Why Consensus Compresses Returns

Key takeaways

• In its best year, Zillow's Homes segment turned $1.715 billion of revenue into $80.6 million of gross profit, or 4.7% before a dollar of operating expense (Zillow's fiscal 2021 Form 10-K, fiscal 2020 figures). That is the margin a crowded consensus leaves behind.
• Volume does not rescue a compressed spread: Zillow sold 189% more homes in 2021 than in 2020 and the Homes gross margin went negative.
• A real edge usually carries a price of admission rivals decline to pay. Costco prints its model in a public 10-K every year; almost nobody copies an 11.1% merchandise gross margin.
• The test for stubbornness is a named falsifier with a date on it. A position that has absorbed three years of contrary evidence without moving a decimal is a belief, not a thesis.

The price of agreement

Returns are a residual. They are what is left after everyone capable of doing the same thing has done it. This is not a moral claim about crowds and it is not a recommendation to be different for its own sake. It is arithmetic. Whatever spread exists between what an asset costs and what it sells for is available to every party who can see it and fund it, and each additional participant bids the buy side up, the sell side down, or both. The spread does not vanish because anyone made a mistake. It vanishes because the thesis became widely held.

The cleanest recent demonstration is instant home-buying, where two well-capitalised public companies executed the same widely agreed thesis in the same years and both left an audited record of what the spread was actually worth. Zillow's fiscal 2021 Form 10-K reports Homes segment revenue of $1,715.4 million for 2020 against Homes cost of revenue of $1,634.8 million, leaving $80.6 million of gross profit on 5,337 homes sold, about $15,100 per house, or 4.7% of revenue, which Zillow reported as a 5% gross margin. That is before sales and marketing, before technology, before general and administrative, and before the cost of the capital tied up in the inventory. And 2020 was the good year.

The following year the company scaled into it, into a rising market. Homes sold rose 189% to 15,436 at an average selling price of $387.6 thousand, up from $320.5 thousand in 2020; Homes revenue rose 251% to $6,015.8 million, and Homes cost of revenue was $6,106.9 million, a gross loss of $91.2 million, reported as a (2)% margin. That figure includes a $211.0 million inventory write-down on homes still held at December 31, 2021. On November 2, 2021 Zillow decided to wind Zillow Offers down; group net loss for the year was $527.8 million.

The control experiment sits in Opendoor's fiscal 2022 Form 10-K. Same thesis, same years, different company, and no wind-down. In 2021 Opendoor sold 21,725 homes on $8,021 million of revenue at a 9.1% gross margin, and its own Contribution Margin, a non-GAAP measure it defines as gross profit adjusted for inventory valuation and then reduced by direct selling and holding costs on homes sold, was 6.5%. Net loss that year was $662 million. In 2022 it sold 39,183 homes on $15,567 million of revenue, gross margin fell to 4.3%, Contribution Margin to 3.4%, and net loss reached $1,353 million with a $458 million current-period inventory valuation adjustment.

Read the two filings together and the lesson is not that either team was incompetent. It is that the best available spread on a thesis everyone agreed with, in the best possible market conditions, was a single-digit gross margin on an asset that costs real money to hold. There was never enough room. The consensus did not make the business hard to run; the consensus was the thing that had already taken the margin.

What compression looks like before the income statement shows it

Margin compression arrives in the accounts last. By the time gross margin has visibly rolled over, the crowding happened eighteen to thirty-six months earlier and the decisions that mattered were made then. The useful skill is reading the leading indicators, all of which are visible from inside an operating business and none of which appear on a financial statement.

The first is the auction. Whatever you buy (clicks, leads, inventory, small businesses, drivers, warehouse space) watch the price of the marginal unit rather than the average. Averages lag because they contain your old, cheap contracts. The second is terms drifting toward the counterparty: suppliers shortening payment terms, customers winning month-to-month where they used to sign annual, sellers of businesses in your niche running competitive processes where they used to take one call. Terms move before prices do, because terms are what people concede when they still want to hold the headline number. The third is the composition of your competitors: when the newcomers are well-funded and professional rather than amateur, the era of easy spread is over. The fourth is social: when the strategy has a name, a conference track, and a cohort of people teaching it, the pricing of it is finished.

Illustrative only: a lead-generation agency signs clients at a $1,800 acquisition cost against $6,000 of gross profit over twenty-four months, a ratio of 3.3x. Eighteen months later the same ad auction costs $3,200 per signed client, and retained gross profit is $5,100 because two competitors now offer month-to-month contracts. The ratio is 1.6x. Nothing about the operator got worse. The founder still runs the same playbook with the same skill. What changed is that the number of people running the identical play roughly doubled, and both sides of the ratio moved against them at once, which is the signature of crowding rather than of execution failure.

The levers, once you see it early: buy the input on a longer contract before the crowd arrives, move to a channel or a customer segment where the auction is thinner, change what you sell so that the comparison shopping stops working, or take the cash out and redeploy. All four are easier at the 3.3x reading than the 1.6x one, which is the entire argument for watching leading indicators.

The risk is the opposite error, and it is common: reading normal competitive noise as terminal compression and abandoning a fine business two quarters early. The distinguishing test is whether the pressure is on price alone or on price and terms and mix simultaneously. One input getting more expensive is a cost problem. Every input and every term moving the same direction at once is a structural one.

An edge that survives being published

If consensus destroys returns, the natural question is what kind of advantage does not get destroyed. The answer is not secrecy. Most durable edges are fully visible and stay durable anyway, because the cost of copying them is one competitors rationally decline to pay.

Costco's fiscal 2025 Form 10-K, for the 52 weeks ended August 31, 2025, is the plainest published example available. Net sales were $269,912 million against merchandise costs of $239,886 million, a merchandise gross profit of $30,026 million, or 11.1% of net sales. Selling, general and administrative expense was $24,966 million. So the retailing operation, after paying for itself, contributed $5,060 million of operating income. Membership fees were $5,323 million. Total operating income was $10,383 million, meaning membership fees accounted for 51% of it. Net income was $8,099 million. The renewal rate was 92.3% in the US and Canada and 89.8% worldwide, across 81.0 million paid members paying a base annual fee of $65 in the US.

Every one of those numbers is filed publicly and has been for decades. There is no proprietary technology in the structure and no trade secret in the arithmetic. A competitor could adopt it tomorrow. Almost none do, and the reason is legible in the same numbers: adopting it means voluntarily capping merchandise gross margin at roughly 11%, which is a fraction of what a conventional grocer runs, and then waiting years for a renewal base large and loyal enough to make the fee line carry half of operating income. The imitation cost is paid immediately and in full; the return arrives late and only if the loyalty materialises. That trade is available to everyone and attractive to almost no one, which is precisely why it does not get competed away.

Generalise the shape. A durable edge is rarely an opinion others have not had. It is usually one of: a cost position rivals cannot reach without dismantling their existing business, a constraint you are willing to accept that they are not, a distribution or trust relationship that took years of unglamorous work, an information advantage that regenerates rather than being consumed once, or a time horizon longer than the one your competitors are measured against. Notice that four of those five are about willingness rather than knowledge.

So the diagnostic question for any claimed edge is not is this a good idea. It is: is there a reason my competitors CANNOT do this, or only a reason they HAVE NOT? Cannot is an edge. Have not is a head start, and head starts have a known expiry date. If the honest answer is have not, the strategy is to bank the abnormal returns while they last and to build something with a cannot underneath them before they end.

Edge or stubbornness: three questions and a clock

The uncomfortable fact about minority positions is that being alone feels identical whether you are early or wrong. There is no internal signal that distinguishes them. Conviction, frustration at being misunderstood, the sense that the market will come around: all of it feels the same from inside a correct thesis and a doomed one. The distinction has to be built externally, in writing, before you need it.

The first question is mechanism. State the thesis as a causal claim, not a preference: not this category is undervalued, but this category is undervalued BECAUSE buyers cannot compare offerings on a single dimension, and that will change when X publishes standardised data. A thesis with a named mechanism can be checked. A thesis without one is a mood, and moods cannot be wrong, which is exactly the problem.

The second question is the falsifier. Write down, in advance, the specific observation that would make you abandon the position: a number, a competitor behaviour, a customer response, a regulatory event. It must be something that could plausibly happen. If you cannot name one, you do not have a thesis; you have an identity, and identities are defended rather than tested. Most people can produce a falsifier if forced, and the exercise itself is diagnostic: the ones who take twenty minutes and produce something vague are usually the ones already in trouble.

The third question is the clock. Attach a date. Early and wrong are separated only by time, so the thesis must say when the mechanism should begin operating, not when you win, but when the first evidence of the mechanism should be visible. Then, on that date, check whether the mechanism has started at all. A thesis that is right but slow shows early mechanism evidence and late results. A thesis that is wrong shows neither, and shows nothing on the next date either.

Illustrative only: an operator holds that small manufacturers will move procurement online, and buys inventory and builds a catalogue against it. The written thesis names the mechanism, that buyers switch when the online price plus lead time beats their incumbent rep, sets the falsifier as fewer than 15% of a hundred surveyed buyers requesting an online quote within eighteen months, and sets the check date. At eighteen months the figure is 9%. The disciplined response is not to abandon manufacturing; it is to accept that the mechanism has not started, stop adding inventory against it, and re-underwrite. The undisciplined response is to explain why the survey was unrepresentative, which is what stubbornness sounds like from inside, every single time.

The practical rule that follows: a position that has absorbed three consecutive years of disconfirming evidence without any change to its numbers is not a contrarian bet. It is a belief with capital attached to it. Real theses move: they get narrower, or the timeline extends with a stated reason, or the size changes. Immobility under new information is the tell.

Holding a minority position for years without becoming a crank

Assume you have done the work and the thesis survives its own falsifier. You now face the actual difficulty, which is operational rather than intellectual: being right early is expensive, lonely, and takes longer than your cash allows unless you design for it.

Size so that being wrong is survivable and being right is meaningful. Those are two constraints, not one, and most people satisfy only the second. The specific failure is committing the capital that funds the thesis and the capital that funds payroll to the same bet, which converts a timing question into a solvency question. Keep the two separate and a thesis that is three years slow is an inconvenience; merge them and it is fatal.

Separate reversible bets from irreversible ones and pay for optionality wherever it is cheap. Rent capacity before buying it. Sign a two-year lease rather than a ten. Use a contract manufacturer before building a line. The point is not caution. It is that the cost of being early is paid in fixed commitments, and a thesis proved right in year four is worth nothing if the infrastructure for year one had a fifteen-year payback.

Keep at least one credible person who disagrees with you and who can still get your attention. Minority positions attract two audiences: people who dismiss you, whom you learn to discount, and people who agree with you, who are usually agreeing for worse reasons than yours. Neither group updates you. What you need is a specific person who understands the thesis, rejects it, and is close enough to tell you when the falsifier has quietly triggered, because you will not notice it yourself. Contrarians almost never fail by getting the analysis wrong; they fail by not noticing when the analysis stopped applying.

Finally, separate being contrarian on facts from being contrarian on price. Disagreeing with the crowd about what is true is a research claim and can be settled with evidence. Disagreeing about what something is worth is a valuation claim and can stay unsettled for years. The first kind is where careful work pays reliably. The second is where being right and being solvent stop being the same thing. General education, not financial or business advice: the material here is a method for testing your own reasoning, not a recommendation about any market or asset.

Put it to work

Pick the assumption your plan shares with everyone else in your market. Write it as a claim with a mechanism, one observation that would disprove it, and a date to check. Then measure your gross margin against the crowd's: if it is thinning while volume grows, you are buying agreement, not advantage. Re-read the note on the date you set.

Sources & references

Linked entries open the named source directly. Entries without a link say exactly what kind of reference they are — and how to check them yourself.

• Zillow Group, Form 10-K for fiscal 2021 (filed February 10, 2022) — Primary source for Homes segment revenue and cost of revenue for 2020 and 2021, the 5,337 and 15,436 homes sold, the $320.5 thousand and $387.6 thousand average selling prices, the $211.0 million inventory write-down, the November 2, 2021 wind-down decision, and the $527.8 million net loss.
• Opendoor Technologies, Form 10-K for fiscal 2022 (filed February 23, 2023) — Primary source for 2021 and 2022 revenue, homes sold, gross margin, the company-defined non-GAAP Contribution Margin, net loss, and the $458 million current-period inventory valuation adjustment.
• Costco Wholesale, Form 10-K for fiscal 2025 (52 weeks ended August 31, 2025) — Primary source for net sales, merchandise costs, SG&A, membership fees, operating income, net income, the 92.3% and 89.8% renewal rates, 81.0 million paid members, and the $65 US base annual fee. The 11.1% merchandise gross margin and the 51% membership-fee share of operating income are arithmetic on those reported lines.
• Worked examples in this briefing — The lead-generation agency ratios and the online-procurement thesis are illustrative composites written for this library, not measurements of real companies.