Best AI keyword research tools compared, and the demand an AI keyword tool still cannot measure

Two completely different products are sold as AI keyword research tools, and the difference decides whether you are getting new information or a nicer interface on the same database.

The first kind puts a language model in front of an existing keyword index. You describe your business in a sentence and it returns grouped suggestions, clusters, intent labels and draft briefs. The data underneath is the same keyword database it always was; the model is doing the expansion and the grouping, which is genuinely useful work and used to take a person a day.

The second kind reads the assistants. It puts questions to ChatGPT, Gemini, Perplexity, Claude or Google AI Overviews and records what came back, which names were mentioned and which pages were cited. That is a different data source, and it is the only one of the two that can see demand a keyword database has never held.

The criterion on this page: is the vendor adding a model, or adding a data source, and does it say which?

What the vendors publish, read on 21 October 2026

Ahrefs is the most useful page in this category to read because it explains its own method in one paragraph. ahrefs.com/brand-radar states "Every prompt is modeled from real user searches in Ahrefs keyword database" and describes the process: Ahrefs "takes real queries from its keyword database, expands them into natural questions via People Also Ask and semantic fanout, then runs the resulting 493 million+ prompts through each AI platform and stores the responses". It publishes per surface counts for that index, with AI Overviews at 277,456,152 prompts, AI Mode at 100,797,074, Gemini at 28,930,170, Copilot at 28,921,226 and Perplexity at 28,837,146. It publishes its own unit economics too: "Claude consumes 8 checks per update", an "AI Visibility Index" at "$199 /mo" for "83 prompts/day" with "+2,500 checks/month" and an "Overage $0.020/check", and Custom Prompts from "$50/mo". It also offers to "Discover the fanout queries AI generates behind the scenes when processing a prompt".

So Ahrefs is doing both kinds at once: a model expanding a keyword database into questions, then real responses collected from the platforms. The expansion is modelled and the response is observed, and the page says which is which.

SE Ranking publishes the second kind as named trackers: an AI Overviews Tracker, an AI Mode Tracker, a ChatGPT Tracker and a Gemini Tracker, plus ChatGPT ads and AI Mode ads trackers, in its product navigation on seranking.com/pricing.html. Its metering puts both units side by side: "2k keywords and 100 prompts to track daily" at "$129.00 /mo", and "5k keywords and 250 prompts to track daily" at "$279.00 /mo". It also publishes an AI Visibility API and "SEO Skills for Claude".

Moz publishes an "AI search and visibility" product "For anyone wanting to track and monitor visibility in AI-generated search", recommended for "Prompt research, AI sentiment analysis, LLM citation tracking, Monitoring brand mentions, Competitive benchmarking", at "Starts at $45/month" on moz.com/products. That is the lowest published entry point for prompt research of any vendor named here.

Semrush leads its own free tools list with an "AI Visibility Checker" on semrush.com/pricing, lists "Prompt Research" in its product navigation beside Keyword Research, and publishes "Unlimited prompts to track daily" against one of its AI visibility rows. Its entry SEO tier at "$139" includes "Track performance in AI search".

DataForSEO sells the second kind as raw data: an "AI Optimization API, LLM visibility, keyword demand, responses, and generative-engine data" with "LLM Responses, LLM Scraper, AI Keyword Data" named inside it, and a free AI Visibility Tracker tool, on dataforseo.com/pricing, on a "$50 minimum top-up" pay as you go model.

Nightwatch publishes "Citation Intelligence, Track when AI recommends your brand" and a "Prompt Research" product on nightwatch.io/pricing, metered as "500 AI prompts" on its "399 euros a month" agency tier and unlimited above it.

The ranking, on the stated criterion

On whether the vendor says plainly which half is modelled and which is observed, the order on 21 October 2026 ran: Ahrefs first, and it is not close, because it publishes the method, the per surface prompt counts, the per engine check cost and the overage rate on one page. SE Ranking second, because it publishes both units per tier and names each surface it reads as a separate tracker. DataForSEO third, because it sells the responses as an endpoint and names the product honestly. Nightwatch fourth, because the prompt allowance is published per tier. Moz fifth on method disclosure and first on entry price. Semrush sixth, because "Unlimited prompts to track daily" is a strong offer and the surfaces behind it are not enumerated in the rendered text we read.

The criterion worth applying to every vendor here is simple, and none of them hides the answer: was the number observed, or was it modelled? Ahrefs says it outright on ahrefs.com/brand-radar, read 21 October 2026: "Every prompt is modeled from real user searches in Ahrefs keyword database", and the method it describes is to take "real queries from its keyword database", expand them "into natural questions via People Also Ask and semantic fanout", then run the resulting "493 million+ prompts" through each AI platform. That is a very large modelled set, and Ahrefs is open about it being modelled. Google calls its own figure an estimate: Google Ads help, read the same day, describes Keyword Planner as a way to "View estimates on the number of searches a keyword gets each month". Neither of those is a flaw. Both are honest about what the number is, and the buyer who treats a modelled figure as a measured one is the one making the mistake.

The thing an AI keyword tool of the first kind structurally cannot find

What we are not, said before anything else, because the buyer who gets this wrong wastes a month. AI Knows Us does not sell a keyword tool, a rank tracker or a SERP API. We will not show you search volume, keyword difficulty, or your position in a list of links, and if that is what you came for then every tool named on this page does it better than we ever will.

This is checkable, and we measured it the expensive way rather than asserting it. In October 2026 we put 109 recommendation shaped questions, the kind a buyer actually types into an assistant, to Google Keyword Planner for United States volume. Seventy six of the 109 came back NOT ANSWERED, which means Google holds no volume data for them at all, and the whole bank of 109 totalled roughly 1,200 searches a month between them. Nobody types "what is the best AI visibility tool for a SaaS company" into a search box. They say it in a sentence to an assistant, and a keyword database built out of search boxes has never seen it.

That is the whole reason the second kind of product exists, and it is why a model sitting on a keyword database can be excellent at grouping and still blind to the question your buyer actually asked. Expansion makes the database more readable. It does not add a query nobody typed into a search box.

What we do sell is the reading of a demand no keyword tool has ever seen. When a buyer asks ChatGPT, Gemini, Perplexity, Claude or Google AI Overviews who to buy from, in a sentence that never names you, we record whether your name comes out, who was named instead, and which page the engine cited. Then we run the content and placement work that changes it. A credit is one buyer question put to one app once. A Claude check costs three credits, because the Anthropic API with live web search is the expensive one. Most checks cost one.

Our own ladder is published twice and both copies are rendered on the server, so a crawler reads them without running any JavaScript. aiknowsus.com/llms.txt is plain text, and aiknowsus.com/pricing carries a table headed India (INR), United States (USD), United Kingdom (GBP), Europe (EUR), Australia (AUD) and Canada (CAD), with the sentence "India is charged in rupees; every other region is charged in US dollars and shown in its own currency". Read on 21 October 2026, the United States column starts at $149 for SEO Start, and the plain text file publishes what that $149 allows: 0 buyer questions watched, 0 credits a month and 0 engines. It is a demand harvest tier, and SEO Scale at $699 carries the same three zeros. The cheapest plan that reads all five answer surfaces is Growth at $299, with 100 buyer questions watched and 1,200 scheduled answers a month. Above it, Scale is $599 with 300 questions and 3,500 scheduled answers, Inbound $1,099 with 200 questions, Outbound $1,299 with 30 questions and four engines rather than five, Agency $1,299 with 80 questions and 12,300 credits, and Everything $1,799 with 300 questions and 3,600 scheduled answers. A top up of 1,000 credits is $49. Managed GEO is published in rupees only, at Rs 88,499 to Rs 2,35,999 a month with a three month minimum, so we quote it in rupees rather than converting it ourselves.

Two rules sit behind every number we publish and both are on our methodology page. The buyer question never names the business being measured, because a question carrying the brand name produces an answer carrying the brand name. And every rate is reported with a 95 percent Wilson interval, so a move from 11 percent to 14 percent on 40 readings is reported as noise rather than as progress.

Where a rival is ahead of us, here it is plainly. Ahrefs Brand Radar, read at ahrefs.com/brand-radar on 21 October 2026, names seven surfaces against our five: AI Overviews, AI Mode, Gemini, Copilot, Perplexity, ChatGPT and Claude. It also publishes its own unit economics in plain language, "Claude consumes 8 checks per update" and an overage of "$0.020/check", which is more disclosure than we publish. SE Ranking, read at seranking.com/pricing.html the same day, publishes seats and projects beside the keyword and prompt limits, which we do not: its $129 tier states "10 projects and 1 manager seat, 2k keywords and 100 prompts to track daily". Nightwatch, read at nightwatch.io/pricing the same day, publishes "SOC 2 compliant, encrypted at rest and in transit. SSO on Enterprise" and white label client reports. Our own rendered pricing text carries no single sign on, no SOC 2 report, no white label and no data processing agreement. Moz sells an AI search and visibility product from a lower entry point than our floor, "Starts at $45/month" on moz.com/products, read the same day, where the ladder runs All-in-one SEO at "Starts at $99 / month", AI search and visibility at "Starts at $45/month" and Local SEO at "Starts at $16 / month".

On this page's criterion we put ourselves first on one narrow and checkable thing, and it is the distinction the page is built on: every prompt we report is a buyer's sentence put to the live app and recorded with the answer, so our unit is observed rather than modelled, and the question never names the business being measured. Where Ahrefs beats us is scale and disclosure of unit economics: 493 million prompts, seven surfaces against our five, and a published per check overage rate. We are the smaller, narrower, observed set, and both of those words matter.

For the vendor by vendor comparison of the second kind of tool, we keep a separate page on the tools that track AI search mentions, where coverage is published per tier rather than per vendor. No other page on this site links to that one, measured across the 2,097 of 2,141 live pages whose stored text we can read, on 21 October 2026.

What buyers ask about AI keyword research tools

Does an AI tool find keywords a normal tool cannot? It finds phrasings and groupings faster. It does not find a query that is absent from the database underneath it.

Should prompts replace keywords? No. They are two demand channels with different units, and the vendors metering both separately have it right.

How many prompts is enough to start? Twenty buying questions, written in the words a customer would use, with your brand name left out of every one of them.