Company Comparison Matrix

Nine loops, compared. The per-company pages carry the detail; this page is for deciding where to put your next ten hours when you have more than one process running.

What is solid here and what is not

Round structure is corroborated by recruiters and candidates and changes slowly. Pattern families are the durable signal. The problem lists on the individual pages are crowd-reported, stale, and biased toward whoever posted them — every table on this page is generated from the same profile data, so it inherits exactly that caveat and no more.

What each loop is actually testing

Company	The binding constraint	Rounds	Approx. interview minutes	Prepare, above all
Google	The follow-up ladder. The first solution is the entry fee.	4	~225	the $O(1)$-space variant of every pattern you know
Meta	Speed. Two problems in 45 minutes, so fluency beats depth.	4	~225	implementation speed and the stdlib
Amazon	Leadership Principles, weighted like the code.	4	~465	six to eight STAR stories
Microsoft	Clean, reviewable code, arrived at collaboratively.	3	~300	naming, edge cases, and test categories
Apple	Whatever the team decides — the least standardised loop here.	3	~210	your own CV, three questions deep
Netflix	Senior judgement and autonomy, not puzzles.	3	~270	the culture memo, and what you declined to do
Uber	Turning a scenario into a model. Graphs and design.	3	~240	weighted graphs and bounded-memory design
Bloomberg	Object-oriented design, and composite data structures.	3	~210	class design with an extension point
ByteDance / TikTok	Raw algorithmic depth. The hardest distribution here.	3	~300	hard problems and exact complexity bounds

The minutes column is the sum of each round’s count times its length, taking the low end of any range — so it is a floor, and it excludes recruiter screens. Amazon’s ~465 is not a mistake: it runs the longest loop here, and a large share of it is behavioural.

Which pattern families are shared

Counted across all nine profiles:

Appears in	Pattern families
4 of 9	hash-tables
3 of 9	grid-traversal-islands-and-flood-fill, linked-lists, strings, top-k-elements
2 of 9	binary-search-on-answer, binary-search-template-and-variants, design-iterators-and-flatteners, design-lru-and-lfu-caches, design-with-stacks-and-queues, greedy-interval-scheduling, merge-intervals, one-dimensional-dp, shortest-paths-dijkstra-bellman-ford-and-floyd-warshall, simulation-and-stateful-iteration, tries
1 of 9	advanced-dp-optimizations, binary-trees-and-bst, bit-manipulation-tricks, bitmask-and-tree-dp, design-trackers-and-feeds, graph-traversal-and-connected-components, matrix-and-grid-manipulation, monotonic-deque, monotonic-stack, number-theory-for-competitive-programming, segment-trees-and-lazy-propagation, sliding-window, sorting-with-custom-comparators, stack-parsing-and-expression-evaluation, string-algorithms-kmp-z-and-rabin-karp, sweep-line-and-event-counting, tree-bfs-and-level-order, tree-dfs-paths-and-sums, two-dimensional-dp-and-knapsack, two-pointers

The most striking thing here is what is missing: no family appears in more than 4 of the nine profiles. There is no universal core that every company is known for leaning on — which is the opposite of what “just do Blind 75” implies.

Two honest readings of that, and they point the same way:

• The profiles record emphasis, not coverage. Every one of these companies will happily ask a hash-map problem; the tag records that Google’s loop is notable for graph and DP laddering, not that Meta never asks about graphs. Absence of a tag is weak evidence.
• So the union is the syllabus, and the intersection is not. Preparing only the families in the top row would leave you exposed at every single company on this page. The per-company lists are for ordering your revision, not for scoping it.

sketch Which pattern families overlap across nine loops p5.js Pause Restart

Nine companies down the side, all thirty-six pattern families across the top, ordered by how many companies lean on each. The left of the grid is the shared core -- study it once and it pays off everywhere. The long right-hand tail is company-specific, and it is where preparation stops transferring. Every cell is read from the same companies.yaml the tables on this page are built from, and the counts are derived on screen.

Distinctive emphases

Families that appear in exactly one profile — the clearest signal of what makes a loop unusual:

Company	Only-here families
Google	graph-traversal-and-connected-components, monotonic-stack, sliding-window, two-dimensional-dp-and-knapsack
Meta	stack-parsing-and-expression-evaluation, tree-dfs-paths-and-sums, two-pointers
Amazon	tree-bfs-and-level-order
Microsoft	binary-trees-and-bst, matrix-and-grid-manipulation
Apple	bit-manipulation-tricks
Netflix	design-trackers-and-feeds, sorting-with-custom-comparators
Uber	sweep-line-and-event-counting
ByteDance / TikTok	advanced-dp-optimizations, bitmask-and-tree-dp, monotonic-deque, number-theory-for-competitive-programming, segment-trees-and-lazy-propagation, string-algorithms-kmp-z-and-rabin-karp

ByteDance accounts for 6 of these on its own — segment trees, DP optimisations, string algorithms, bitmask DP, monotonic deque, number theory. That cluster is a competitive-programming syllabus, and it is the single clearest profile difference in the whole matrix. If you are interviewing there and nowhere else, this course’s Phase 16 and 17 move up your list; if you are interviewing everywhere else, they move down.

Bloomberg has no only-here families, which is not a gap in the data — its distinctiveness is format rather than content. It shares the design and data-structure families with others and then asks you to express them as classes.

How to sequence a multi-company search

Situation	Where the next ten hours go
Several loops, no offers yet	The union of the shared families. Breadth first — a gap at one company is a gap at all of them
Amazon in the mix	STAR stories, immediately. It is the only loop here where the behavioural half alone can reject you
Google or ByteDance next	Hard problems and follow-up depth. Both punish a thin top end, for different reasons
Meta next	Timed two-problem drills. Fluency, not new patterns
Bloomberg next	Class design on problems you can already solve
Apple next	Ask the recruiter for the loop shape, then prepare that
Netflix next	The culture memo and your ownership stories. Least algorithmic loop here

“Breadth first” is worth putting a number on. Counting every (company, family) pair across the nine loops gives 58 in total. The 16 families that appear at two or more companies are only 44% of the 36 families on the list, but they account for 38 of the 58 pairs — 66% of the demand. Six families, taken in order of breadth, already touch all nine companies. Verified against companies.yaml, which is also what the tables above are built from.

That is the whole case for the shared core when you have several loops and no offer yet: the first twelve families cover 30 of the 58 pairs and the last twelve cover 12 — the same study time is worth 2.5× more at the front of the list than at the back.

sketch Study the shared core first: 44 percent of the families, 66 percent of the demand p5.js Pause Restart

Take all 58 company-family pairs across the nine loops as the total demand. Now study families in order of how many companies want them. The curve is the fraction of that demand covered, and it is steep at the start and flat at the end -- which is the whole argument for breadth before depth when you have several loops and no offer yet. The counts, the crossings and the curve are all computed from companies.yaml on screen.

Pitfalls

• Reading an absent tag as “they never ask this”. The profiles record emphasis, not coverage. No family appears in more than 4 of 9, so absence is weak evidence and the union is the syllabus.
• Preparing the intersection. It is nearly empty. Scoping revision to only the most-shared families would leave you exposed everywhere.
• Treating the problem lists as question banks. They are crowd-reported and stale. The families are the durable part; that is why they sit above the problem tables on every page here.
• Optimising for the company instead of the round. Amazon’s coding bar is lower and its behavioural bar is higher. Prepare for the round that can reject you, not the company’s reputation.
• Assuming a standard loop at Apple or ByteDance. Both vary — Apple by team, ByteDance by region. One recruiter question replaces a lot of guessing.
• Ignoring the shape of the follow-ups. Google ladders deeper on the same problem; Uber goes sideways into systems; ByteDance goes into complexity. Same pattern knowledge, three different second halves.

Interview follow-ups

They ask	What they’re checking	The answer
“Why do you want to work here specifically?”	Whether you researched the team, not the brand	Name what the team builds and connect it to something you have done. This is the question Apple’s variable loop makes most important
“How are you preparing?”	Whether your process is deliberate	Patterns rather than problem counts, a log tagged by pattern, timed drills, mocks near the end. Specific beats enthusiastic
“You interviewed elsewhere. How did it go?”	Candour and self-assessment	Be honest and specific about what you learned; name a gap you then closed. Never disparage another company’s process
“What is your weakest area?”	Self-awareness	Name one and what you are doing about it. A pattern-tagged log makes this answerable in a sentence
“Do you have offers or deadlines?”	Timeline, and leverage	Answer factually. Timelines are a normal part of the conversation; inventing one that does not exist is a bad trade

Recall card

• The loops differ in what they bind on, not in the patterns. Google ladders, Meta races, Amazon scores stories, Microsoft reviews code, Bloomberg wants classes, ByteDance wants depth.
• No pattern family appears in more than 4 of the 9 profiles. There is no universal core — the union is the syllabus, the intersection is not.
• Absence of a tag is weak evidence. The profiles record emphasis, not coverage.
• Amazon runs the longest loop (~465 minutes) and is the only one where the behavioural half alone can reject you. Prepare STAR stories first.
• ByteDance owns 6 of the only-here families — a competitive-programming cluster. Phases 16-17 matter there and almost nowhere else.
• Bloomberg’s difference is format, not content — the same families, expressed as classes.
• Apple varies by team, ByteDance by region. Ask the recruiter; it replaces a lot of guessing.
• Follow-ups differ in direction: Google deeper, Uber sideways into systems, ByteDance into complexity.