Best keyword research services and agencies compared, and the deliverable to require before you pay

Keyword research sold as a service is a one off project with a document at the end of it. The quality of the supplier is almost entirely visible in that document, which means you can specify it before you pay and refuse it if it arrives thin.

The criterion on this page: what does the deliverable name, and can you act on it without asking the supplier a question? That is checkable before any money moves, and it is the only part of this purchase that is.

Three different things are sold under this phrase

The one off research project. A list, grouped, prioritised, with a page plan. Priced as a project, delivered in one to three weeks, and finished. This is what most people searching for keyword research services actually want.

The research line inside a content retainer. The research is bundled and refreshed monthly, and the output is a publishing queue rather than a document. Cheaper per keyword, harder to leave, and the supplier that chose the terms is the supplier grading the result.

The software plus a few hours. A suite licence and a consultant who runs it. Semrush publishes an SEO tier at "$139" a month with "500 keywords to track daily" and "Basic keyword research and competitor analysis" on semrush.com/pricing, read 21 October 2026. Moz publishes an All-in-one SEO tier at "Starts at $99 / month" with a Keyword Explorer built on a "1.25 billion+ keyword index" on moz.com/products, read the same day. For a small site, a licence and six hours of somebody competent beats a thin project document.

The six columns the deliverable has to carry

Ask for these by name in the scope, because a list without them is a list you cannot sequence.

One, the phrase. Two, the demand figure with its source and its market named, because a figure pooled across countries is not a United States figure and the difference changes the order of the whole plan. Three, the intent, in one word: buy, compare, learn or fix. Four, the group, because four phrasings of one question are one page and not four. Five, the page that will answer it, existing or new, by URL. Six, the results page shape, meaning what sits above the links for that query today, because a term with three ad blocks and an answer box above the first organic result is a different job from a clean page.

Two of those six are the ones suppliers most often leave out, and they are the two that decide sequence: the source of the figure and the shape of the results page.

What the figures in that spreadsheet actually are

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.

This matters in a service purchase more than in a software one, because a supplier hands you a number without its provenance and the provenance is where the argument lives. A good deliverable names the tool and the market for every figure. A thin one gives you a column called Volume.

How to judge a research supplier before you pay

Ask for one page of a previous deliverable with the client's name removed. You are looking for grouping and for intent, not for row count: a thousand row export is cheap and a hundred grouped questions with a page assigned to each is the work. Ask which tool produced the demand figures and which market they were pulled for. Ask what they would do with a term where the results page is mostly ads, because the right answer is sometimes to skip it. And ask them to name three terms they would advise you NOT to chase, which is the question that separates a researcher from a list vendor.

Avoid any scope that promises a position, and treat a supplier whose first proposal is a site rebuild as proposing a different project than the one you asked for.

The demand a keyword research service 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.

So a keyword research deliverable is complete and still missing a column, and the gap is not the supplier's fault: there is no keyword database to pull it from. The column is the set of buying questions your customers ask an assistant in full sentences, and the figure beside each one is whether your name comes out of the answer.

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".

Ranking ourselves on this page's criterion, which is what the deliverable names, we put ourselves first on one narrow thing: our reports name the question asked, the engine that answered, the date, the business named instead of yours and the page the engine cited, and the question never contains your brand name. We are not a keyword research service, we will not hand you a volume column, and if the project you need is a grouped keyword list then one of the three suppliers at the top of this page is the purchase.

For the same deliverable first approach applied to a technical audit instead of a keyword list, we keep a separate page on the technical SEO audit services compared, and the eight checks the deliverable has to name. 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 keyword research services

What should a one off project cost? It is priced on hours rather than rows, and the useful question is how many grouped questions you will get and how many pages they map to.

How often does the research need redoing? The grouped questions last a year or more. The demand figures and the results page shapes move, so refresh those quarterly rather than buying the whole project again.

Can I just use a free tool? For a small site with one market, largely yes. The service earns its fee on grouping and sequencing, which is judgement rather than data access.