Andi is not a search engine you point at a topic. It is one you point at a decision. That difference sounds like marketing, but it changes how you should type. Put a bare keyword into Google and you get ten doors to open. Put the same keyword into Andi and you get a shrug, because the engine is built to compose an answer and it has nothing to work with.
Andi has been around since 2022, built by a small team that wanted search without ads, without tracking, and without the ten blue links. If you want the background on where it came from and how it differs from Perplexity or Bing Chat, the explainer on Andi’s answer-first model covers it properly. What follows is the part most people miss: the actual mechanics of getting good output from it. Six habits, one worked example, and a query template you can steal.
Write the query the way you’d ask a colleague
The single biggest upgrade to your results costs nothing. Stop typing fragments and type a sentence with your constraints baked in.
- Weak: “csv python”
- Better: “How do I append rows to an existing CSV in Python without rewriting the header row every time?”
The second query tells Andi three things: the task, the language, and the failure mode you’re trying to avoid. The first tells it almost nothing, so it guesses and you get a generic tutorial about the csv module.
Constraints are the whole trick
Anything Andi cannot infer from the words should be stated. Budget, country, skill level, deadline, what you have already tried. “Cheap laptop for university” gives you a shopping list. “Laptop under £600 for a computer science student who runs virtual machines and commutes by bike” gives you something you can act on. Same engine, wildly different output.
Treat every answer as the first turn, not the last
Andi keeps the thread. Most people fire one query, read the answer, and leave. The useful work happens in turns two through five.
Say you asked about switching your small business from spreadsheets to a proper accounting tool. The first answer will be broad. Then you narrow:
- “What changes if I’m VAT registered and invoice EU clients?”
- “Which of those options has the cheapest plan for two users?”
- “What does migrating six years of transaction history actually involve?”
By the third follow-up you’re past the marketing copy and into the questions that decide whether you buy anything at all. That is a fifteen-minute session, and it replaces an afternoon of tab-hopping.
Make it compare, not just explain
Andi handles side-by-side questions well, but only if you name both sides and name the context. Comparison queries without context produce comparison tables that could have been written by either vendor.
Compare this pair:
- “Notion vs Obsidian”
- “Compare Notion and Obsidian for a two-person consultancy that needs shared client notes and offline access on a train”
The second one usually returns a table plus a recommendation with reasoning attached. If you want the table, ask for it explicitly: “give me a table with rows for price, offline support, and sharing, and put a one-line verdict underneath.” Andi follows formatting instructions reliably, which is not true of every AI search tool.
Know when to reach for something else
Andi is a generalist. For narrow technical work it is the wrong instrument, and recognising that saves you twenty frustrated minutes.
Debugging a stack trace or working through framework documentation is what Phind was built for, and it will cite the specific version of the library you’re on. Long, citation-heavy research projects, the kind where you need to trace a claim back through four sources, are better served by a research-focused AI search tool with deeper source control. Use Andi for decisions, explanations, and anything where you want one clean answer instead of a reading list.
A worked example: three days in Lisbon for two people
Here is a realistic chain I ran recently. Budget was £600 for two, late November, and neither of us wanted to rent a car.
- “Three days in Lisbon in late November for two people on a £600 budget, no car, we like food markets and walking”
- “Which major museums in Lisbon close on Mondays?” (Most of them, which reshuffled the whole itinerary.)
- “What’s the weather realistically like in Lisbon in the last week of November?”
- “Cheapest way to get from Lisbon airport to Alfama for two people with cabin bags”
Four queries, maybe six minutes. The Monday closures detail alone saved a wasted morning, and it is exactly the kind of fact that a generic itinerary page buries in paragraph nine. Notice that none of these are keywords. They’re questions, and each one carries the constraint that matters.
Check the citations before you repeat anything
Andi cites sources inline. Read them. The failure mode with every AI search engine is the confident, fluent, slightly wrong answer, and the only defence is opening the link.
This matters more now than it did two years ago. Systems are producing plausible-looking output on genuinely hard problems, and Astra’s recent run at ten long-standing maths problems is a good reminder that fluency and correctness are separate qualities. If a claim would change what you spend money on, verify it at the primary source, not at the summary.
One more thing worth doing: citations tend to cluster around well-indexed English-language sources. If your question touches on research coming out of other regions, widen the net manually. Following researchers directly helps, and Chinese AI researchers have become notably more visible on X over the past year, which makes that easier than it used to be.
Two habits that compound over time
Ask it to argue against itself
After any recommendation, add: “What would have to be true for this advice to be wrong?” The answer is often more useful than the original suggestion, because it tells you which assumption is load-bearing. If the whole recommendation collapses when one number changes, you now know which number to check.
Keep a prompt file
Every time a query produces a genuinely good answer, paste the wording into a notes file. After a month you’ll have a dozen reusable patterns, and you’ll write better queries in every other AI tool too. The skill transfers.
The one-line template worth memorising
If you take nothing else from this, take the structure. Every strong Andi query has four parts:
[Who I am or what I’m working with] + [the decision or task] + [constraints and what I’ve tried] + [the format I want back].
So: “I run a two-person design studio. We need to pick a project management tool. We invoice monthly, work with three regular clients, and I’ve already ruled out Monday.com because the pricing jumps at five seats. Give me three options with a short reason each.” That is one sentence, and it will outperform anything you type as keywords. The engine is only as specific as you are.