GoldieBench Blog · 6 min read
KGR Calculator And AI Models: Which Ones Build Keyword Tools And Landing Pages, With Real GoldieBench Scores
Landing and Web-OS task scores matched to the jobs around KGR research, and the point where a person has to decide.

A KGR calculator is a tiny one-page tool, and building tools like it, plus the keyword briefs that follow, is work AI models now do in a single shot.
Book a free SEO strategy session →
The ratio is allintitle results divided by monthly searches, used on phrases with 250 or fewer searches a month.
This page matches the jobs around KGR research to GoldieBench scores, reading every number off live goldiebench.com pages on 12 November 2026.
If you want people choosing your keywords instead of models, book a free strategy session with Julian's team.
KGR calculator
KGR = allintitle results / monthly searches. Use it for keywords with 250 or fewer monthly searches.
Under 0.25 = great. 0.25 to 1 = ok. Over 1 = hard. Runs in your browser only; nothing is sent anywhere.
What GoldieBench Measures, And What It Leaves Out
The GoldieBench methodology gives every model the same fixed prompt, one shot, with no iteration and no best-of-N picking.
The output has to be one self-contained HTML file that runs in a browser.
That matches a calculator page closely.
What the bench does not test is keyword judgement, search volume or ranking, so none of the scores below say anything about which phrases will rank.
The KGR Jobs A Model Can Take On
- Tool pages: a working calculator or checker in one HTML file.
- Landing pages: the page you publish for a low-ratio phrase.
- Sheets and formulas: the sorting sheet with bands and intent labels.
- Briefs: titles, outlines and FAQs for each phrase that passes.
The bench has scores for the first two, through its Page category tasks.
Overall Leaderboard Averages For The Shortlist
These averages come from the GoldieBench models page.
| Model | Avg /10 | Tasks scored |
|---|---|---|
| Fusion | 8.59 | 47 |
| Claude Opus 5 | 8.27 | 50 |
| GPT-5.6 Sol | 8.16 | 50 |
| Grok | 8.09 | 47 |
| GLM-5.2 | 7.77 | 47 |
| Gemini 3.6 Flash | 7.08 | 50 |
An overall average mixes games, simulations and visuals, so the task-level scores below matter more for page work.
Landing Page Scores: The Page You Publish
The Landing task asks for a modern marketing landing page in one shot.
The task page lists a field average of 8.11/10 and names Fugu Ultra as the winner.
| Model | Landing score |
|---|---|
| Fusion | 9.0/10 |
| Fugu Ultra | 9.0/10 |
| GLM-5.2 | 9.0/10 |
| Grok | 9.0/10 |
| GPT-5.6 Sol | 8.7/10 |
| Gemini 3.6 Flash | 8.6/10 |
| Claude Fable 5 | 7.2/10 |
Book a free SEO strategy session →
For a KGR page, the design is the easy part; the answer in the first two sentences is what the model needs to get right.
Web-OS Scores: Interactive Tools That Actually Work
A calculator needs inputs, a button and an output that respond.
The closest bench test is the Web-OS Desktop task: a working desktop with windows, a dock and Notes, Paint and Terminal apps.
Its task page lists an average of 7.98/10, with Fusion as the winner.
| Model | Web-OS score |
|---|---|
| Fusion | 9.0/10 |
| Grok | 9.0/10 |
| GPT-5.6 Sol | 8.4/10 |
| Claude Fable 5 | 8.0/10 |
| Gemini 3.6 Flash | 8.0/10 |
| Kimi K2.7 | 8.0/10 |
| GLM-5.2 | 7.5/10 |
Fusion and Grok sit at the top of both page-style tasks, which makes them the first pair to try for a one-shot calculator build.
A One-Shot Prompt For Your Own Calculator
Build one self-contained HTML file: a KGR calculator.
Inputs: allintitle results (0 or more), monthly searches (1 or more).
Output: KGR = allintitle / searches, rounded to 3 decimals.
Bands: under 0.25 great, 0.25-1 ok, over 1 hard.
Show a note when searches are above 250.
No external requests. Accessible labels. Works on mobile.
Run the same prompt through two of the top scorers and keep whichever passes your own test cases.
Test Cases Every Model-Built Calculator Must Pass
A one-shot build can look perfect and still divide wrongly.
Run these five checks before you publish anything a model produced.
| allintitle | Searches | Expected result |
|---|---|---|
| 20 | 100 | 0.2, great |
| 50 | 100 | 0.5, ok |
| 300 | 200 | 1.5, hard |
| 10 | 400 | 0.025, plus an over-250 note |
| 5 | 0 | A clear error, not a crash or "Infinity" |
The last row catches most weak builds, because division by zero is the edge case models skip.
Also check the page works on a phone and that the result is announced to screen readers.
Book a free SEO strategy session →
Sheets And Briefs: No Bench Score, So Test It Yourself
GoldieBench has no task for spreadsheets or keyword briefs.
So no score on this page should be read as proof that a model writes better briefs.
The fair test is simple: give two models the same phrase and the same brief prompt, then compare which outline goes deeper than the current results.
Keep the winner as your default and re-check whenever a new model joins the leaderboard.
What The Inputs Really Are
No model fixes soft inputs.
Google's Keyword Planner help calls its volume figures estimates and says very low-volume phrases may be missing.
Google's search operators overview says operators sit within indexing and retrieval limits, and allintitle: isn't on the list in Google's Refine Google searches page.
KGR itself is a community rule of thumb, with no Google page behind it.
Pairing Two Models For One KGR Page
A simple split works well for most sites.
Let the strongest page builder you have access to produce the layout, the calculator or the comparison table.
Then hand the draft copy to a second model and ask it to cut anything that repeats the keyword unnaturally or promises a result.
Google's spam policies count keyword stuffing as spam, so that second pass is worth the extra minute.
Finally, a person reads the page against the live results and adds the detail no model can know, such as a photo or a real customer question.
The bench scores tell you who should do the build; they do not replace that last read.
Where Models Stop And People Take Over
A model can build the tool, the sheet and the brief.
It cannot decide which phrases bring buyers, or whether the live results page is beatable.
Google's hiring-an-SEO page says no one can guarantee a #1 ranking, and that holds whichever model drafted the page.
Google's AI features page adds that normal SEO best practice applies to AI Overviews and AI Mode, so the page still needs to be indexed and useful.
Our post on ranking in Google AI Overviews goes further on that side.
Book a free SEO strategy session →
For brief prompts, the free 200+ AI SEO Prompts pack is a quick start, and the free link building book covers getting other sites to cite the page.
The Pick For Each KGR Job
- Calculator or checker tool: Fusion or Grok, both 9.0/10 on Web-OS.
- Landing page for a phrase: Fusion, Fugu Ultra, GLM-5.2 or Grok, all 9.0/10 on Landing.
- Keyword judgement: a person, with Search Console data.
Want to learn the full stack? Join the AI Profit Boardroom and learn AI SEO yourself.
Or choose a time directly: Book a free SEO strategy session.
Book a free SEO strategy session →
Also On Our Network
- Agent Operatorsthe six-step operator workflow for clean KGR counts and a sorted keyword batch
- AI Income Deskwhether a KGR check is worth a small business owner's time, with a bakery example
- AI Tool Verdictthe Keyword Golden Ratio method rated out of 10, strengths, limits and verdict
- agentos.guidefour Agent OS prompts that seed, score and brief KGR keywords
Let the top-scoring models build the tool and the pages, keep the keyword calls with people, and book a free strategy session when you want a team acting on your kgr calculator.
FAQ
Does GoldieBench test keyword research?
No. It scores one-shot builds such as landing pages and interactive apps, which is why this page only uses it for the tool and page jobs.
Which model should build my own KGR calculator?
Fusion and Grok both scored 9.0/10 on the Web-OS task, the closest test of a working interactive tool.
Can a model tell me if a keyword will rank?
No. Rankings are never guaranteed, and the inputs to KGR are estimates, so keep keyword decisions with a person.
Where did these scores come from?
From the live GoldieBench models, Landing task and Web-OS task pages, read on 12 November 2026.


