Kagi’s AI-Off Switch: A Controlled Search Comparison Protocol
Kagi lets users disable AI features in Search. This protocol holds account settings, queries, lenses, domain rules, and environment constant without inventing a result.
From this desk
Start with the most useful page
The featured page offers a strong introduction. The remaining cards help readers move through this editorial section without turning the archive into a service menu.
Read the featured method
Compare page representations without changing the task
Use a preregistered protocol to compare accessibility trees, HTML, rendered pages, and screenshots while keeping tasks and scoring consistent.
Protocols
Plan the test before seeing the result
Declare the question, sample, environment, measures, exclusions, and stopping rule before interpretation begins.
Experiments
Change one material variable
Preserve the baseline, observation window, confounders, and decision rule so a result remains inspectable.
Evidence
Keep every claim close to support
Classify facts, observations, interpretations, recommendations, and opinions before drafting conclusions.
Latest research
Methods, tests, and evidence reviews
Each article states the method, retained evidence, unresolved limits, and the boundary on what the result can establish.
-
-
Which Schema.org Types Are Actually Used Across the Web? July 2026 Data
Official May–July data shows Product and Offer entering the 10M+ domain bucket in June and Article entering it in July, with important coverage limits.
-
How to Test Whether a Public ChatGPT Site Is Indexable
OpenAI allows eligible users to publish ChatGPT Sites publicly, but public access is not proof of indexability. Use this controlled technical audit before reporting results.
-
How to Read Brave’s 1,500-Query AI Search Benchmark
Audit Brave’s 1,500-query AI answer comparison by examining query sampling, interface parity, LLM judges, pairwise order, ownership, and replication.
-
Who Participates in IndexNow in 2026? A Live Registry Audit
A dated audit of IndexNow’s public participant registry, linked metadata, and the difference between being listed and receiving real-time URL notifications.
-
A URL Can Return HTTP 200 and Still Fail Before the Client Reads HTML
A reproducible client matrix shows Node rejecting oversized response headers before HTML while curl still succeeds, even though the server returns HTTP 200.
-
How to Read a Search Patent Without Calling It a Ranking Factor
Trace a search patent’s identity, family, legal status, claims, specification, and external implementation evidence before making a product or ranking claim.
-
Google, Bing and ChatGPT Report Different AI Visibility Signals
Use a five-layer crosswalk to separate crawler access, eligibility, AI appearances, referrals, and on-site outcomes across Google, Bing, and ChatGPT.
-
How to run small SEO experiments without overclaiming
Treat most small-site tests as structured observations: define one falsifiable question, save the baseline, change one material variable, log confounders, and prewrite the decision.
A useful standard
How to read SearchEngineAnswer research
A useful study makes its conditions visible. Treat each result as evidence within a declared sample and environment, not as a universal platform rule.
Method before conclusion
The question, inputs, environment, measures, and stopping rule should be understandable before the result is interpreted.
Failures stay in the record
Timeouts, absent results, rejected cases, and contradictory observations are data, not material to hide.
Limits remain visible
A measured association or one platform observation does not automatically establish causation or a permanent ranking rule.
References and next steps