Keyword difficulty is the number most SEO tools put next to a keyword to say how hard it is to rank for. It is useful, widely misread, and only half the picture now that many buyers ask an AI assistant instead of scanning ten blue links.
This post explains what the score measures, why tools disagree, how to turn a generic score into a judgement about your own site, and how to read difficulty in AI answers, with examples from our own tracked questions.
What does keyword difficulty actually measure?
It measures the strength of the pages ranking today, as a tool models it. Most tools lean on links. Ahrefs describes its method plainly: it pulls the top 10 ranking pages for a keyword and looks up how many websites link to each of them, then maps that onto a 0 to 100 scale. In its own words, around 0 to 5 means the top pages barely have backlinks, around 50 means a couple of hundred, and 90 or more means thousands.
Other tools blend in their own authority metrics, the presence of SERP features, or the share of big brands on page one. None of them is a Google number. Google does not publish a difficulty score, so every difficulty figure is a model of the results page, built on the tool's own link index.
That has three consequences:
- Scores are not comparable across tools. A 40 in one tool and a 40 in another are different measurements.
- The score is absolute, not personal. It describes the average site's chances, not yours. Ahrefs makes the same point: you still have to judge your own site's authority against it.
- It is blind to everything except links. Intent, page type, freshness and brand dominance all affect how hard a query really is.
How do you judge personal difficulty?
Turn the generic score into a judgement about your site with four checks.
Compare it to your own link profile
If the pages ranking for a query have links from a few hundred sites and yours has links from a few dozen, the query is hard for you whatever the score says. If your strongest pages already rank for similar queries, it is easier than it looks.
Look at what kind of page ranks
Open the results. If Google shows tools, calculators or product pages and you plan a blog post, the difficulty is not links; it is that you are building the wrong thing. Match the page type to the intent, which our search intent post covers.
Check who dominates
Results where one brand holds several positions, or where one platform's own help pages answer the question, are harder than their link counts suggest. You are not competing with pages; you are competing with the default.
Check your topical depth
A site that already covers a topic well, with owner pages linked to each other, ranks new pages on that topic faster than a site writing about it for the first time. Difficulty falls as your coverage of the neighbouring questions rises.
How should keyword difficulty shape what you write?
Use it to sequence, not to select.
- Start where you can win now. Low personal difficulty, high intent: specific questions close to a purchase that your site is already credible on.
- Build towards harder terms. Each owner page you win on a narrow question strengthens the cluster around a broad one.
- Skip what you cannot win this year. If one dominant source owns the answer, put the effort elsewhere and come back when your coverage is deeper.
Revisit the scores every quarter. Difficulty changes as competitors publish, as results pages gain or lose features, and as your own site earns links, so a keyword you parked in spring may be within reach by autumn.
The most expensive mistake is the opposite: chasing the highest-volume head term because it looks valuable, publishing a page with no chance, and concluding that SEO does not work.
What does difficulty look like in AI answers?
AI answers are not ten ranked links, so link-based difficulty does not transfer. The pages they cite overlap only partly with Google's first page (see why brands go missing from AI answers for the outside data).
A better measure for AI answers is concentration: how much of an answer's citations one source holds. We read it in our own product for our own site's tracked questions. These numbers are from the attensira.com workspace, read over the Attensira MCP on 4 October 2026, for 4 September to 4 October 2026.
The table at the top of this page shows four of those questions, each with the share of its citations held by its single top source.
Read it like a difficulty score:
- Concentrated questions are hard. When OpenAI's help centre holds about two thirds of the citations for "how to get cited by ChatGPT", no third-party page is going to displace it on ChatGPT. That question is the AI equivalent of a 90+ keyword for anyone who is not OpenAI. "What is generative engine optimization" is close behind: Google's developer documentation held 708 of the 1,915 ChatGPT citations, more than a third.
- Spread questions are open. On Perplexity, the most cited source for "best AI search visibility tools" held 45 of 577 citations. Many small pages share the answer, so a strong owner page or a few placements on the roundups can change it.
- Near misses are the cheapest wins. For "AI search visibility starting price" on ChatGPT, our own pricing page was cited 69 times across 57 answers and ranked 11th among the deciding sources by weight, behind G2 with 116. The page is already read. A rewrite that answers the price question in one liftable sentence is the lowest-effort move on our whole list.
The open case is easiest to see in a single answer. Asked for the best AI search visibility tools for agencies on 4 October 2026, Perplexity listed eight tools and cited nine roundup and vendor pages. Across 45 Perplexity answers to that question from 5 September to 4 October, the six sources that weighed most held between 17 and 33 of 468 citations each, and we were named in none of the 45. No page owns that answer yet, so one good placement counts.
The same question can also be hard on one engine and open on another. The concentrated ChatGPT answers above look very different on Perplexity, where community pages, videos and independent explainers share the citations. Judge difficulty per engine, never averaged. An average of a hard ChatGPT answer and an open Perplexity answer describes neither, and leads to a plan that fits neither engine.