Playbook
How Artificial Analysis gets its Google traffic
Artificial Analysis’s model section accounts for an estimated 52.9% of its main-brand-filtered top-20 keyword coverage.
- Estimated US organic visits/month
- 40,520
- Estimated ranking keywords
- 6,715
- Estimated non-brand top-20 keywords
- 4,032
- Estimated AI Overview citation searches
- 1,146
- Indexed referring domains
- 15,578
- Domain rank
- 79
The playbook
Artificial Analysis draws search visitors into comparisons, rankings and individual model pages; /models holds an estimated 52.9% of its main-brand-filtered top-20 keyword coverage.
Its useful asset is measured data: the same results support searchable comparisons and pages that publishers cite.
The first move to copy is one narrow comparison page with original measurements, clear methods and links to detailed results.
Scorecard
These are estimates and search-index observations, not first-party analytics.
| Metric | Finding |
|---|---|
| Estimated monthly US organic traffic | 40,520 visits |
| Estimated ranking keywords | 6,715 |
| Estimated brand share of top 100 keywords’ traffic | 30.7%: about 9,559 of 31,109 estimated monthly visits in that sample |
| Estimated non-brand keywords in the top 20 | 4,032 with the broader brand exclusions; 4,037 with the main-brand filter used for section comparisons |
| Estimated AI Overview citation coverage | 1,146 searches, including brand searches; not visits |
| Referring domains, indexed count | 15,578 currently |
| Referring-domain 12-month change | Exact year-over-year change: unknown. The 12 monthly observations rise from 3,979 in November 2025 to 14,612 in October 2026: +10,633, or 267.2% |
| Domain rank | 79, a third-party authority score |
| Previously measured sitemap pages | 15,178; not a count of indexed or traffic-earning pages |
Brand classification includes Artificial Analysis spelling variants, Artificial Index, Intelligence Index, GDPval-AA and AA LCR. Other companies’ model names count as non-brand for this site; ambiguous “ai analysis” stays non-brand.
Traffic engines
Shares below measure estimated keyword coverage, not visits. Section counts use the same main-brand exclusions: “artificial analysis,” “artificialanalysis” and “intelligence index”; five additional branded keywords remain in the site total.
| Section | Page type and how it is made | Estimated top-20 keywords after main-brand filter | Share of site total | Top page | Top keyword |
|---|---|---|---|---|---|
/models | Comparison hub, model profiles and comparisons; reusable templates filled from data | 2,134 | 52.9% | Model comparison | ai benchmarks |
/leaderboards | Ranked tables and metric summaries; templates filled from data | 308 | 7.6% | LLM leaderboard | ai rankings |
/evaluations | Benchmark results; data templates with explanatory background | 216 | 5.4% | Terminal-Bench result URL | unknown |
| Rest of site | Homepage, agents, media, providers and other pages; mixed formats | 1,379 | 34.2% | Homepage — highest sampled row | intelligence ai |
The rest is the arithmetic remainder, not a separately measured section. Shares are rounded; template classifications are inferred from the visible page structure.
| Page inspected | Search it answers | What the page does | Path into the product | Visible publication date |
|---|---|---|---|---|
| Model comparison | Compare AI models by capability, cost and speed | Combines charts, selectors, short metric summaries and FAQs | Model detail links and “Try it out” lead into deeper analysis and MicroEvals | unknown |
| LLM leaderboard | Which models rank highest | A filterable table puts comparable measurements side by side | Each row links to model and provider analysis; a trial link opens MicroEvals | unknown |
| Terminal-Bench 2.1 | Compare performance on a named benchmark | Explains the benchmark, methods and example tasks alongside cost and speed charts | Links to methods and other evaluations keep visitors exploring the analysis | unknown |
The ranked Terminal-Bench URL redirects to the inspected page. These pages make the analysis itself the entry point; a separate article is not required before someone can use it.
Top pages
Estimated traffic sums the unique organic keyword rows inspected for each URL, not all its traffic. “Top keyword” means the largest traffic row in this sample.
| URL | Section | Top sampled keyword | Estimated monthly searches, 12-month average | Observed absolute position | Estimated monthly organic visits in sampled rows |
|---|---|---|---|---|---|
| /leaderboards/models | Leaderboards | ai rankings | 4,400 | 1 | 6,171 |
| /models | Models | ai benchmarks | 8,100 | 1 | 6,019 |
| Homepage | Homepage | intelligence ai | 301,000 | 3 | 3,301 |
| /models/glm-5-2 | Models | glm 5.2 benchmarks | 720 | 1 | 317 |
| /evaluations/humanitys-last-exam | Evaluations | humanity’s last exam leaderboard | 320 | 1 | 167 |
| /evaluations/gpqa-diamond | Evaluations | gpqa diamond | 880 | 2 | 154 |
| /agents/coding-agents | Agents | agent benchmark | 390 | 1 | 125 |
| /evaluations/ifbench | Evaluations | ifbench | 390 | 1 | 125 |
Search patterns
Each volume is an estimate for one keyword, not an additive market total. Several phrases can represent the same demand.
| Pattern | Example keywords | Estimated monthly searches, respectively | Section answering it |
|---|---|---|---|
| Category + benchmarks | ai benchmarks; artificial intelligence benchmark | 8,100; 6,600 | /models |
| Category + rankings | ai rankings; llm rankings | 4,400; 4,400 | /leaderboards |
| Compare + category | compare ai models; ai model comparison | 590; 1,300 | /models |
| Best + category | best ai models | 4,400 | /leaderboards |
| Model + benchmarks | glm 5.2 benchmarks | 720 | /models |
| Named test + leaderboard | humanity’s last exam leaderboard | 320 | /evaluations |
| Named test | gpqa diamond; ifbench | 880; 390 | /evaluations |
| Job category + benchmark | agent benchmark | 390 | /agents |
Link magnets
Counts are observed domains in a sample of 50, sorted by source-domain authority, with one link per domain. Removing a search page, redirect/interstitial pages and a social channel leaves 45; these are not sitewide totals or endorsements of every source.
| Page type | What attracts the link | Sampled linking domains | Examples of sources and destinations |
|---|---|---|---|
| Homepage | A general reference for independent AI analysis | 17 | AWS, Meta, MIT Sloan → homepage |
| Model comparisons and profiles | A specific model’s measured performance or a comparison resource | 7 | Google Developers → Gemini profile; ZDNet Japan → Claude profile |
| Evaluation pages | A named benchmark or index that supports a claim | 6 | OpenAI → Terminal-Bench Science; Google Blog → GDPval-AA |
| Main leaderboards | A reusable ranking of models | 5 | The Guardian, Cloudflare Blog, AI Stack Exchange → model leaderboard |
| Articles | Analysis of releases and domain-specific evaluations | 4 | DigitalOcean → GLM analysis; Harvard Law → Harvey LAB-AA article |
| Image and speech result pages | Rankings for a specific medium | 3 | Microsoft News → streaming speech-to-text; SourceForge → text-to-speech |
| Documentation, methods and trends | Reference material behind the numbers | 3 | GitHub → documentation; xpr.media → benchmark methods; Daum → trends |
The clearest link pattern beyond the homepage is reusable evidence: model measurements, benchmark results and rankings. This sample shows where links land, not that those links caused the rankings.
Moves to copy
-
Build a comparison people can use immediately. The model section has an estimated 2,134 top-20 keywords after the main-brand filter; “ai benchmarks” ranks first. Start with a small set of alternatives and measurements that matter to one buyer. Answer “[category] comparison” on a single page. First page: “Customer-support AI models compared on accuracy, response time and cost,” using original tests and published methods.
-
Give each tested option a useful detail page. The GLM-5.2 page ranks first for “glm 5.2 benchmarks,” with an estimated 720 monthly searches. Use the same measurement fields for every option, add interpretation, and link back to the comparison. First page: “[Most-requested product] performance and cost benchmark.” Publish it only once there is enough evidence to help a decision.
-
Publish results for a named test. Evaluation pages cover an estimated 216 top-20 keywords after the main-brand filter and attract six domains in the strongest-link sample. A smaller startup can define one repeatable test around a customer task, explain its limits and publish comparable results. First page: “Invoice extraction accuracy benchmark,” with a fixed test set and error examples.
-
Turn measured results into a maintained leaderboard. The main leaderboard ranks first for “ai rankings” and receives five sampled linking domains. Target “[narrow category] rankings” with sortable results, dates and clear criteria. First page: “Support automation leaderboard by resolution accuracy and cost.” The useful pattern is a maintained decision aid; a claim to be “best” without evidence will not reproduce it.
What won’t copy
Brand demand already contributes an estimated 30.7% of the top-100 sample’s traffic. A new comparison page cannot inherit searches for Artificial Analysis or its named indexes.
The backlink index records 15,578 referring domains, including links from major technology companies, publishers and universities. That accumulated recognition cannot be reproduced by copying the layout. Nor does the previously measured sitemap count of 15,178 pages establish that publishing thousands of pages caused the traffic.
Independent measurements also require ongoing work. The visible pages combine methods, benchmark results and usable comparisons; the reusable template is only the presentation layer.
Method
Data from OpenSEO (DataForSEO estimates, US market), reviewed on 2026-11-18, with English search data and direct inspection of the linked pages. Traffic and search volumes are estimates; ranking and link observations can predate the review. Section shares use 4,037 main-brand-filtered keywords; broader exclusions reduce the site total to 4,032. Page traffic uses 30 sampled organic rows, with duplicates removed. AI Overview coverage includes brand searches. Unknown values were not inferred.