AIO Copilot

How to Do Keyword Research With AI: What It Does Well, What It Makes Up, and the Tools Worth Paying For

Language models handle the wordy half of keyword research well: expanding a list, sorting it, labelling intent, drafting the brief. They have no access to search counts, so any volume they offer on their own is made up. A process that keeps the two apart, with prompts, and what the tools cost in November 2026.

Updated 2 November 202614 min read

Can you do keyword research using AI?

Yes, with one hard limit. A language model can do the language work of keyword research, suggesting phrasings, sorting a list into topics, labelling intent and drafting the brief, but it cannot tell you how often anything is searched. ChatGPT, Claude and Gemini have no access to Google's query logs, so a search volume one of them produces unprompted is a guess shaped like a number, not a measurement.

That limit decides how the work should be split. The counts come from data: the queries Search Console already reports for your site, the volumes in Google Keyword Planner, and the modelled estimates in tools such as Semrush and Ahrefs. The model sorts, labels and writes around those numbers, and says "missing" when one is absent.

What follows is why the line falls where it does, a process that runs from a Search Console export to a finished brief, prompts for the steps a model handles well, and what the main tools cost and do as of November 2026. Keyword research ends in site structure, so this process stops where our guide to building topic clusters picks up: each group of searches assigned to one page.

Why can't ChatGPT or Claude tell you a keyword's search volume?

Because the counts live in one place, Google's query logs, and Google does not publish them. The volume figures in SEO tools come from three narrower sources: what Google releases through Keyword Planner, what Search Console reports about your own site, and third-party data that tool vendors buy and model. A language model has access to none of them.

Keyword Planner is the nearest thing to a first-party count, and Google hedges it. Its help page on Keyword Planner statistics says average monthly searches cover a keyword "and its close variants", are "averaged over a 12-month period" by default, and that "your search volume statistics are rounded." Accounts that spend little get less. In 2016 Google said that "advertisers with lower monthly spend may see a limited data view," in which a keyword searched 500 times a month shows as the range 100 to 1K, Search Engine Journal reported at the time. Google did not say what counts as lower spend.

Search Console counts real impressions, but only for your own site, and not all of them. Google omits anonymized queries, which it defines as "those that aren't issued by more than a few dozen users over a two-to-three month period."

The commercial tools model the rest. Semrush says it calculates volume with "machine learning algorithms" over third-party data and "historical clickstream data acquired from reliable sources," then divides 12 months of searches by 12, according to its knowledge base. Ahrefs uses "Google Keyword Planner data as a base to build our model," its Academy says, and its help centre adds Google Trends and other third-party sources, concluding that "we don't believe there's a real 'accurate' search volume."

Even these measured estimates are rough. In a 2021 study of 72,635 keywords, Ahrefs compared volumes with Search Console impressions for the same month and counted a figure as roughly accurate if it was within 50%. Keyword Planner met that bar for 45% of keywords and overestimated in 91% of cases; Ahrefs' own figures met it for about 60%. Ahrefs ran the study and sells the tool that came out ahead, but the yardstick, Search Console, is Google's.

Keyword Planner's volume was within 50% of real impressions for 45% of keywords, and Ahrefs' for about 60%.
Google Keyword Planner45%
Ahrefsabout 60%
Share of keywords with a roughly accurate search volume estimate, by source
Google Keyword Planner45%
Ahrefsabout 60%
Share of 72,635 keywords (1K to 10K monthly volume range) whose estimate was within 50% of Search Console impressions for the same month. Source: Ahrefs, GSC vs. GKP study, November 2021. A vendor-run study.

A language model has none of these inputs. Without a tool connected, it generates text by predicting what usually comes next; asked for a volume, it writes a number that looks like the volumes it has seen, and nothing in that process looks a query up. Researchers at OpenAI and Georgia Tech argued in September 2025 that models "guess when uncertain" because "the training and evaluation procedures reward guessing" over admitting doubt. The monthly search count for one phrase in one country is exactly the kind of specific fact a model cannot know and will still answer. Ahrefs says so in its own guide to AI keyword research: "ChatGPT has no keyword database of its own, so any volume, KD, or SERP figures it generates unprompted are fabricated." Ahrefs sells a keyword database, but the mechanism holds whoever describes it.

The exception is a model connected to data. The same Ahrefs guide describes Claude using the Ahrefs MCP server to "query real search data mid-conversation," and any model can read an export you paste or upload. The numbers are then the tool's estimates, with the tool's error, and the model is doing the sorting. The process below is built on that split.

What is AI good at in keyword research, and what is it bad at?

A model is reliable where the answer is in the text you hand it, and unreliable where the answer sits in a database it cannot see.

TaskUse a model?Where the facts come from
Brainstorming seed terms and phrasingsYesCheck each one in Keyword Planner or a paid tool
Sorting a measured list into rough topicsYes, as a first passSERP overlap decides the borderline pairs
Labelling intent across hundreds of queriesYesThe live results page, for queries that matter
Drafting a content briefYesYour exports, the ranking pages, People Also Ask
Search volume, difficulty, cost per clickNoKeyword Planner, Search Console, Semrush, Ahrefs
What ranks for a query todayOnly with live search or a connected toolThe results page itself
Whether interest is risingNoGoogle Trends, which scales interest from 0 to 100 rather than counting searches

Ahrefs' guide lists the same strengths: brainstorming variations, clustering "by meaning and intent," tagging intent and turning clusters into briefs. Each is a job of reading and rewriting. On the last row, Google's Trends help page says each data point is divided by total searches for its place and time, then "scaled on a range of 0 to 100," which makes Trends good for direction and seasonality and useless as a volume.

Where should AI-driven keyword research start?

In Search Console, with the searches Google already shows your site for. It is the only source with real impressions, clicks and positions for your own pages, and it costs nothing.

The Performance report holds 16 months of data, enough to see a full year of seasonality. Export the queries for the pages or folder you are working on. The interface exports at most 1,000 rows; the API allows "50,000 rows per day per site per search type," according to the same Google post on data limits. For a small site the interface is enough. For a large one, use the API or the bulk export.

Three patterns in that export are worth more than any seed list. Queries where the site sits on page two collect impressions without clicks, which means Google already connects the page with the search. Pages that appear for queries they were never written for point to a missing page. And question-shaped queries, usually long and specific long-tail keywords, tell you what each section should answer. Our guide to Search Console regex filters has the patterns for pulling question and conversational queries out of the report.

The long tail is also where the volume tools run dry. Google's Keyword Planner guide says "keywords with very low search volumes" cannot be discovered or forecast, yet the same searches appear in Search Console as long as enough people make them to escape anonymization. A model can sort this list well, because every row in it is real.

How do you expand a keyword list without inventing demand?

Add candidates from sources that carry a count, and treat anything a model suggests as unverified until a tool has measured it.

Google Keyword Planner is free, but it lives inside Google Ads: Google says you must complete account setup, enter billing information and create a campaign before you can use it. It has two modes, according to its guide: "Discover new keywords" offers suggestions from a seed term or a website, and "Get search volume and forecasts" measures a list you already have. Paste your Search Console queries, and anything a model suggested, into the second.

The People Also Ask box on Google's results gives you questions in searchers' own words. It carries no counts, so use it to decide what a page should answer, not which pages to build.

The paid tools now generate ideas with AI too, attached to their own estimates. Ahrefs' AI Suggestions, which "helps you quickly generate new keyword ideas," is available on Standard plans and above, according to its list of AI features. Semrush's Keyword Magic Tool, which draws on a database of about 27 billion keywords, has an "AI-powered filter bar": enter a domain and it surfaces opportunities for that site, with a Personal Keyword Difficulty score that Semrush says is calculated using large language models judging how relevant the domain is to each keyword.

A chat model is fine for brainstorming. Ask it for 50 ways a buyer might phrase a need, run them through Keyword Planner, and keep the ones that come back with a volume or already have Search Console impressions. A suggestion with neither may still be real demand, too rare to measure, but you have no evidence for it yet, so it goes to the bottom of the list rather than into a plan.

How do you cluster keywords: by SERP overlap or by meaning?

By SERP overlap, whenever the decision is one page or two. If two searches return many of the same URLs in Google's top results, Google is answering them with the same documents and one page can compete for both. If they share few or none, Google has split them, however similar the words.

Grouping by meaning, which is what a language model does when you ask it to cluster, answers a different question, whether the phrases say similar things, and usually Google agrees. Where the two part company is intent. "Wrongful dismissal lawyer toronto" and "wrongful dismissal ontario" sit side by side in meaning, but the first searcher wants to hire someone and the second wants to know their rights, and a page built for one serves the other badly. You are trying to rank in Google's results, so Google's grouping is the one that counts.

The tools that decide page boundaries work this way. Semrush says its clustering "compares the top 10 organic results for each keyword" and groups two keywords "when two keywords return similar URLs." Its Keyword Strategy Builder, which needs a paid SEO Toolkit subscription, takes up to 10,000 keywords a list. Keyword Insights groups keywords that share at least three URLs by default, and now compares the top seven results because, its documentation says, "Google no longer consistently shows 10 traditional links." Ahrefs' Keywords Explorer clusters by Parent Topic, according to a 2024 Ahrefs post: it takes the page ranked first for a keyword and finds the keyword sending that page the most traffic. Ahrefs itself calls this "arguably less accurate than some keyword clustering tools because it clusters by Parent Topic," that is, from one page rather than the whole results page, and faster for the same reason.

A model is useful one step earlier: sorting a long, already-measured list into rough piles, so you run overlap checks only on the pairs that matter. This prompt does that and keeps the numbers out of its hands.

Below are two files for the same topic: a Search Console export
(query, clicks, impressions, position) and a Keyword Planner export
(keyword, avg_monthly_searches). Join them on the query text.

Sort the queries into candidate groups by the job the searcher wants
done. For each group return: a short name, its queries, and the sum
of impressions and of avg_monthly_searches taken from the files.

Rules:
- Use only numbers that appear in the files. If a value is missing,
  leave it blank and write "missing". Never estimate one.
- Do not add queries that are not in the files.
- List every pair you are unsure belongs together. I will decide
  those by comparing their top 10 Google results, not by meaning.

[paste both files]

Here is the overlap check for Harbour & Finch LLP, a sample Toronto law firm. The queries are realistic; the shared-URL counts are sample data, not a measurement.

Pair of queriesClose in meaning?Shared URLs in top 10 (sample)Decision
wrongful dismissal lawyer toronto / wrongful termination lawyer torontoYes7One page: the dismissal service page
wrongful dismissal lawyer toronto / wrongful dismissal ontarioYes1Two pages: the service page and a guide to employees' rights
what is constructive dismissal / constructive dismissal ontarioYes6One guide
severance pay ontario / severance lawyer torontoPartly0Two pages: a guide and the severance review service

Grouped by meaning, these seven queries could easily land on one employment law page. The sample counts send them to five: two service pages and three guides, split by what the searcher wants to do next. Check pairs near a threshold by reading both results pages; two or three shared URLs is where a cluster flips. Our topic cluster guide covers thresholds and how to treat chains of queries that link A to B and B to C.

How do you classify search intent with AI?

Give a model the list and a fixed set of labels, make it mark the queries it is unsure about, and check those against the live results page. Intent labelling suits a model well: the answer depends on reading a phrase, and nothing has to be looked up.

Use the labels the tools use, so your map lines up with their exports. Semrush tags each keyword in the Keyword Magic Tool with one of four types: informational, navigational, commercial and transactional. Ahrefs' AI Search Intent, on Standard plans and above, works from the other side, grouping the results on the page by intent. That approach is closer to the truth. For ranking purposes, search intent is what the results page shows Google believes the searcher wants, and a label taken from the words alone is a prediction of it.

For a quick pass without a model, our free search intent classifier sorts a pasted list into the same four types by matching signal words such as "how", "best" and "buy", and gives each query a confidence score. It does not read the results page either, so treat its low-confidence rows the way you treat the model's: open the results and look.

You are labelling search queries for a keyword map. Below is a CSV
with one column, query. For each row add four columns:
- intent: informational, commercial, transactional or navigational
- page_type: the page that would satisfy it (guide, comparison,
  service page, product page, tool, local page)
- confidence: high, medium or low
- reason: under 12 words

Rules:
- Do not add, remove or reword any query.
- Do not add search volume, difficulty, CPC or any other number.
  If the file has number columns, copy them through unchanged.
- Mark confidence low when a query could reasonably be read two
  ways. I will check those against the live results page.
Return CSV only.

[paste queries]

How do you map clusters to pages and check for cannibalization?

Give every cluster one owning URL, existing or planned, then check Search Console for queries where two of your pages already split the impressions.

The map is a table with one row per cluster: the main query, the other queries in the cluster, the intent, the measured demand (Search Console impressions and Keyword Planner volume, each in its own column), the owning URL and the next action. Nothing a model generated goes in the demand columns.

Keyword cannibalization is two of your own pages competing for one search so that neither holds a good position. To check, filter the Performance report to a query and open the Pages tab. Two pages with impressions is not a problem in itself; two pages swapping places week after week while neither ranks well is. The fixes are to re-scope one page to a different cluster, or to merge the weaker page into the stronger and redirect it permanently. Semrush lists keyword cannibalization analysis among the features of its Pro+ plan, at $299 a month on monthly billing (Semrush's pricing page, November 2026). For a site with a few hundred queries, the Search Console check and a spreadsheet do the same job.

On a large export a model can do the joining: paste the query-and-page export with the map and ask for every query where more than one URL has impressions, numbers copied from the file. That is filtering, and a pivot table gives the same answer, which makes the model's output easy to verify.

How do you turn a keyword cluster into a content brief?

Give the model the measured cluster, the page that will own it, the pages that rank now and the questions searchers ask, and have it draft a brief that names the source of every fact. Of all the steps, this is where a model's drafting saves the most time with the least risk, because every input is in front of it.

Write a content brief for one page from these inputs:
1. The cluster: [queries, with impressions and avg monthly searches]
2. The page that will own it: [URL, current title and headings, or "new"]
3. The top five ranking pages for the main query: [URLs and their H2s,
   copied from the live results]
4. People Also Ask questions for the main query: [paste]

The brief must give: the main query and a page title; each H2 as the
question it answers, in the searcher's words; which queries each
section covers; the facts each section needs and the kind of source
each must come from; what the ranking pages leave out; internal links
to and from the page.

Rules: every number in the brief must come from the inputs, with the
input named. Where the brief needs a figure the inputs do not contain,
write NEEDS DATA instead of a number.

The NEEDS DATA rule does more work than it looks. Asked for a brief without it, a model that guesses when uncertain tends to fill the gaps with plausible figures: a market size, an average fee, a statistic with no source. With the rule, the gaps become a research list for the writer.

AI keyword research tools: what they cost and what the AI does

As of November 2026, Google Keyword Planner is free with a Google Ads account, Ahrefs starts at $29 a month, Keyword Insights at $58 and Semrush's SEO plan at $139, and each vendor puts its AI features at a different price point.

Keyword research costs nothing in Google's own tool and $29 to $249 a month in the paid ones; Ahrefs' AI suggestions start at the top of that range.
Google Keyword PlannerFree
Ahrefs Starter$29
Keyword Insights Basic$58
Ahrefs Lite$129
Semrush SEO$139
Ahrefs Standard$249
Ahrefs plan that adds AI SuggestionsOther plans
Monthly price in US dollars, monthly billing, November 2026
Google Keyword PlannerFree
Ahrefs Starter$29
Keyword Insights Basic$58
Ahrefs Lite$129
Semrush SEO$139
Ahrefs Standard$249
Lowest price of each plan on monthly billing, as listed on Semrush's pricing page, Ahrefs' pricing page and Keyword Insights' pricing page, checked 2 November 2026. Keyword Planner is free but needs a Google Ads account with billing details and a campaign (Google).
ToolPrice, November 2026AI and clustering featuresFits
Google Keyword PlannerFree, inside Google AdsKeyword ideas from a seed term or URL; Google's rounded volume averagesEveryone, as the volume source
SemrushFree account, 10 reports a day; SEO plan $139 a month ($117.33 billed yearly)AI-powered domain filter, Personal Keyword Difficulty, intent labels, Keyword Strategy Builder clustering by top 10 overlapTeams wanting volumes, intent and SERP clustering in one suite
AhrefsStarter $29; Lite $129; Standard $249 a monthAI Suggestions, AI Search Intent, AI Translations and keyword clusters from Standard; Parent Topic clusteringTeams already using Ahrefs for links and site data
Keyword InsightsBasic $58 a month (10,000 credits); $1 for a 7-day trialClustering by shared URLs, one credit per keywordSorting large lists into pages
AIO Copilot search intent classifierFreePattern matching into four intent types, with a confidence scoreA quick first pass on intent

Semrush suits teams that want volumes, intent labels, domain-specific difficulty and SERP-based clustering in one place. The SEO plan costs $139 a month, or $117.33 a month billed yearly, and the free account allows 10 reports a day across its main tools, according to its pricing page. Because the Keyword Strategy Builder's clusters come from comparing top 10 results, they are a sound basis for page decisions rather than only topic labels.

Ahrefs sorts its AI features by plan. Starter, at $29 a month, includes Keywords Explorer with 200 credits a month, 250 rows per report, one month of history and no exports, and renews monthly only, according to its help page on the plan. Lite costs $129 a month. AI suggestions, keyword clusters and search intents are listed among the features that begin on Standard, at $249 a month (Ahrefs' pricing page, November 2026; annual billing saves up to 17%). Our comparison of Ahrefs and Semrush goes further into how the two suites differ.

Keyword Planner is the only volume source that comes from Google itself. Use it for the numbers, knowing they are rounded averages that group close variants together.

Keyword Insights does one job, SERP-overlap clustering, and prices it by the keyword. The Basic plan is $58 a month for 10,000 credits, clustering costs one credit per keyword, and a 7-day trial costs $1 for 5,000 one-time credits, according to its pricing page in November 2026. For a site that needs a few thousand keywords sorted into pages once a quarter, that is cheaper than a suite.

The choice follows the size of the job. A single business can run this whole process with Search Console, Keyword Planner and a chat model that is never asked for numbers. Agencies and content teams producing many pages get their money's worth from a suite for the volumes and a SERP-based clusterer for the page decisions, whether that is Semrush's builder or Keyword Insights.

What is the most affordable keyword research software?

Google's own. Search Console and Keyword Planner cost nothing, and between them they cover the searches your site already gets and Google's own volume figures.

Among the paid tools here, the cheapest entry is Ahrefs Starter at $29 a month, with the tight limits above. Semrush's free account gives 10 reports a day, enough to check volumes for a short list, and Keyword Insights' $1 trial will cluster up to 5,000 keywords once. The costlier mistake is building pages for demand a model made up.

Are there voice search keyword research tools?

Not in any measured sense. Google does not report spoken searches separately to site owners: Search Console's search type filter covers web, image, video and news, with no voice option, so spoken queries arrive in the export as ordinary text.

What gets sold as voice keyword research is, in practice, question research: long, conversational queries phrased as full questions or ending in "near me". Pull those out of Search Console with a regex filter, add the People Also Ask questions, and measure them like any other keyword. Our guide to voice search optimization separates what is known about spoken answers from what is marketing.

Is an AI keyword research service worth paying for?

Only if it can show where every number came from and how it decided which searches share a page. Before paying an agency or a tool for AI keyword research, ask four questions:

  • Which source does each volume come from: Keyword Planner, Search Console or a named tool?
  • Are clusters decided by comparing results pages, or by meaning alone?
  • Does the plan check the site's existing pages, and its Search Console data, for searches two pages already compete for?
  • Did a language model generate any number in the deliverable?

The right answer to the last one is no. The research in AIO Copilot's monthly plan is built on that rule: keyword research joined to a site's own Search Console and Google Ads data, with estimates kept apart from measured figures, ending in a page brief rather than a spreadsheet.

Whoever builds the list, the test is the same. Pick three numbers in it at random and ask where each came from. If the answer is a file, keep working. If the answer is the model, go back to Search Console and start again.

Questions

Is ChatGPT good for keyword research?

It is good at the parts that are about language: suggesting phrasings, sorting a pasted list into topics, labelling intent and drafting briefs. It is not a source of search volume, difficulty or cost per click, because it has no keyword database, so figures it gives unprompted are guesses. Connect it to a keyword tool through MCP or paste in an export, and take every number from that file.

How accurate is search volume in Ahrefs and Semrush?

Both vendors describe their figures as estimates: 12-month averages modelled from Keyword Planner, clickstream and other third-party data. In Ahrefs' 2021 study of 72,635 keywords, its estimates were within 50% of Search Console impressions about 60% of the time, and Keyword Planner's 45% of the time. Use them to compare keywords with each other, not to forecast visits.

Is Google Keyword Planner free?

Yes. Google describes it as a free resource, but it sits inside Google Ads: you need a completed account setup, including billing information, and a campaign. Its volumes are rounded 12-month averages, and in 2016 Google said accounts with lower monthly spend may see ranges, such as 100 to 1K, instead of a figure.

How many keywords should one page target?

As many as Google answers with the same pages. If two searches share several of their top results (Keyword Insights uses three shared URLs as its default), one page can serve both. If they share none, they need separate pages even when they mean almost the same thing. There is no fixed number per page.

Does Search Console show every keyword a site ranks for?

No. Google leaves out anonymized queries, those not issued by more than a few dozen users over two to three months, and the interface exports at most 1,000 rows. The API returns up to 50,000 rows a day per site per search type, and the bulk export to BigQuery includes all performance data except anonymized queries.

New studies, by email

Our own studies, and what each new release changes for SEO, AEO and Google Ads.

Or get your price for the app