Research · 5 September 2026
Keyword difficulty is a weak signal. I checked 12,512 search results pages to see how weak.
Every keyword tool sells a 0 to 100 difficulty score. I bought the actual first page of Google for 12,512 queries and counted who was standing on it. Here is the entire effect of the score, as 100 first pages at each end.
A first page a small site could get onto: none or one major brand on it.
The same test. Half of the pages the score called hopeless had room.
What I tested
I build small tools that work in the browser, and each one is a page that has to rank for a search like “typing test” or “subnet calculator”. Every keyword tool sells a number that claims to say how hard that will be. I was about to trust it with months of work, so I measured it instead.
I expanded seed phrases into 23,400 queries and bought the difficulty score and the search volume for 14,698 of them. Then the step almost nobody does: for 12,512 of those queries I bought the actual first page of Google results and had a script classify each one by counting how many of the top results belong to a major brand: Adobe, Canva, YouTube, Google Play, Apple, Microsoft, Amazon, Wikipedia, the big file converters, and so on. Three verdicts:
- Open. None or one major brand on the page. A small site is already standing there, so a small site can.
- Contested. Two.
- Locked. Three or more. Nobody without a brand is getting on.
For 4,065 of the sites standing on those pages I also bought the backlink profile, which let me measure something the score does not: the weakest site on each first page. The score is DataForSEO's, the same family of backlink-derived score the big tools sell; the results pages cost about a tenth of a cent each through the same API.
The score barely moves the odds
Of the 12,512 first pages, 64% were open, 16% contested and 20% locked. Split them by the score they were given and the bars barely leave the average.
The line is the average across every band. The bands differ from each other far beyond chance, and none of them gets far from it.
Table view
| Score | First pages | Open | Contested | Locked | 95% interval |
|---|---|---|---|---|---|
| 0 to 10 | 2,967 | 72% | 17% | 12% | ±1.6% |
| 11 to 20 | 2,160 | 65% | 16% | 18% | ±2.0% |
| 21 to 30 | 1,882 | 63% | 17% | 21% | ±2.2% |
| 31 to 40 | 1,605 | 66% | 16% | 18% | ±2.3% |
| 41 to 50 | 1,415 | 65% | 16% | 20% | ±2.5% |
| 51 to 60 | 868 | 58% | 15% | 27% | ±3.3% |
| 61 to 70 | 678 | 53% | 15% | 32% | ±3.8% |
| 71 to 100 | 937 | 49% | 18% | 33% | ±3.2% |
Read it the way a person choosing keywords would. Trust a low score and you are wrong 28% of the time. Skip a high score and you walk away from an open page 49% of the time. The same tool sells both numbers with the same confidence.
What the score cannot see
The difficulty score is built from backlinks. It cannot see a brand, and brands are what lock a page. Here is one first page the score called easy, and who was standing on the 2,468 locked pages overall.
Table view
| Site | Locked pages it appears on | Share of locked pages |
|---|---|---|
| youtube.com | 1,228 | 50% |
| adobe.com | 1,144 | 46% |
| play.google.com | 1,116 | 45% |
| canva.com | 961 | 39% |
| reddit.com | 576 | 23% |
| freeconvert.com | 498 | 20% |
| ilovepdf.com | 493 | 20% |
| smallpdf.com | 478 | 19% |
| cloudconvert.com | 417 | 17% |
| en.wikipedia.org | 325 | 13% |
| drawboard.com | 294 | 12% |
| apps.apple.com | 278 | 11% |
Now the reverse. “world clock”: 301,000 searches a month, difficulty 100. The score says impossible. Page one is timeanddate.com, nist.time.gov, worldtimebuddy.com, play.google.com. No brand in sight, and one of them holds its place on a handful of links.
The number that does predict something
If the score cannot tell you whether a first page has room, what can? The weakest site already standing on it. If a site with nine referring domains holds a slot, a slot exists that nine referring domains can hold. For the 5,043 open first pages with a measured weakest site, here is how many links that site had.
Median 151 referring domains. One open page in 8 is held by a site with ten links or fewer; on one page, a site with 30,042,975 referring domains sits beside one with 1.
Table view
| Weakest site has | Of every 100 open pages | Cumulative share |
|---|---|---|
| ten links or fewer | 13 | 13% |
| eleven to one hundred | 27 | 40% |
| one hundred and one to four hundred | 32 | 72% |
| more than four hundred | 28 | 100% |
| Median | 151 referring domains | |
Does the difficulty score at least know about the weakest site? Barely. The rank correlation between the score and that site's links is 0.11 across 4,980 pages.
The median climbs from 110 at a score of 0 to 10 to 338 at 71 to 100, and the middle half of every band overlaps the middle half of every other. The score and the weakest site are measuring different things, and the weakest site is the one that says whether there is a door.
Table view
| Score | Pages | Median links | 25th to 75th percentile |
|---|---|---|---|
| 0 to 10 | 1,333 | 110 | 20 to 328 |
| 11 to 20 | 874 | 164 | 39 to 535 |
| 21 to 30 | 710 | 134 | 50 to 404 |
| 31 to 40 | 720 | 167 | 68 to 405 |
| 41 to 50 | 541 | 144 | 44 to 553 |
| 51 to 60 | 318 | 234 | 84 to 453 |
| 61 to 70 | 223 | 234 | 36 to 502 |
| 71 to 100 | 261 | 338 | 90 to 622 |
Limits
- One vendor's score. Other vendors' scores are built the same way, from backlinks, so I would expect the same shape, but I have only measured this one.
- The pool is queries with tool intent above a volume floor, not a random sample of all search. Within it, every query had already passed the screens people actually use, which is the point.
- “Major brand” is my list. Move the line and the shares move; the data file carries the rule as it stood.
- A first page is a snapshot of one moment, one country, one device type. Pages move.
- Open means a small site can get onto the page. It says nothing about reaching the top three.
The data
The aggregates behind every number and chart on this page are one file, the same file the page is rendered from, so a figure quoted from here can be checked against it. It holds the band tables, the distribution of the weakest site's links, the domain counts and the examples above. It does not hold the per-query list.
Download the data as JSON