AI Keyword Research: What Assistants Do Well and What They Invent

Keyword research with AI means using a general assistant such as ChatGPT, Claude or Gemini for the language side of keyword work, then checking its output against real search data. An assistant is fast at expanding a seed list into hundreds of realistic phrasings, sorting a messy export into intent clusters, and drafting the sub-questions a page has to answer. It has no access to search volume, competition or difficulty, so any number it gives you is generated rather than measured, and it will produce those numbers confidently if you ask for them.

The working method is to let the assistant handle wording and grouping, then verify every phrase and every figure against Search Console, Keyword Planner and the live results page before anything reaches a content plan. Used that way it saves hours. Used as a data source it produces a plan built on figures nobody can trace.

Most people who try AI keyword research for the first time ask an assistant for keywords with search volumes attached, receive a tidy table, and take it away. The table looks exactly like the output of a paid tool. The phrases in it are often good. The numbers beside them are invented, and there’s nothing in the formatting to tell you which column is which.

That’s the whole problem in one sentence, and it’s fixable. This page sets out what an assistant genuinely does better than a keyword tool, what it makes up, how to catch it, and a prompt sequence you can copy and run today. If you want the underlying process first, covering seed terms, volume, difficulty and mapping keywords to pages, start with our keyword research guide and come back here for the AI layer on top.

What Is Keyword Research with AI, and How Is It Different From Using a Keyword Tool?

AI keyword research is the use of a general-purpose language assistant to do the interpretive parts of keyword work: generating phrasings, grouping terms by what the searcher wants, and turning a topic into the set of questions a page should answer. It’s a different job to the one a keyword tool does. A tool measures demand. An assistant writes and sorts language.

Keyword tools such as Semrush, Ahrefs and Google Keyword Planner hold databases built from clickstream data, ad auction data and search logs. When they tell you a phrase gets 1,900 searches a month, that figure comes from a measurement, however imperfect. An assistant doesn’t hold one. It produces text that resembles the output of those tools because it has read a great deal of that output, which isn’t the same thing at all.

Once you’ve accepted that split, the two stop competing. The assistant does the part tools are weak at, which is language and grouping. The tools do the part assistants cannot do at all, which is telling you whether anyone actually searches the phrase. The same division applies across search engine optimisation generally: machines are good at volume of output, people and measured data are good at judgement.

This also connects to how answers get assembled now. AI SEO work depends on covering the cluster of related sub-questions behind a query rather than one headline phrase, because assistants and AI Overviews break a question into parts before they search. That makes the question-generation side of AI keyword research more useful than it was a few years ago, when the unit of work was a single phrase and its volume.

What Does an Assistant Do Faster Than a Keyword Tool?

Four things, all of them language work rather than data work.

Phrasing variety. Give an assistant a seed term and it will produce the plain-English, jargon, regional and beginner versions of the same idea in one pass. A tool can only show you phrases that are already in its index, which systematically misses newer wording and the way customers describe a problem before they know the industry term for it.

Intent grouping at volume. Sorting an 800-row export into informational, commercial, transactional and navigational buckets is a full morning by hand if you’re doing it yourself. An assistant does it in a minute and is right most of the time, which leaves you correcting a handful of rows instead of sorting all of them.

Question generation. Ask 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 rather than just a ranking.

Translation between vocabularies. Businesses describe themselves in supplier language. Customers search in problem language. An assistant moves between the two quickly, which is where a lot of missed demand hides.

None of those four require the assistant to know a single search volume, which is exactly why they work.

How Do You Expand a Seed List With an Assistant?

AI keyword research works better from a grounded starting point, so start with terms you can defend. Pull the queries you already appear for from Search Console, take the words your sales team hears on calls, and add the phrases your competitors use in their page titles. Ten to fifteen seeds is plenty, and they don’t need to be clever.

Then ask for expansion by category rather than volume. A request for “100 keywords” produces a list padded with near-duplicates. A request for specific types of variant produces something usable: how a beginner would phrase it, how someone comparing suppliers would phrase it, how someone ready to buy would phrase it, and how someone in Northern Ireland or Ireland would phrase it if location changes the wording.

Two rules make the output better. Tell the assistant to leave search volume out entirely, which stops it generating numbers you would then have to strip. And tell it to flag anything it is unsure is a real phrase, which gives you a column to check first rather than a list that hides its own guesses.

How Do You Cluster a Keyword List by Search Intent?

Clustering is where AI keyword research earns most of its time back. Paste the list in and ask for groups where every phrase in a group would be satisfied by the same page. That framing matters more than the standard intent labels, because the practical question isn’t “is this commercial” but “do these two phrases need one page or two”.

Ask for three things per cluster: a name, the phrase that best represents it, and a one-line description of what the searcher wants. Then ask the assistant to name any clusters that look close enough to merge. Overlapping clusters are how sites end up with four thin pages competing for one query, which is the same cannibalisation pattern that shows up in most content audits.

Check the merge suggestions yourself by searching two or three phrases from each pair. If Google returns broadly the same results for both, they belong on one page. If the results differ, they don’t, whatever the assistant said.

How Do You Draft the Question Set a Page Needs to Answer?

Take one cluster at a time and ask for the questions someone searching that cluster would want answered on the page, ordered from most to least important, with the obvious ones separated from the ones most pages skip. That second group is where the useful material sits.

Then cross-check against sources that are not generated: the People Also Ask box for your main phrase, the autocomplete suggestions, and the questions your own inbox gets asked. An assistant’s good at producing plausible questions and less good at knowing which ones people actually ask, so the generated set is a starting draft rather than a finished brief.

The output of this step is what you hand a writer. It travels better than a keyword list, because a writer given twelve questions produces a page that answers twelve questions, while a writer given twelve keywords produces a page that mentions twelve keywords.

What Does AI Invent During Keyword Research?

Five things, reliably, and all five look identical to real output. This is the part of AI keyword research that causes actual damage, because nothing in the formatting separates a measured figure from a generated one.

Search volumes. The most common and the most damaging. Assistants have no live connection to search data unless you have explicitly connected one, and they will still produce a monthly volume for any phrase you name. The figures are usually plausible, internally consistent across a list, and unrelated to reality.

Keyword difficulty scores. Difficulty is a proprietary metric, calculated differently by every tool that publishes one. There’s no general difficulty score for an assistant to know. A number out of 100 with no tool behind it is a guess wearing a uniform.

Cost-per-click figures. Same problem, with the added risk that someone builds a budget on them.

Dead phrases. Grammatically correct keyword-shaped strings that no human has typed. These are the hardest to catch because they read perfectly. “Best commercial waste compaction solutions provider” is a real-looking phrase and an empty one.

Claims about what ranks. Ask which pages currently rank for a phrase and you may get titles and domains that do not exist, or did once. Unless it’s searched the live web during that answer, it is describing a memory of the results page rather than the results page.

The tell is uniformity. Real search data is lumpy: odd numbers, big gaps, phrases with surprising demand. Generated data is smooth, rounded and evenly spaced. If a list of thirty keywords has volumes that all sit in tidy multiples, none of it’s measured.

How Do You Verify an AI Keyword List Against Real Data?

Verification is the step people drop first, and it’s the one that makes AI keyword research safe to use. It isn’t slow either. Four checks, in this order.

1. Your own Search Console data first.Google Search Console shows the queries your site already appears for, with real impressions and positions. Any generated phrase that also shows up there is confirmed as a real search, and phrases where you already sit between positions 8 and 25 are usually the fastest wins available. This is the only data in the process that’s both free and specific to you.

2. Volume, with the limits understood. Google’s Keyword Planner reports average monthly searches for a term and its close variants, averaged over a twelve-month period by default, and those historical figures are shown for exact matches whatever match type you enter, according to Google’s own documentation. Accounts with little or no ad spend see bucketed ranges rather than exact counts, a restriction Search Engine Land reported when Google introduced it. A range of 1K to 10K is still infinitely better than a generated 4,400.

3. The results page itself. Search the phrase. If the results do not match the intent the assistant assigned it, the clustering is wrong. If the results are a mess of unrelated pages, the phrase is probably dead, whatever it looked like on the list. Thirty seconds per phrase, and you only need to do it for the ones you plan to build pages around.

4. Autocomplete and People Also Ask. If Google suggests the phrase as you’re typing it, somebody types it. If nothing appears and the phrase returns nothing sensible, drop it regardless of how good it looked in the list.

Run those four and the generated list becomes a verified one. Skip them and you have a content plan whose foundations nobody can check, which isn’t an easy thing to explain to a client six months later.

What Does a Full AI Keyword Research Prompt Sequence Look Like?

Four prompts, run in order, in one conversation so the assistant keeps the earlier output in view. Replace the bracketed parts and paste them as they are.

Prompt 1: expand the seed list

I am doing keyword research for [business type] serving [location and customer type]. Here are my seed terms: [paste 10 to 15 terms].

Expand these into a list of related search phrases, grouped under four headings: beginner phrasing, comparison and research phrasing, ready-to-buy phrasing, and location-specific phrasing.

Do not include search volume, difficulty or CPC figures of any kind. You do not have access to that data and I will get it from a tool.

Mark with an asterisk any phrase you are not confident is something a real person would type, so I can check those first.

Prompt 2: cluster by intent

Group that list into clusters where every phrase in a cluster would be satisfied by the same single page.

For each cluster give me: a short cluster name, the phrase I should treat as the primary target, and one line describing what the searcher actually wants.

Then list any two clusters that are close enough that building separate pages would make them compete with each other, and say which one you would keep.

Prompt 3: draft the question set

Take the cluster called [cluster name]. List the questions a page targeting it needs to answer, ordered by importance.

Split them into two sections: questions almost every page on this topic already answers, and questions most pages skip but a reader would still want answered.

Phrase each one the way a person would type it into a search box, not the way a marketer would write a heading.

Prompt 4: audit your own output

Review everything you have produced in this conversation and list every claim that would need checking against a real data source before I use it, including any figure, any statement about what currently ranks, and any phrase you are uncertain is genuinely searched.

Be specific about what you do not know rather than reassuring me.

The fourth prompt is the one people skip and the one that saves the most trouble. Assistants are noticeably better at identifying their own uncertain output when asked directly than at avoiding it in the first place.

Where Should AI Stop and a Person Take Over?

At the point where the question stops being about language and starts being about money. An assistant can tell you that “emergency plumber Belfast” and “24 hour plumber Belfast” are the same intent. It can’t tell you that one of those jobs is profitable for you and the other isn’t, that you won’t take work beyond a certain postcode, or that the cluster with the most searches is full of people who never buy.

It also hasn’t any view of your capacity. A keyword plan’s 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’s a scheduling decision, not a search decision.

Teams that get good results from AI keyword research treat the assistant as a fast junior who produces a great deal of material and needs all of it checked. Teams building the skill in-house tend to find that practical AI training pays for itself quickly, mostly because the difference between a useful prompt and a useless one is larger than people expect.

FAQs About AI Keyword Research

Getting This Working in Your Own Process

AI keyword research is worth adopting for the language work and worth distrusting for anything numerical. The prompt sequence above takes about twenty minutes end to end, and the verification pass afterwards takes about the same. That is still a fraction of what the same work costs by hand, and what comes out the other side is a plan built on figures somebody can check. Get in touch if you would rather we ran it across your site.