Kagi gives paying search users more control over their sources

Kagi is a paid search service with an unusually direct proposition: users fund the product and receive tools for shaping their own search experience. Its AI Assistant sits alongside that search foundation. The most useful evaluation is whether the combination helps a person find and inspect better evidence during everyday work. A long list of model names matters less than how easily the user can control sources, revisit a result and recognize the limits of an answer.

01 / ProductSearch is the foundation and Assistant is an additional mode

Kagi's current pricing page¹ describes an advertising-free, subscription-funded service with both search and Assistant access. Search returns the pages a person can open and evaluate. Assistant adds conversational synthesis with optional web access, file context and configurable model choices. These are related workflows, but paying for search does not require every question to become an AI-generated response.

The Assistant documentation⁴ describes switching web access on or off and applying Lenses or personalized results to Assistant searches. That lets a user decide whether a conversation should draw on the web at all. For a question about a supplied document, the appropriate starting point may be the document itself; for a current product feature, fresh external evidence is usually essential.

Kagi also provides Quick and Research assistants. The Research Assistants guide⁵ describes Quick as the faster option and Research as a multi-step process with additional planning and tools. Research is part of Ultimate. This is a product distinction to evaluate with the work you actually do, rather than assuming that a slower answer is inherently a better answer.

02 / AudienceA fit for people who search repeatedly within a subject

Kagi is most attractive to frequent search users: developers, analysts, writers and other knowledge workers who return to familiar subjects and sources. Domain preferences can gradually reduce repeated encounters with unhelpful sites. Lenses can preserve a useful source scope so the user does not rebuild the same constrained query every day.

The value is smaller for someone who searches occasionally and is content with their existing results. Paying for search should solve an observable problem, such as repeatedly filtering irrelevant sites or struggling to keep a research task within authoritative documentation. A subscription is easier to justify when those friction points recur throughout the workday, not because the product presents an appealing philosophy alone.

Compare Perplexity's blueprint when conversational research is the main interface you want. Exa's blueprint is relevant to developers building retrieval into an application rather than improving their personal search environment. These are different buying decisions. Kagi's strongest consumer argument is control over the everyday relationship between search results and the person reading them.

03 / WorkflowUse source scopes without narrowing away the answer

A proposed trial is a week of research into a software migration. Begin with a small set of real questions: which behavior changed, which versions are affected and what migration path the maintainer recommends. Run ordinary searches first and record the useful sources. Then create a Lens for the official documentation and release-note domains that repeatedly matter.

Kagi's Lenses guide² describes include and exclude lists for sites and keywords, along with region, file-type and date controls. A custom Lens can include up to ten websites and exclude up to ten. It is a constrained view of search, not a separate verified database. Its results remain dependent on what those sites publish and what the search system can retrieve.

Keep a broad discovery pass

Use the official-source Lens to establish supported behavior, then temporarily broaden the search to discover terminology, common failure reports and questions you may have missed. A community report can identify a symptom worth investigating even when it is not sufficient evidence for a final technical claim. Separating discovery from verification is more useful than permanently excluding every unofficial source.

For each migration question, keep a short note with the source page, applicable version and unresolved condition. Do not rely on the eventual AI summary to preserve every qualification. If a release note says a feature is available only behind an option, the note should retain that condition explicitly.

Personalize recurring noise, then compare without it

Kagi's personalized-results guide³ lets users block, lower, raise or pin domains. Personalization can also be disabled for one search. Use that temporary comparison when a query unexpectedly produces little evidence. A preference that helped one task may suppress a useful source for another, especially when a publisher covers several unrelated subjects.

A Lens and a domain preference solve different problems. The Lens defines the scope for a particular task; a domain preference expresses an ongoing ranking choice. Keep the task-specific restriction in the Lens when you do not want it affecting unrelated searches. This makes the search environment easier to reason about months later, when the original reason for a global preference may no longer be obvious.

Move to Assistant when synthesis adds value

After locating the relevant documents, ask Assistant to compare the old and new behavior or explain a difficult passage. Keep source links in the response and verify the claim that determines the migration decision. A readable explanation can reduce effort, but the maintainer's condition still governs the implementation. If the question is already answered by one clear paragraph, an additional research cycle may add little.

04 / PricingCurrent plans separate search allowance from AI depth

The pricing page¹, checked 15 September 2026, lists Starter at $5 per month, Professional at $10 and Ultimate at $25, plus applicable sales tax. Starter includes 300 searches; Professional and Ultimate include unlimited search. Ultimate adds Research mode and premium model access. Annual billing is advertised at a 10% discount.

Assistant remains subject to a usage allowance even on a plan with unlimited search. Its usage documentation⁴ describes a value-based token budget and an in-product notice as the limit approaches. A month of short questions and a month of large-file analysis can therefore produce very different AI usage on the same subscription.

For a practical budget decision, separate how often you search from how often you need substantial synthesis. Professional can make sense for someone whose main benefit is daily source discovery. Ultimate is easier to justify when deeper research or premium models replace work the person regularly does elsewhere. The decision should follow observed use during the trial rather than an assumption that the highest tier is necessary for useful search.

The free trial includes 100 searches, which can be enough to compare a representative set of tasks if used deliberately. Choose questions from the actual workweek, including some whose answers you already know well enough to judge. A trial filled only with broad curiosity questions can make the product pleasant to explore while leaving its practical value unresolved.

05 / DistinctionsSource control is the meaningful differentiator

Kagi's domain controls make ranking preferences explicit. That is useful because a search user can identify a recurring problem and change the treatment of the source directly. A site that consistently provides thin summaries can be lowered, while a source that repeatedly supplies the needed original material can be raised. The benefit is cumulative when the user returns to the same subject area.

The tradeoff is that a customized search experience is no longer a neutral baseline for comparing results with a colleague. When sharing research, share the query and Lens scope as well as the answer. Another person may have different domain preferences and therefore see a different set of pages. A shared Lens creates a copy of its settings, not a permanently synchronized research policy for everyone who uses it.

The Research assistant's toolkit adds a different kind of value. Its guide describes searching, reading source material, processing files and using computational tools. That can help with a task combining discovery and calculation, but the evaluation should identify which part benefited. A correct chart built from an unsuitable source still leaves the research conclusion weak.

06 / QuestionsRetention and coverage need specific checks

Kagi states that Assistant content is not used to train models and that account information is not shared with model providers. Its default thread behavior removes conversations after 24 hours of inactivity unless the user changes retention. The separate LLM privacy guide⁶ describes provider-dependent processing and retention. A temporary thread setting should not be confused with a universal zero-retention promise across every provider involved.

For the migration trial, save or export the final evidence note in the team's normal documentation system. A disappearing chat is convenient for temporary questions but unsuitable as the only record of why a production decision was made. Preserve the result that matters without assuming the search provider's conversation history is a permanent project archive.

Also test subjects where your existing search engine works well. A service can feel dramatically better on one technical topic and offer little improvement for a local query or a specialized database. Keep those cases separate in your assessment. A universal verdict obscures the practical choice: which service should be the default for the work you repeat most often?

07 / DecisionChoose Kagi for a search environment you will maintain

Kagi is a strong candidate when frequent search is part of your work and you want to shape recurring sources explicitly. Start with the ordinary search experience, add one useful Lens and make only a few deliberate domain adjustments. Then assess whether Assistant or Research improves the parts of the workflow that need synthesis.

The best outcome is not a perfectly personalized list of familiar sites. It is a search routine that finds dependable sources faster while still allowing a broad check when the evidence is incomplete. Choose the plan around that routine and the AI allowance you actually need.