TL;DR (Click to expand)
A little over a year with Kagi has made it obvious how much rubbish I had learned to ignore in ordinary search results. Kagi removes the ads, lets me block repeat offenders and turns up more personal websites. I recently cancelled a longer-running Perplexity subscription after it capped premium searches. Kagi Assistant now does most of the same work and publishes the numbers behind its allowance, including the 20% margin.
I recently cancelled Perplexity after it added a monthly cap to premium searches. My Perplexity subscription predates Kagi, and I was slow to reconsider it.
For a while Perplexity was the king of metasearch. A vague question would produce a decent briefing and enough citations to start checking the answer, and nothing else I tried was as good at searching across several sources and turning a loose query into a useful place to start. I still checked those citations, as anyone should with LLM output, but the core product was excellent, which makes each new cap and plan distinction particularly irritating.
When the cap appeared, I checked Perplexity's public plan comparison. Pro offers “extended access”, with weekly and monthly limits based on “average use”, but the page does not say what those limits are. The actual cap only appeared in my account.
I understand why AI services have limits because running the models and searching the web costs money. What annoyed me was having to translate the plan language before I could work out what I was buying, especially when the cap affected the main reason I had subscribed.
Unlimited, subject to vibes
This is the boring middle of enshittification, before a service becomes bad enough to inspire migration. The product still works and each individual change is defensible, but the plan becomes worse and harder to describe over time.
Perplexity has expanded into a browser, an agent, file creation and image generation. Those features are now spread across Pro and Max with different limits. I subscribed for search with better models and sources, and the plan had become harder to justify for that use case.
The monthly cap finally overcame the inertia that keeps most software subscriptions alive. I cancelled Perplexity and moved those questions to Kagi Assistant, which was already included alongside the search engine I had used for a little over a year.
That move was easy because Kagi had already become my default search engine. I now notice the difference whenever I use another browser, where the results page makes me sort through considerably more before reaching a useful link.
Orion, Kagi's WebKit browser, deserves an honourable mention. It has improved my Mac's battery life considerably, and I no longer have to put up with Safari's never-ending UX changes. It has zero telemetry, blocks ads and trackers by default, and supports many Firefox and Chrome extensions. It carries the same priorities into the browser itself, and I am glad somebody is trying something beyond another Chromium shell.
The personal internet hit list
Most of the improvement comes from the results themselves and a handful of controls that become more useful over time. Kagi has no ads and says it downranks pages heavy with ads and trackers while promoting independent sources. This removes plenty of pages engineered to rank first and provide an answer somewhere much further down.
I had developed a whole defensive routine around ordinary searches: ignore the sponsored block, skip the marketplaces pretending to be reviews, open several copies of the same press release, add “reddit” when the results became hopeless. Kagi made that routine less necessary, and although I still use bangs to search elsewhere when it misses, doing so is now occasional.
Kagi's domain controls have been more useful than I expected. Any site can be raised, lowered, pinned or blocked from later results, so a publisher that repeatedly wastes my time can simply be removed. This explicit form of personalisation works better for me than one inferred from tracking.
Over time this has changed the mix of results I see. Documentation and useful forums appear more often, familiar rubbish sinks, and Small Web surfaces personal sites that larger search engines tend to miss. Kagi Lenses can also narrow a query to a chosen group of sites, which is useful for technical searches and when an event fills the general results with rewrites of one wire story.
Finding personal websites again has been the most valuable part. Platform feeds separate writing from its original site and place it among unrelated posts, while visiting the site itself gives me an archive and links to follow. I have spent more time following those links during the past year, and the web feels larger even though Kagi often shows fewer results.
The improvement is difficult to represent in a benchmark because most of it consists of irritation that no longer happens. I spend less time deciding which result is useful, fewer searches end on affiliate pages, and the results page generally sends me on to the web without trying to hold my attention.
Twenty per cent, written in ink
Kagi Assistant covers most of what I used Perplexity for. I can choose a model, change it during a thread, enable web search when the question needs current information and apply the Lenses or result preferences already used by Kagi Search, and having those features in the same subscription is convenient.
Kagi's pricing is also easier to understand because Assistant usage is tied to the value of the subscription. Its worked example says a $25 monthly plan provides up to $25 of model usage: $20 of raw token cost and $5 for search, infrastructure and development. The $5 is a disclosed 20% margin, and the same documentation explains what happens as the allowance runs out.
The allowance can disappear quickly when using premium models, and long threads cost more as their context grows. Kagi also disables new interactions when the allowance runs out, although I can see enough of the cost to choose a cheaper model when the task does not need an expensive one.
That level of detail is rare among consumer AI subscriptions. Terms such as “extended”, “priority” and “reasonable” suggest generous access without explaining how much is included. Kagi publishes the allowance and its margin, which makes the subscription easier to evaluate.
Kagi could still change its pricing or make the product worse, though its current business model starts from a useful premise: the customer pays for the service, removing the need to monetise search history or attention. I know what I am paying for and can leave if that changes.
Paying for search felt absurd when I started because the previous price had been hidden for most of my life. A little over a year with Kagi has made the direct version easier to justify, and the monthly cap forced me to look closely at both subscriptions before keeping the one whose limits I could understand.