The last piece in this series derived a floor. In a market with zero cost, perfect fit and equal numbers on both sides, 36.79% of possible matches still fail. Every supplier in that model was identical, which is what made the arithmetic clean. It is also the one assumption a brand exists to break. Break it, and the same equation says a supplier holding 5% of a market's attention ends a period with nobody 62 times less often than a supplier holding a fair share. Procter and Gamble put $9.2 billion into advertising last year to hold positions like that. What follows is which part of that money buys something unavailable by any other route, and which part is lying in the open.

Start with the question most people actually ask, because it turns out to have a testable answer. When a very large brand wins, is it winning because the product is better?

The assumption the last piece made on purpose.

The urn-ball model (a hundred balls thrown into a hundred buckets, counting how many buckets stay empty) assumes every ball is thrown at random. No bucket is more attractive than any other. That assumption is doing enormous work. It is what produces 1 over e, and it describes a market in which no supplier has any prior claim on anyone's attention.

Real buyers do not throw at random. They throw at what they can already name.

So replace the uniform throw with a weighted one. Each supplier now carries a probability of being chosen, call it p, its share of consideration (the chance that a given buyer, at the moment of choosing, thinks of you at all). A supplier with no reputation holds whatever a fair split gives them, 1 divided by the number of suppliers. A brand holds more. Everything in this piece falls out of that one substitution.

The chance a supplier finishes a period with nobody becomes (1 minus p), raised to the power of the number of buyers.

That is the entire model. One line.

What share of consideration actually buys.

Put a hundred buyers into a market against a hundred suppliers, hold everything else identical, and vary only p.

At a fair share of 1%, a supplier gets nobody 36.60% of the time, which is the floor from the last piece arriving on schedule. At 2%, that falls to 13.26%. At 5%, it falls to 0.59%. At 10%, it is indistinguishable from zero.

Read those twice. Five times the share of consideration does not produce five times the result. It produces sixty-two times fewer failures. The returns are superlinear (the output rises faster than the input that produced it), and that single property explains why the spending looks irrational from outside and is perfectly rational from inside.

A big brand does not take its share. It takes the tail.

Now watch what the same arithmetic does to everybody else. Let one brand hold a share w, and let the remaining ninety-nine split whatever is left.

When the big brand holds 1%, everyone fails 36.60% of the time and nobody has an advantage. When it holds 10%, its own failure rate is 0.003% and the others have moved to 40.12%. When it holds 40%, its failure rate is effectively zero and the others sit at 54.45%.

The big brand's position improves by very nearly everything available. Each individual small supplier is worse off by about eighteen points. The prize is concentrated in one place and the damage is spread thinly across ninety-nine, which is precisely why nothing ever organises against it. This is the congestion externality from the last piece wearing different clothes. The extra share is not stuck in the jam. It is the jam, for everybody behind it.

How much is actually enough.

Here is the part that makes this a workable problem rather than a complaint.

You do not need a large brand's share. You need enough to clear a threshold, and the threshold has a closed form. Set the failure rate to a coin flip, solve, and the required multiple of a fair share comes out as the natural logarithm of 2, multiplied by the number of suppliers per buyer.

In a market with one supplier per live buyer, you need 0.7 times a fair share. You are already past it. At five suppliers per buyer, the crowded case from the last piece where 81.87% of suppliers end up empty, you need 3.5 times a fair share. At ten per buyer, 6.9 times.

Three and a half times. Not four hundred times. In the crowded market the last piece spent seven thousand words describing, the distance between invisible and a coin flip is a factor of three and a half in one variable.

That number is the reason this piece exists.

So is it the product, or is it the fame?

Everything above is arithmetic. Arithmetic can tell you that a difference in share of consideration produces a violent difference in outcome. It cannot tell you where that difference came from. If large brands hold attention because their products are genuinely better, the advice is simple and boring: make a better product.

Marketing science has been testing exactly that question for sixty years, and the answer is not the flattering one.

The finding is called double jeopardy (the observation that a smaller brand is punished twice, once for having fewer buyers and again because those buyers are also slightly less loyal). William McPhee described the pattern in 1963, Andrew Ehrenberg found that it generalised to brand purchasing, and Ehrenberg, Goodhardt and Barwise formalised it in 1990. It holds, with few exceptions, across a very wide range of categories, countries and time periods.

The load-bearing part is what it implies. If product quality were driving brand advantage, loyalty would vary independently of size. A small, excellent brand would show low penetration and high loyalty. That combination is rare to the point of being an anomaly. Loyalty tracks size. The Ehrenberg-Bass Institute states the consequence plainly: growth comes from acquiring category buyers rather than from holding on to the ones you have, and acquisition is roughly twice as important as reduced defection.

Which means large brands do not look more loved because they did something clever about being loved. They look more loved because they are large.

There is a second reason to treat this as structure rather than opinion. The model underneath double jeopardy is the NBD-Dirichlet (Goodhardt, Ehrenberg and Chatfield, 1984), which predicts a brand's performance measures from its market share alone. NBD stands for negative binomial distribution, which is a Poisson process with the rate allowed to vary from person to person. The last piece derived its floor from a Poisson process. This is not a framework borrowed from a neighbouring field. It is the next term in the same equation.

What the largest brand in the category actually spends.

It is worth being exact here, because the folk version of this story is wrong in a way that will lose you the room.

Procter and Gamble's annual report for the year ended 30 June 2025 states it directly. Advertising costs, charged to expense as incurred, include television, print, radio, digital and in-store advertising expenses, and were $9.2 billion in 2025, $9.6 billion in 2024 and $8.0 billion in 2023.

Against roughly $84.3 billion in revenue, that is 10.9%. The filing is explicit that consumer promotions, product sampling and aids sit outside that figure, on top of it.

For comparison, the most recent CMO Survey puts marketing budgets at 9.0% of revenues, and that number covers the entire marketing budget rather than advertising alone.

So the large brand is not quietly spending less. On the single line that buys share of consideration, it spends more than the average company spends on everything. Anybody planning to argue that the giants win through restraint should stop here.

What they do differently is not the amount. It is the timing, and the target.

The $200 million that went missing, and the 10% that appeared.

In 2017, Procter and Gamble removed roughly $200 million from its digital advertising, about $100 million in the quarter to June and another $100 million through December. The first tranche, by the company's own account, had little appreciable impact on the business. The reduction was reported to have increased the company's reach by 10%.

They spent less and were seen by more people.

One measurement from that review explains it. The average dwell time on an advertisement in a mobile newsfeed was 1.7 seconds.

The money had been going into precision. Narrower targeting, more intermediaries, more measurement of an event lasting under two seconds. Removing the precision did not remove the presence. It removed the tax on the presence.

This is the single most useful thing in this piece for anybody without a budget, because it is an experiment run at nine-figure scale on the question you cannot afford to test yourself: which half of the spending was carrying the value? The answer was the cheap half. Broad, unglamorous presence carried it. The expensive apparatus in front of it came out without loss.

Reach, retrieval and fit.

Share of consideration is not one thing, and taking it apart is where a small supplier finds a lever. It separates into three factors that multiply together.

Reach is the fraction of buying situations you are present in at all. Bought, with money, at scale. It is the nine billion dollars. Treat it as unavailable. Retrieval is the chance you are brought to mind, or returned by a system, given that you are present. Partly bought, partly engineered. Fit is the chance you survive the shortlist once retrieved. Almost entirely a function of whether your specifics are legible, precise and checkable. Free.

Hold reach completely fixed. Change nothing about budget or presence. Move only the last two terms, in a market running five suppliers per buyer.

A supplier at 0.2% share fails 81.86% of the time. Three times better on retrieval and fit puts it at 54.78%. Five times better puts it at 36.60%.

Same money. Same reach. From four failures in five to fewer than two in five. That is the three-and-a-half-times threshold from earlier, arrived at from the other direction, and it sits entirely on the terms that carry no price.

The ninety-five per cent they are paying for and you are not.

Now the timing, which is the real difference between the two kinds of company.

Buyers on a five-year purchase cycle are in the market for a small fraction of any given period and out of it for the rest. The Ehrenberg-Bass work behind the 95:5 rule puts it at about 5% in a given quarter. The last piece worked that as a duty cycle: an annual event plus a note to the base leaves you present for roughly 1 of the 20 quarters in which the decision actually gets made.

A large brand's advertising is not aimed at the 5% buying today. It is aimed at the 95% who are not, so that the memory is already in place when their moment arrives. It is not a purchase of demand. It is the purchase of an asset that pays out later.

The small supplier does the reverse, almost always for a defensible reason. Nothing is spent during the 95%, because nothing is happening and budget is short. Then a live opportunity appears and money goes into interrupting a buyer at the most crowded and most expensive moment there is, against everybody else who also just noticed.

That is the whole asymmetry, and it needs no numbers at all: they pre-pay, you pay at the counter, and the counter is where the price is highest.

Why a machine changes the arithmetic.

Which brings us to the one genuinely new thing, and the reason any of this is actionable rather than merely true.

Mental availability lives inside human memory, and there is no way to install yourself in a few million of those without a few billion dollars. If human recall were the only retrieval mechanism, this piece would end with a shrug and an invoice.

It is no longer the only mechanism, and the replacement has a different failure mode.

An ungrounded language model composing an answer out of what it absorbed in training is a machine that has read the internet, and the internet over-represents large brands enormously. Asked to compare suppliers using nothing but its own memory, it reproduces the fame distribution. It hands the incumbent the shortlist at no charge. Ungrounded generation amplifies the brand advantage rather than levelling it, which is the last piece's warning restated: if the evaluating model is ungrounded, it composes the comparison, with a purchase order attached.

A grounded system does something else. It queries live records at the moment of the question. It holds no memory of who is famous, no impression of who is established, no accumulated sense that one name feels safer than another. It has a query and an index. A name absent from the index does not rank low. It is simply not there.

That is the opportunity, and it is narrow and specific. Retrieval by a machine is the only channel found so far in which p is not weighted by fame. An index is built once and answered from at the moment of the question, at effectively no marginal cost. It is a small supplier's version of pre-paying, the same structural move a large brand makes with memory, executed against a system that can be written to rather than a population that has to be persuaded.

The window is open because it is early. It will not stay open, because indexes fill.

What actually moves the number.

Stop paying for precision before you have presence. The nine-figure test has already been run, and narrow targeting in front of a 1.7-second impression was the removable part.

Be present during the 95%, rather than louder during the 5%. The duty cycle is the lever, and it costs time rather than money.

Make your specifics machine-legible. Prices, stock, minimums, lead times and capabilities, in a form a system can retrieve and quote. This is the fit term, and it has no price attached.

Concentrate rather than spread. Match probability depends on density in the one place a buyer actually looks, and a flat budget across seven channels thins the reagent in all seven.

Stop trying to be recalled and start being retrievable. The first needs a budget you do not have. The second needs a decision you have not made.

None of this closes a nine-billion-dollar gap. It does not have to. The threshold was three and a half times, not four hundred.

The house that was built for this.

Every lever above is a cost removed from somebody else, or a term moved that money cannot reach. That is the design brief, and it is why the infrastructure looks the way it does.

70,000+ products across 200+ brands in an MCP-ready catalogue, queryable by an agent in real time, so the answer is retrieved from live records rather than composed from memory. A costed answer on the page, with the working shown, instead of a range and a phone call. Blind-ship fulfilment with NDA-standard discretion. Single-source accountability, strategy through manufacturing through delivery. One operator, one line, one number to call when it matters.

We do not add noise to a saturated market. We answer the question at the moment it is asked, from the record.

If the last piece told you the market is structurally against you, this one puts a number on the distance and names the two terms that do not require a budget. The gap is real. It is also smaller than it looks, and it is not made of product quality.

If you are building from what is left, you are not finished.

Jenny Huang Goodman MPA, MSc, MHSA Principal Huang Goodman | Hako Shikin | POPS4 Virginia Beach, since 1997 ASI #217876 | DUNS 18-204-6339 (757) 390-0424 | jenny@huanggoodman.com