What search was always supposed to be
Google used to work.
Now it doesn't. I'm Nic Weyand. A year ago I started building my own search engine, because I was mad that Google used to work great and now it's horrible. As I kept building, Google kept getting worse. So I kept making more, deliberately the opposite of Google and Silicon Valley's rot.
The Argand Principle
"The user finds EXACTLY what they were looking for, the first time they look."
That sentence is the test every feature has to pass. It is not a slogan. It is the question I ask about every decision: does this actually help someone find what they came for, the first time they look? If the answer is no, or even "probably", the thing doesn't ship.
Anti-enshittification, by architecture
Four things Google does that Argand cannot.
Enshittification, a term coined by the writer Cory Doctorow, is the documented pattern of a platform getting worse for users over time to wring more money out of them. Argand blocks it by making it architecturally impossible. Not by promising not to.
No paid rank. There is no place in Argand's code to sell a higher spot. The document data has no field for a sponsor or a bid, so a paid result has nowhere to live. It is the shape of the system, not a promise it makes.
Tap, swipe, or use the arrow keys.
Prefer to read? The four guarantees in words
The internal record for a document has no field where a sponsor or a bid could be stored, and the ranking function takes no parameter for business logic. So there is nowhere to inject a paid result, because the data type has no slot for one. A later change that tried to add such a slot would be caught by the spec that forbids it.
Every result links you to the source page through a /yw/ referral, so the click lands on the site that did the work. The yw is short for "you're welcome", because the whole point is
handing that site its traffic instead of keeping it. Argand shows no
AI overviews and no featured snippets that answer the question in place of
sending you on. Structured data like a recipe card or an FAQ is there to
help you choose which result to open, never to replace opening it. The one
thing shown inline is pure computation, like a unit conversion or a sum.
Every time the index is rebuilt it is scored against a set of at least a thousand questions that people have judged by hand. Two standard search-quality measures, and , are computed for the new build and compared with the one already in production. If either measure drops by more than half a percent, the build is rejected automatically. The only way past the gate is to update the judged set with fresh human ratings. No configuration flag can skip the check, so quality can only hold steady or rise.
Argand keeps no persistent identifier for you, and it forgets your search history the moment your session ends. It never tries to guess who you are from how you behave. Ask the same question in the same language as anyone else and you get the same answer back. The only thing that shapes your results is the filters you set on purpose. You can see and change those at any time, and they clear when the session ends.
How it works
The search flow
A search feels instant, but a fair amount happens in well under a second before you see a result. Here it is, step by step.
You type anything, typos and clumsy phrasing included. Argand cleans the text up and quietly adds related words to widen the net, before it looks anything up.
Tap, swipe, or use the arrow keys.
Prefer to read? The whole search flow in words
You can search for anything. Plain language, technical questions, odd phrasing, and typos are all welcome. The first thing Argand does is tidy the text up. It fixes obvious mistakes and quietly adds related words to widen the net, so a query phrased one way can still find a page that phrased the same idea another way. That cleanup happens before the search itself runs.
Then three searches run at the same time. The first looks for your exact words. The second looks for closely related words that Argand has learned tend to go together. The third looks past the words altogether and matches on the overall meaning of what you asked. Each of the three returns its own short list of likely pages.
Those three short lists are merged into a single set of candidates, so the best of each approach ends up in one place. A second, slower model then re-reads the top of that combined list against your actual question and floats the genuinely best results up to the top. This re-read step is the same one the Benchmarks section measures, and it is where most of the quality comes from.
When Argand is confident it has found the right page, Argand Navigate sends you to the exact spot on that page and highlights the passage that answers you. When it is not confident, it stays silent rather than guess, and it never makes up an answer. Either way the click goes to the source site, not to Argand.
The ranking also learns from how searches actually go. If people consistently pick the second result over the first, or skip past the first page, that pattern feeds back into the ranking and helps Argand do better next time. The signal is anonymous and stays inside the engine. It is used to improve search, never tied to you, and never turned into a profile.
A promise
Argand will never generate an answer with AI.
Argand will never write an answer for you with AI. No AI overview sitting on top of the results, no confident summary that might be wrong. The job of a search engine is to find what already exists, not to manufacture it. Everything below is why the engine is deliberately built to run small, on hardware you might already own.
Most AI search runs in warehouses packed with power-hungry graphics cards. Argand made the opposite choice and runs on ordinary processors, the kind already inside a cheap laptop or an old phone.
Tap, swipe, or use the arrow keys.
Prefer to read? The whole story, with sources.
The promise, in plain terms. Argand finds pages that already exist. It will never generate an answer for you, the way a chatbot does. There is a reason that matters beyond taste: if a confident summary sits on top of the results, the ranking underneath can quietly rot and nobody notices, because they read the summary and leave. So Argand shows you real sources and sends you to them.
Built for a potato. Most search engines assume you have fast wifi and a recent phone. Argand assumes you might have neither. It needs no graphics card, only an ordinary processor and a small, hand-tuned index. It is tested at 2G mobile speeds, roughly 50 to 100 kilobits per second, the kind of connection you still hit on rural networks and anywhere bandwidth is scarce or expensive. A typical home broadband or 5G connection is hundreds of times faster. The point is simple: if Argand is quick on a potato, it is quick on whatever you actually own.
What the data-centre boom actually costs. The alternative, the one most AI products chose, is enormous. According to the International Energy Agency's Electricity 2024 report, data centres used about 460 terawatt-hours of electricity in 2022 and could pass 1,000 terawatt-hours by 2026, which the agency notes is roughly the entire annual electricity use of Japan.
Those buildings also consume water to stay cool. A study by Li and colleagues, peer-reviewed in Communications of the ACM in 2025, estimated that training a single older model, GPT-3, in US data centres could use about 5.4 million litres of water, and that a short exchange of a handful of questions is on the order of a 500 millilitre bottle. The companies' own numbers point the same way: Google's 2023 environmental report put its company-wide water use at about 5.6 billion gallons in 2022, up around 20 percent in a year, most of it for cooling its data centres, and Microsoft reported a 34 percent jump over the same period.
It lands on real neighbourhoods, and on real authors. These warehouses are not abstract. In Chandler, Arizona, residents living near a large data centre complained for years about a constant hum, and the city ended up requiring noise-mitigation plans before new ones could be built. In Virginia, home to the world's largest concentration of data centres, state legislative researchers and environmental regulators are now scrutinising the thousands of diesel generators that sites keep on standby for backup power, with stricter emissions controls expected on new permits from mid-2026. These are documented complaints and regulatory questions, not a settled verdict on health.
There is a consent problem too. The AI models that power generated answers are trained on text and images scraped from the open web without the authors' permission. The New York Times sued OpenAI and Microsoft in December 2023 over its articles, and a group of visual artists filed a class action in January 2023 over their work. Both cases are ongoing and the claims are contested, but the underlying practice is why Argand does not build on generated answers at all.
Argand's answer. Argand does the expensive work once, when it crawls and indexes the web. After that, answering your search runs on a plain processor and sips power, on a host that states it runs on renewable energy. There is no graphics-card bill, so there is no money pressure to cut corners later. There is no new data centre being built in someone's backyard on Argand's behalf. And there is no generated answer, because the whole design is to point you at the real thing instead.
The Argand ecosystem
Search is just the beginning.
One platform. 59 microservices and 47 in-house Rust crates in development. A few of the near-term surfaces are below; the full twelve live on the roadmap.
Tap any service to see why it exists and what it does.
Dig deeper
The whole story, in pieces.
This page is the short version. Each piece below is its own page — link to it, share it, come back to it.