Two surfaces, not one ‘Kagi AI’

Kagi Search is a search surface with result listings, Quick Answer, vertical answers, and user controls. Kagi Assistant is a separate conversational surface with a selected model; web access can be enabled for one message and then disabled. For auditing, a Search result cannot automatically be attributed to Assistant, and a model name does not prove that web was used.

Section sources:[2] Kagi[1] Kagi

What Kagi calls a source

Kagi lists Teclis and TinyGem, anonymized calls to major search providers, Marginalia, and vertical sources such as Wikipedia, Wolfram Alpha, Apple, Yelp, and weather data. This describes source families, not the full candidate set for one query. A report should capture the visible result and avoid turning the source catalog into a claim about every answer.

Section sources:[1] Kagi

Lenses as an explicit space constraint

A Search Lens can set region, included and excluded sites, keywords, file type, and date boundaries. Academic, Forums, PDFs, and Kagi Documentation therefore change the search space. A lens is an observable user control, not a hidden SEO signal: compare the same prompt with and without it while recording the exact configuration.

Section sources:[3] Kagi

Quick Answer, links, and citation evidence

Quick Answer is described as AI extraction and a brief summary of important result content with links to source material. Separate four events: a URL appears in Search; it appears in Quick Answer; the link opens the source; and the source supports the specific proposition. The last requires reading context and cannot be inferred from visual co-occurrence alone.

Section sources:[1] Kagi

Assistant, models, and context

Assistant offers model selection, custom assistants, file uploads, a web toggle, lenses, and personalized results. These settings shape a particular thread. An uploaded file, pasted URL, and web result are different input classes and should be logged separately. The existence of a web toggle does not mean every subsequent answer calls Search.

Section sources:[2] Kagi

Search API versus the consumer interface

The Kagi Search API provides programmatic access to premium search results and inherits account settings such as promote/block and snippet length. An API response is not equivalent to browser Assistant: endpoints, formats, limits, and context differ. A study should name the surface rather than saying generically that ‘Kagi chose a source’.

Section sources:[4] Kagi

Plans and access

The documentation says Assistant is available to members, while premium models depend on Ultimate; Research and Quick differ in depth and tools. Do not transfer that statement to anonymous browsing or the API without checking plan documentation. Record plan, sign-in state, market, language, and date in the protocol.

Section sources:[2] Kagi[5] Kagi

Reproducible test and unknowns

For each URL, make four records: Search without a lens, Search with a saved lens, Quick Answer, and Assistant with web on. Repeat in RU and EN on the same day; preserve prompt, answer, link order, excerpts, account, and configuration. Unknowns include the full retrieval graph, domain weights, Teclis/TinyGem refresh rules, reasons for a particular citation, and the share of answers without web. Label the next statement observation, inference, or hypothesis.

Section sources:[1] Kagi[2] Kagi[3] Kagi[4] Kagi[5] Kagi

What this does not prove

  • The documentation describes sources and features, but not the complete index, candidate weights, personalization, Assistant retrieval internals, or causality between crawling and citation.

Sources

  1. 1
    Search SourcesKagi · official · 11 Sept 2026
  2. 2
    Kagi AssistantKagi · official · 11 Sept 2026
  3. 3
    LensesKagi · official · 11 Sept 2026
  4. 4
    Search APIKagi · official · 11 Sept 2026
  5. 5
    Kagi Research AssistantsKagi · official · 11 Sept 2026

Correction history

No material corrections have been published.