is searched, and that is why both work well.
What a keyword tool measures that an assistant cannot
Keyword tools hold databases built from search logs, advertising auctions and clickstream panels (records of how real people browse). When a tool reports that a phrase gets a certain number of searches each month, that figure comes from a measurement, however imperfect. An assistant holds no such database, so any number it offers is generated text that resembles a tool’s output.
Once that split is clear, the two stop competing. The assistant handles wording and grouping, and the tools tell you whether anyone searches the phrase at all. If you want the underlying process first, covering seed terms, volume, difficulty and mapping keywords to pages, start with this guide to what keyword research involves and come back here for the AI layer on top.
Why the question set matters more for AI search
Assistants and Google’s AI Overviews break a query into smaller questions before they look for answers. A page that covers the cluster of follow-up questions behind a search has more chances of being matched and cited than one written around a single headline phrase.
That shift makes the question-drafting side of keyword research more valuable than it was a few years ago. It is also the reason ProfileTree’s SEO Northern Ireland work now starts from the questions a buyer asks, with the headline phrase chosen afterwards, and why the same thinking runs through its AI search visibility work for clients across the UK.
Four jobs an assistant does faster than any keyword tool

Each of the jobs below is language work, which is where assistants are strongest. Used for these four tasks and nothing numerical, the software can save a marketing manager most of a working day on a single topic.
Producing phrasing variety from a short seed list
A keyword tool can only show you phrases already in its index. That misses newer wording and the way customers describe a problem before they know the industry term for it, which is often where the least contested searches sit.
An assistant has no such limit. Working for a kitchen showroom, ask an assistant how a homeowner might describe a cramped galley kitchen and it will produce phrases no supplier would think of, some of which turn out to be searched. Your job is to check which ones are real before any of them shapes a page.
Sorting long exports by search intent
Search intent is the goal behind a query: learning something, comparing options, buying, or finding a specific site. Sorting an 800-row export into those groups takes a full morning by hand, and a language model gets through a first pass in about a minute.
The first pass is right most of the time, which leaves you correcting a handful of rows. That is a much better use of an hour than labelling every row yourself, and it frees time for the judgement calls no model can make.
Drafting the questions a page should answer
Request a list of what someone searching a phrase still needs to know afterwards, and you get the follow-up questions that belong on the same page. This is the most directly useful output of the whole exercise, because it maps to how a page earns citations in AI answers as well as rankings.
The questions travel well to a writer too. A writer handed twelve questions produces a page that answers twelve questions, while a writer handed twelve keywords tends to produce a page that mentions twelve keywords.
Turning supplier language into customer language
Businesses describe themselves in trade terms, and customers search in problem terms. An accountancy practice offers “management accounts”, while its future clients search for help understanding why the bank balance never matches the profit figure.
Regional wording adds another layer for firms selling across the island of Ireland. Everyday speech in Belfast, Derry and Dublin differs in ways that show up in searches, a point made plain in ConnollyCove’s guide to Irish slang. An assistant can suggest local variants quickly, and a person from the area can tell you within seconds which ones a customer would actually type.
A working method from seed list to page brief
The method below takes a list of grounded starting terms through to a page brief a writer can use. Run it in a single conversation, so the assistant keeps its earlier output in view as you move from one stage to the next.
Building a seed list you can defend
Start with terms you already know are searched. Pull the queries your site appears for from Google Search Console, which reports real impressions and positions, then add the words your sales team hears on calls and the phrases used in the page titles of businesses that rank above you.
Ten to fifteen seeds is plenty, and they do not need to be clever. Expansion from real phrases strays far less than expansion from invented ones, so the time spent here pays back at every later stage.
Expanding the list by type of searcher
A request for “100 keywords” produces a list padded with near-duplicates. Ask for specific kinds of variants, and the output becomes usable: how a beginner would phrase the need, how someone comparing suppliers would phrase it, how someone ready to buy would phrase it, and how location changes the wording.
Two instructions improve the result further. Instruct the model to leave search volume out entirely, which stops it generating numbers you would have to strip later. Then tell it to mark anything it doubts is a real search, which gives you a column to check first.
Clustering phrases that one page can satisfy
The practical question in clustering is whether two phrases need one page or two, and the standard intent labels only help so far. Have the model group phrases where every phrase in a group would be fully answered by the same page, and to name any groups close enough to merge.
Overlapping clusters are how sites end up with several thin pages competing for one query, a problem known as keyword cannibalisation. Test each suggested merge by searching two or three phrases from each group. If Google returns broadly the same results for both, they belong on one page, whatever the assistant said.
A four-prompt sequence to copy
These four prompts cover expansion, clustering, the question set and a final audit. Replace the wording in the first prompt with your own business details and paste each one in order.
Prompt one: “I am doing keyword research for a kitchen showroom serving homeowners in Northern Ireland. Here are my seed terms. Expand them into related search phrases under four headings: beginner phrasing, comparison phrasing, ready-to-buy phrasing and location phrasing. Do not include search volume, difficulty or cost figures of any kind. Mark with an asterisk any phrase you are not confident a real person would type.”
Prompt two: “Group that list into clusters where every phrase would be satisfied by the same single page. For each cluster, give a short name, the primary phrase and one line on what the searcher wants. Then list any clusters close enough to compete and say which you would keep.”
Prompt three: “Take the largest cluster and list the questions a page targeting it needs to answer, in order of importance. Split them into questions most pages already answer and questions most pages skip. Write each one the way a person would type it into a search box.”
Prompt four: “Review everything you have produced in this conversation and list every claim that needs checking against a real data source before I use it, including any figure, any statement about what ranks, and any phrase you are unsure is searched. Be specific about what you do not know.”
The fourth prompt is the one people skip and the one that saves the most trouble. In ProfileTree’s experience, assistants are better at flagging their own uncertain output when asked directly than at avoiding it in the first place.
What AI invents during keyword research and how to catch it
Assistants reliably invent a handful of things, and all of them look identical to real data. Nothing in the formatting of a table separates a measured figure from a generated one, so the checks below are the part of the process that protects the plan.
Volumes, difficulty scores and click costs
Unless an assistant has been connected to a live data source, it has no access to search volume, and it will still produce a monthly figure for any phrase you name. The figures are usually plausible and consistent across a list, and they bear no relation to what people search.
Difficulty scores are a separate problem. Each keyword tool calculates difficulty its own way, so there is no general score for an assistant to know, and cost-per-click figures carry the extra risk that someone builds a budget on them. A useful tell is uniformity: measured data is lumpy, with odd numbers and big gaps, while generated data comes back smooth and neatly rounded.
Even measured figures need reading with care. When the team checked the UK numbers for this topic in November 2026, Google’s advertising data showed about ten searches a month for “AI for keyword research”, while one paid keyword database estimated 150 for the same phrase. Google’s Keyword Planner documentation explains that its average monthly searches include close variants, are averaged over twelve months by default and are shown for exact matches only. Two real sources can disagree that widely, and a generated figure has no source at all.
Dead phrases and out-of-date claims about rankings
Dead phrases are grammatically perfect, keyword-shaped strings that nobody has ever typed. They are the hardest invention to catch because they read well, and a phrase like “best commercial waste compaction solutions provider” looks exactly as convincing as one with real demand.
Claims about what ranks are the other weak spot. Ask which pages currently rank for a phrase and it may name titles and domains that no longer exist or never did. Unless it searched the live web during that answer, it is describing a memory of the results page.
Four checks that turn a generated list into a verified one
Verification is the step people drop first, and it is quicker than most expect. Run these four checks in order, and only for the phrases you plan to build pages around.
First, look in your own Search Console data. Any generated phrase that already shows impressions there is confirmed as a real search, and phrases where your site sits between positions 8 and 25 are usually the quickest wins available. Second, check volume in Keyword Planner, accepting that accounts with little ad spend may see ranges in place of exact counts. A range is still worth far more than a generated figure.
Third, search the phrase yourself. If the results do not match the intent the assistant assigned, the clustering is wrong, and if the results are a jumble of unrelated pages, the phrase is probably dead. Fourth, watch autocomplete and the People Also Ask box. If Google suggests the phrase as you type, somebody types it, and if nothing appears, drop it however good it looked on the list.
Connecting an assistant to live data sources
Some assistants can now be connected to Search Console, analytics platforms and keyword databases, so they read real figures during the conversation. That changes the picture for the data half of the job, because the assistant is reporting a number it fetched and no longer generating one.
Connection does not remove the need to check. Ask the assistant to name the source and date range behind every figure it gives you, and spot-check a few against the original platform. ProfileTree keeps a note of where each number in a keyword plan came from, so the work can stand up to questions from a client months later.
Where a person takes over, and how ProfileTree helps

An assistant stops being useful at the point where the question moves from language to money. The decisions below depend on knowledge of the business that no model holds, and they are where an experienced team adds the most.
Decisions that depend on profit and capacity
An assistant can tell you that “emergency plumber” and “24 hour plumber” share the same intent. It cannot tell you that one kind of job is profitable for you and the other is not, that you do not work beyond a certain area, or that the biggest cluster is full of people who never buy.
Capacity matters just as much. A keyword plan is only useful if somebody writes the pages, so the right size of plan depends on how much your team can publish in a quarter. That is a scheduling decision, and it belongs with the people who know the workload.
How this fits a content and SEO plan
ProfileTree is a Belfast web design and digital marketing agency, founded in 2011 and based at the McSweeney Centre. On client work, the team uses assistants for the first pass of expansion, clustering and question drafting, then verifies everything against Search Console and live results before a single brief is written.
Take a professional services firm that wants more enquiries from its website. The assistant might turn fifteen seed terms into several hundred phrasings in an afternoon; the team cuts that down to the clusters with real demand and a sensible margin behind them, and each surviving cluster becomes a page brief built around the questions buyers ask. Businesses that would rather hand the whole process over can use ProfileTree’s content marketing service, which runs from research through to published pages.
Training your own team to run the process
Plenty of teams prefer to keep keyword research in-house, and the gap between a useful prompt and a useless one is wider than most people expect. Teams that treat the assistant as a fast junior, who produces plenty of material and needs every piece checked, tend to get the best results.
“In ProfileTree’s SEO work for Northern Ireland businesses, the AI does the research, the structure and the first draft of the keyword list, and a person who knows the business decides what is true and what the customer needs. Take the person out and the result reads like everyone else’s.” Ciaran Connolly, founder of ProfileTree
ProfileTree’s digital training works on your own website and your own Search Console data, so your team leaves with a verified keyword list and a repeatable process.
Getting AI keyword research working for your business
Keyword research with AI is worth adopting for wording, grouping and question drafting, and worth distrusting for anything numerical. The prompt sequence takes about twenty minutes, the verification pass takes about the same, and the result is a plan built on figures somebody can trace. If you would like ProfileTree to run the process across your site or build it into your SEO strategy, get in touch with the team today.
FAQs
Can ChatGPT do keyword research?
ChatGPT can do the language half well, including expanding seed terms, grouping phrases by intent and drafting questions. It cannot supply real search volumes unless it is connected to a data source, so check every figure elsewhere.
Why do I get different search volumes each time I ask an assistant?
The figures are generated fresh each time you ask. A measured number would stay the same each time, so variation between answers is a reliable sign that nothing real sits behind them.
Does AI keyword research replace paid keyword tools?
No. It replaces the manual brainstorming and sorting around those tools, while the tools still supply the measured data.
How many keywords should one page target?
One primary phrase, plus the related phrasings that share its intent. Splitting a cluster across several pages makes those pages compete with each other.
Can AI tell me how hard a keyword is to rank for?
Not as a reliable score, because each tool calculates difficulty differently. An assistant can explain why a term looks competitive, which is useful context, but any number it gives is a guess.