Google Interview Process & Rounds
In candidate reports from April to October 2026, a Google software engineering application starts with the Google Hiring Assessment, a work-style questionnaire with no coding, and some new grad and early career candidates also take a two-problem coding assessment on HackerRank. The interviews then come in two rounds. Round 1 is usually two back-to-back 45-minute video interviews, one coding and one Googleyness and leadership. Round 2, the onsite, is usually two coding interviews for L3 and L4, with system design added at L5 and above and a Solution Design round in some L4 loops from September 2026; in most 2026 reports the onsite was held in person at a Google office. Each coding interview is usually one main problem, often a reworked LeetCode problem, followed by harder follow-ups, written in a shared document or plain editor where the code is not run. After the onsite a hiring committee reviews the loop, and in most reports that describe the order, team matching comes after approval, so the whole process can take months. Google DeepMind runs its own loops, built around your research, ML foundations and coding.
Part of the full catalog of Google interview questions, covering every round.
AceOffer is not affiliated with, endorsed by, or sponsored by Google. Questions are reconstructed from interview experiences candidates shared, not from Google materials.
Key facts
- •7 distinct round types
- •48 questions reconstructed from 111 candidate reports
- •Reports span Apr 2026 – Oct 2026 (DeepMind from Aug 2025)
- •Refreshed monthly · last updated November 2026
The Google loop, from candidate reports
Reported Google software engineering loops start with the Google Hiring Assessment, a work-style questionnaire with no coding, and some new grad and early career candidates also take a two-problem coding assessment on HackerRank. Round 1 is usually two back-to-back video interviews, one coding and one Googleyness and leadership. The onsite is usually two coding interviews for L3 and L4, held in person in most 2026 reports; L5 and above add system design, and from September 2026 some L4 loops add a Solution Design round. Coding is LeetCode-style with Google settings and follow-ups: the earliest time everyone in a friend log is connected and what changes when people unfriend, Number of Islands and its variants, Meeting Rooms variants, course schedule variants, and the fewest cars for a set of rental reservations. Machine learning loops add an ML domain interview, on topics from LLMs and Transformers to your own projects, and an ML design such as a recommendation system. Google DeepMind runs its own loops, with research conversations or a research presentation, ML foundations questions and, in most reported loops, coding rounds.
Google interview timeline
| Step | Typical timing |
|---|---|
| Application to first contact | A day to ten months (10 reports); several candidates were contacted four to ten months after applying |
| Hiring Assessment to result | Twenty minutes to two days (10 reports); one reply says it took up to a week |
| Round 1 to result | Same day to two weeks (15 reports); one candidate waited about three and a half weeks |
| Passing Round 1 to the onsite | About two and a half to five weeks (6 reports) |
| Onsite to result | Same day to about a month (13 reports) |
| End to end | Three weeks for one team-hire loop with no team matching, to about four and a half months from referral to signed offer |
The fastest report is a team-hire L4 loop: recruiter contact in mid-March, a first round with the hiring manager two weeks later, a two-interview onsite a week after that and an offer call the next day, with no team matching. The slowest steps are usually scheduling and team matching. One L5 candidate was referred in early June, interviewed in late June and late July, was approved in late August, matched with a team in mid-September and signed in mid-October. Several candidates describe a contract recruiter handing them to a Google recruiter for the onsite, and waits of a week or two to book a call. Interviewers or recruiters did not show up or rescheduled in at least six reports, and in two reports a visa check call came before the final feedback.
The Google interview process, step by step
1. Application, recruiter contact and the Google Hiring Assessment
An online multiple-choice questionnaire on work style, with no coding; the invitation says about 30 minutes, it is pass or fail, and two candidates had four or five days to complete it.
Reports describe three parts. The first gives three statements at a time and asks which is most like you and which is least like you, and since April 2026 it has replaced an opening section that asked you to agree or disagree with one statement at a time. The second asks how strongly you agree with single statements. The third gives workplace scenarios, such as a gift from a client, a request for confidential client details or a conflict of interest, and asks which action you are most and least likely to take. You can leave and come back, but one candidate notes you cannot go back to change an answer. Results come by email within twenty minutes to two days in most reports. Passing does not guarantee interviews: several candidates heard nothing for weeks, one was rejected on their resume afterwards, and one recruiter said at least one hiring manager has to pick up your resume before interviews are scheduled. In eight reports first contact came from a contract recruiter, and several were later handed to a Google recruiter for the onsite. One recruiter ended a process before any interview because the candidate's current job title was not software engineer.
- →Decide on a consistent work style before you start. Five candidates say the same traits come back in different wording, though one passed without being strictly consistent.
- →In the scenario section, choose the option that follows the stated policy. Four candidates describe these as the easiest questions, with one clearly correct action.
- →Leave about an hour. Three candidates took close to an hour, against the 30 minutes stated.
- →Keep other applications moving after you pass. Several candidates waited weeks with no contact.
2. Online assessment (some early career roles)
HackerRank, two problems in 90 minutes in 2026 software engineer II early career reports, with a built-in AI assistant allowed on some problems.
Coding assessments appear in new grad, early career (L3) and intern reports. Other early career candidates went straight from the Hiring Assessment to Round 1 with no coding assessment, and an L4 AI/ML candidate had none. In spring 2026 early career reports, one problem hands you a feature-rollout codebase to debug until its tests pass, with an AI assistant you may use (3 reports), and the other computes an email verification checksum (2 reports). One candidate found the assistant weak and unwilling to answer questions about the logic; another found it good at locating the source of a failing test but unwilling to answer some questions about Python string methods, and recommends trying the sample problems with the assistant first. The invitation email for that candidate said the two problems take about 70 minutes, but they needed the full 90. Results came the same evening, two days later and two weeks later in reports that give them.
- →Try the sample problems with the AI assistant before you start the scored assessment, so you know what it answers well.
- →Finish the checksum problem quickly. One candidate suggests five to ten minutes on it, leaving the rest for the debugging problem.
- →Use the failing tests. One candidate pasted each test error into the assistant to find where the bug was, and fixed most of the codebase that way.
3. Round 1: coding interview (virtual)
45 minutes on video, in a shared Google Doc in the reports that name the tool, usually back to back with the Googleyness interview.
In 47 reports, Round 1 is one coding interview and one Googleyness and leadership interview, and a reply in one thread says this is the standard first round below L5. The coding interview is usually one problem with follow-ups, though some screens had two problems or a warm-up before the main question. Candidates describe a document with no highlighting, no autocomplete and no way to run code; one had to write their own tree node class. Interviewers type notes throughout, and some start the problem with no introductions. Problems are often stated aloud or loosely, and one L4 candidate felt the ambiguity was deliberate. Several candidates say you are expected to clarify constraints, choose an approach with the interviewer, dry run your own test cases and explain complexity. In two reports of the same 2D clustering problem, the interviewer wanted the brute force written first. Some first rounds differ: one L5 AI/ML candidate got two design interviews instead, a Cloud screen paired behavioral questions with a security monitoring design, and machine learning roles may pair an ML domain interview with Googleyness. A reply in one thread says the recruiter waits for both interviews' feedback before booking Round 2.
- →Practise writing code in a plain document, without running it, and dry run it out loud line by line.
- →Agree on the approach and the edge cases before you write code. One rejected candidate's feedback named missed corner cases and communication, and their own lesson was to align with the interviewer first.
- →Ask whether the interviewer wants a simple version first. In the clustering reports, the interviewer preferred brute force over the optimised neighbour search the candidate proposed.
- →Expect a follow-up on very large or streaming input. One L4 candidate was asked what happens when the input is very large in two different rounds.
4. Round 1: Googleyness and leadership
45 minutes on video with an engineer or manager: past-experience and hypothetical questions, each followed by detailed follow-ups.
The most reported topics are disagreement and conflict (19 reports), your proudest or most challenging project (11), why Google and culture questions (11), a mistake or critical feedback (7), too much work and prioritisation (6), mentoring or helping teammates who struggle (6) and ambiguity or changing plans (5). Interviewers dig into each answer: candidates describe being asked to tell more and what they would do differently, a STAR answer broken into four separate questions, and one 45-minute interview spent on a single project. Several candidates also got hypothetical or role-related questions, such as how to test an image detection app, how to verify a smile detection feature in Google Photos, or how to improve a team with a poor culture. This interview can decide the outcome: one candidate was told their coding feedback was strong but the behavioral interviewer gave no hire.
- →Prepare six or seven stories with results in numbers and the reason for each decision. One candidate who did this kept rushing the result when rehearsing, and recommends recording yourself.
- →For each story, prepare what you would do differently. Follow-ups in several reports ask for it.
- →Prepare for one or two hypothetical questions, including how you would test or verify a product feature.
5. Round 2 onsite: coding
For L3 and L4, usually two 45-minute coding interviews back to back, held in person at a Google office in most 2026 reports, typing on a Chromebook.
In 23 reports the onsite was in person, at offices including Mountain View, Sunnyvale, New York and Seattle, and some candidates could choose the office; Google booked or reimbursed travel in three reports. Some candidates moved it to video by citing another company's process or having no vacation left, while others asked and were refused, including one who was unwell. Code is typed in a text editor on a Chromebook, with a whiteboard and paper for examples in one report; one candidate notes there is no way to run the code, so you dry run it line by line. One PhD candidate's onsite also included lunch with an engineer, which they understood not to be scored. Problems are usually LeetCode-style with a Google setting and follow-ups that change the problem: the most reported are the earliest time everyone in a friend log is connected, then unfriend events (8 reports), Number of Islands and its variants (6), Meeting Rooms variants (6), course schedule variants (5), the fewest cars for a set of rental reservations (4) and the longest non-decreasing subarray if one element may change (4). Others are custom, such as checking a mahjong hand or counting hand moves on a piano. In some loops a follow-up only had to be discussed; in others it had to be written.
- →Ask your recruiter early whether the onsite is in person, and give your reason if you need it to be virtual.
- →Practise the follow-up ladder, not just the base problem: deletions or unfriend events, adding land, K lists instead of two, a stream, and input too large for memory.
- →Ask whether a follow-up needs code. One candidate passed after only discussing theirs; another, rejected, did not finish writing the union-find version of the islands follow-up.
6. Round 2 onsite: system design and Solution Design
One design round at L5 and above; from September 2026, some L4 loops add a 60-minute Solution Design round.
L5 and above loops in reports are two coding interviews and one system design. Reported prompts include a rate limiter for AI output or billing tokens (2 reports), a news aggregator, Google Calendar event conflicts, a customer support system where a virtual agent hands off to a person, a web crawler and, for L6, Find My iPhone with privacy and security. Machine learning loops ask ML design, most often a recommendation system. An October 2026 L4 candidate quoted the Solution Design description: break an ambitious problem into parts, define APIs, structure code and data for specific sub-systems, justify infrastructure and data choices, and show prioritisation, testing and documentation. Reported Solution Design prompts are a client app that stores edits offline and syncs them with version conflicts, and a review ingestion platform that summarises reviews with an LLM, with questions on prompt injection, personal data and trust and safety. One candidate found it like system design but less ambiguous, drawn on a whiteboard in person. In video rounds design is often done in a Google Doc, which one L6 candidate found hard to use without diagrams.
- →Read the round descriptions your recruiter sends. One L4 candidate noticed only on a closer read that Solution Design was not the system design they had prepared for.
- →Drive the depth yourself. One candidate felt their web crawler interviewer valued taking the discussion deeper without waiting to be asked.
- →For Solution Design, prepare APIs and data models for a small system, plus how you would test it and handle conflicts or abuse.
How the Google loop differs by role
Google DeepMind: research scientist and research engineer
DeepMind runs its own process, and a reply in one thread says it does not share a pipeline with Google, so a DeepMind rejection does not stop a Google application. Loops in reports start with a recruiter or HR call on your background, why DeepMind and practical questions such as visa status. Research scientist loops center on your research: one candidate presented their own paper and was questioned on why they did it, alternatives to each design choice, what happens if an assumption fails and what comes next, with no coding at all; another had three research chats and a hiring-manager interview, all very specific about their past work and the team's. Several research conversations turn into ML foundations: one intern interview was planned as 25 minutes of fundamentals (normalisation, Transformers from end to end, losses and embeddings), five minutes of research and 20 minutes of AI system design. Other research loops do include coding. A Gemini research scientist's first round was a 30-minute ML debugging exercise on a model with problems to find, which a reply describes as loss or attention code with planted bugs, and another research scientist had an LLM post-training coding round. A research engineer had two ML rounds of an hour each, mixing their research with harder questions and some system design, then two coding rounds; one coding problem was described as a Hard graph search variant, and they were rejected. A London research engineer's recruiter described two LeetCode-style coding rounds of medium to hard difficulty and an ML quiz round. Hiring-manager interviews ask about your proudest project, the limits of current models and what you would work on if you joined. One AI for science candidate had five rounds from June to August and was turned down at the end of September.
Google DeepMind: coding, ML design and hiring-manager rounds
A new grad DeepMind loop had a one-hour ML foundations interview with no code, from overfitting to LLMs, and a 50-minute coding interview with two short problems, complexity questions and test cases you write yourself. Other coding rounds in reports are not LeetCode problems: writing a Python generator over a list with unit tests and then sampling by given probabilities, or a game strategy problem solved with graph search. An ML design round asked for a model to predict a reaction factor for pairs of molecules, covering data exploration, features, splits, model choice and generalisation, with ML quiz questions along the way. One hiring-manager interview mixed an intro and project questions with quick technical questions such as regularisation methods and multiplying matrices split across two GPUs. A DeepMind Android loop had an easy-to-medium graph problem from the hiring manager and an Android design of the Gemini app, run as rapid questions where the interviewer sets the pace. One candidate went from an HR call in mid-September to an offer in mid-October, through system design (recommendation), coding, an ML fundamentals project deep dive and a hiring-manager conversation with a business case, but was offered L4 instead of L5. One candidate who finished every round says there was no separate behavioral interview, and one software engineer heard back within a week of the phone screen.
Software engineer AI/ML and machine learning engineer
Google's AI/ML software engineer and machine learning engineer loops add an ML domain interview. Recruiters offered some candidates a topic, such as ML fundamentals, LLMs, deep learning, recommender systems or NLP, though one candidate's questions were all about their resume. Reported domain interviews cover all of the LLM stack from data to post-training and Transformer variants, a project deep dive, or an ML design such as a mobile app that identifies leaves from photos. ML system design most often asks for a recommendation system, including YouTube Watch Next, where one interviewer asked for loss functions and optimiser formulas. One machine learning engineer had ML coding (a Transformer block with SwiGLU), a recommendation design, a web crawler coding round and behavioral questions, and later added team-lead and hiring-manager rounds. One L4 candidate was rejected on the ML design round alone, according to their recruiter.
New grad and early career (L3)
Reports from 2026 describe the Hiring Assessment, for some a two-problem HackerRank assessment, a virtual Round 1 of coding and Googleyness, and an onsite of two coding interviews. One new grad who passed in April 2026 was assigned a team without any match calls, while others were still waiting to match weeks later. One general hire new grad had four video rounds, each with behavioral questions and a coding problem. One early career candidate's Round 1 coding and behavioral interviews were a month apart, with the onsite two months after that.
L5 and above
L5 loops in reports pair a Round 1 coding interview with a leadership interview, then an onsite of two coding interviews and a system design. Some first rounds are domain interviews instead: one L5 candidate was told theirs would cover distributed system debugging, and an L6 candidate's screen was a Googleyness interview and a system design. Coding at this level includes less common problems, such as implementing a tee iterator, a custom future with multithreading, or packing servers into racks. Down-levelling is common across levels in reports: at least eight candidates were offered or moved one level lower, at points from after Round 1 to the hiring committee. One L5 candidate's request for a virtual onsite was refused.
Forward Deployed Engineer
Four candidates describe a role they understood to be new this year, interviewed in two rounds of about an hour where the length is given: a coding round (Number of Islands with a dynamic follow-up, a trie autocomplete, a parking problem then overlapping rectangles, or string cleanup) and an agent or AI solution design round, such as an enterprise AI adoption plan, a manufacturing agent that must avoid hallucination and enforce permissions, or a proposal to a small company's CEO. In two reports the candidate chose between an enterprise agent scenario and image or video generation. One recruiter said the FDE process is separate from other Google interviews and that FDE shares the software engineer pay band.
Interns and research interns
Software engineering intern reports describe an assessment and two 45-minute interviews, mostly coding with a few minutes of behavioral or resume questions; one interviewer asked for three short problems in one interview. Research intern loops can be entirely research: a Cloud research intern had two video interviews on their research and interests with no algorithm problem, and a DeepMind student researcher says these internships rarely include coding.
After the final round
After the onsite, the interviewers' feedback goes to a hiring committee. In six reports that describe the order, the committee's approval came first and team matching followed; one L5 candidate's approval email said the panel had completed its review of the whole loop, and an L5 machine learning candidate was told they could not match at L5 before the committee's result. Some loops add an interview before a decision: an extra coding round in three reports, an extra AI/ML round in one, and team-lead and hiring-manager rounds in one. Team matching means calls with several teams, which can include technical questions; one candidate spoke to six teams, mostly in Mountain View and Sunnyvale. It can be slow: replies in two threads say many approved candidates are still waiting to match, and one early career candidate took another offer before matching. In one team-hire loop, where the hiring manager ran Round 1, there was no team matching. Recruiters shared feedback with at least eight rejected candidates, from minor syntax errors to communication flagged as a concern, and one candidate whose L4 loop the committee moved down to L3 was given no appeal. Two candidates were told they could reapply after six and eight months.
What does Google ask in each round?
Google interviews span 7distinct round types, shown below. Counts reflect distinct questions per round across the loops we’ve indexed.
- •Preparing the follow-up, not just the base problem. The friend-log question moves on to unfriend events, Number of Islands to adding land and inland lakes, and Meeting Rooms to free slots and on-duty segments; one candidate says finishing the main problem is only the pass line.
- •Having several specific stories for the Googleyness interview. Reports list disagreement, a project you are proud of, a mistake, and too much work, with follow-ups that ask for more detail and what you would do differently.
- •Agreeing on the approach and edge cases before coding, then dry running your own tests out loud. Code is written in a document or plain editor without running it, and several candidates name clarifying and talking through their thinking as what interviewers look for.
- •Coding before aligning with the interviewer. One rejected candidate's feedback named missed corner cases and communication, and one L4 candidate who solved both onsite problems was rejected for a lack of communication.
- •Skipping the brute force when it is wanted. In both 2D clustering reports, the interviewer preferred or asked for the brute-force version after the candidate discussed faster neighbour searches.
- •Answering the Hiring Assessment's repeated work-style questions inconsistently. Five candidates advise picking a consistent work style, since the same traits come back in different wording.
- •Preparing DeepMind research rounds from research alone. Reports describe ML foundations questions, from overfitting to Transformer internals and normalisation, alongside the research discussion.
How to prepare for the Google interview
1. Take the Hiring Assessment with a consistent work style
Decide which work style describes you before you start, keep your choices consistent, and follow stated policy in the scenarios. Leave about an hour.
2. Practise coding in a plain document, out loud
Write solutions without running them, clarify the problem first, agree on the approach, then dry run your own test cases and state the complexity.
3. Drill the most reported problems and their follow-ups
The friend-log connectivity problem with unfriend events, Number of Islands with added land, Meeting Rooms variants, course schedule variants and car rental, each pushed to deletions, streams and input too large for memory.
4. Prepare six or seven Googleyness stories
Cover disagreement, a proud project, a mistake, too much work, mentoring and ambiguity, with numbers and reasons, plus one or two hypothetical product questions.
5. Ask which design round you have
L5 and above have system design; some L4 loops now have Solution Design; ML roles have ML design. For DeepMind, prepare to present your research and answer ML foundations questions as well as code.
FAQ
How long does the Google interview process take?▾
What is the Google Hiring Assessment?▾
Is there a coding online assessment?▾
How many rounds are in the Google interview?▾
Is the Google onsite in person?▾
Does Google ask system design at L4?▾
Can I use AI in the Google interview?▾
How is the Google DeepMind interview different?▾
How fresh is this Google interview data?▾
Get the full Google catalog
Every question, every candidate-reported follow-up, and what passers actually do. Monthly refresh.