Your CAC Did Not Break. Someone Else Raised.

Acquisition cost is not a property of your funnel. It is a price, set inside an auction by whoever will pay the second-most for the same attention. When a competitor closes a round, that bidder’s ceiling moves, and your number moves with it while every input you control sits where you left it.

So the first question is not what to fix. It is which of four things moved: the auction, your funnel, the price of the underlying input, or your mix. Three belong to you. One does not, and it is the one most often written up as creative fatigue.

Why these three models

The Wire Model scores the features of a decision, routes them to a small ensemble of formal models, then makes the ensemble produce dated actions. What scored here:

• Price set by strategic counterparties inside a formal mechanism (0.9). You do not post a price for a click. You submit a bid into a rulebook, and the rulebook decides what you pay.
• Duration uncertainty (0.8). Every response depends on whether the elevated price runs six weeks or three years, and that answer sits on someone else’s balance sheet.
• Signal extraction from a noisy series (0.8). One number moved and at least four causes could have moved it.
• Historical-analog density (0.7). Funded entrants repricing a category is well documented, in filings you can read.
• Cognitive distortion (0.6). Real, and it points one way. Founders explain their own numbers with their own actions.

That routes to position-auction equilibrium for the mechanism, Markov regime-switching for the duration, and prediction-market price formation for the signal. Three outcome types, equilibrium, cycle-regime and random, so their errors do not line up.

Behavioral did not earn a card. Self-attribution has no lever of its own here and its correction is a control group, which the third model already prescribes. The governance question, whether your team will stop working on what the diagnosis clears, sits in the closing section.

The framework: a price you did not set, a state with a length, and a series with a signal in it

1. The mechanism: you pay the bid below yours

Sponsored search runs a generalized second-price auction. An advertiser pays a price per click equal to the bid of the advertiser in the position below, plus a minimum increment, typically a cent.¹ Varian formalises the consequence: an advertiser earns a profit per impression of its own value minus the bid of the agent immediately below, and in equilibrium every bid is fenced by that bid.²

Read that as a margin statement. Your margin on paid acquisition is your value minus another company’s bid. Nothing in it concerns your creative, your page or your product.

Your cost per click is another company’s bid, plus a cent.

Now add the funded entrant. Their ceiling is not their lifetime value. It is their round divided by the account target on a board slide. Eight million dollars and a promise to triple accounts in twelve months is a bid ceiling set by a commitment, not by a payback calculation, and they can hold it for as long as the cash lasts.

The useful conclusion follows from the same algebra. Raising your bid against a higher ceiling loses twice: you pay more for the positions you still win, and you still do not take the slot, because a commitment is not a valuation you can out-argue. The equilibrium response for a value bidder is to bid lower on the contested slot and let it go.

Do not expect the platform to intervene. Bulow and Klemperer showed that attracting one more bidder beats setting the optimal reserve price, and Ostrovsky and Schwarz tested that live on Yahoo sponsored search, where revenues rose substantially once reserves were reset upward.³ The seller of attention gains more from your new competitor than from anything it could do for you.

None of this needs a bidding interface. Mobile-money agent commissions are an auction. So is a WhatsApp reseller’s margin, a Kampala drive-time slot, a distributor’s shelf, a broker’s referral fee. The funded rival pays the incumbent price plus a step in each, and your acquisition cost moves in channels that have no dashboard at all.

Position-auction equilibrium, the mechanism lens

Assumes: a repeated auction with quality-weighted bids and rivals bidding somewhere near equilibrium.

Fits because: mechanism-set price scored 0.9, and the price you pay is defined by the bid beneath you.

Breaks when: the platform changes the rules, or bundles bidding into an automated objective so no advertiser sets a slot bid at all. Then the bid below you stops being an observable.

Counteracts: reading a cost line as a verdict on your marketing team.

May reinforce: fatalism, and neglect of quality score, which genuinely does move your price.

2. The duration: a spike and a new state look identical for three weeks

Treat the category as moving between two states with different price levels. Entry into the funded state is quick, because the spend starts when the money lands, often before the announcement you are reading. Exit is slow, and their capital governs it rather than your effort.

Uber wrote the transition rule into its own prospectus. Where a rival uses incentives to offset a network advantage, it said it would generally match those incentives even at a negative margin, and that if competitors shifted toward shorter-term profitability by cutting theirs, it would not need to invest as heavily.⁴ The state ends when their strategy changes. Not when your funnel improves.

Then check how long these states run. Seven years on, Uber’s annual report still carries the same risk factor: well-capitalized competitors in nearly every major geographic region, many offering discounted services, incentives, consumer discounts and promotions.⁵ In a large category the elevated state is measured in years.

The competitor’s runway is now your cost curve.

Treat that as a planning input. Estimate their runway from round size, disclosed hiring plans and the date of the raise. Ten million dollars on a thirty-month plan means the elevated price has roughly two years left unless discipline arrives early. Set your horizon against that number, not against your patience.

Markov regime-switching, the duration lens

Assumes: the category occupies a small number of states with different price levels, and transitions are exogenous to you.

Fits because: duration uncertainty scored 0.8, and the exit condition sits on the rival’s cash position.

Breaks when: there is no discrete state at all. Slow drift from many small bidders, a platform pricing change or plain seasonality can produce the same chart with nobody having raised anything.

Counteracts: budgeting a multi-year repricing as a one-quarter blip.

May reinforce: waiting out a state that will comfortably outlast your own runway.

3. The signal: read the price series the way you would read a market

A market price aggregates dispersed private information. That is why prediction markets work, and why manipulating them mostly fails: across the documented attempts, none had much discernible effect on prices except during a short transition phase, because the profit motive brings informed traders in on the other side.⁶

Your cost per thousand impressions is the same kind of object: a running estimate of what attention in your category is worth. When it rises, either the market learned something real, or a large bidder with a non-informational motive pushed it.

Then the analogy breaks, and the break is the point. A prediction market corrects the manipulator because the contract keeps trading. An ad impression is consumed. Nobody can sell that attention back to you at the informed price, so the distorted price is simply delivered, daily, for as long as the deep pocket keeps bidding.

Separate the causes before you touch anything. Acquisition cost is one price term over three conversion terms: cost per thousand impressions, divided by click rate, landing conversion and close rate. One of the four is the auction. Three are yours.

• Auction pressure. Impression cost up, click and conversion rates flat, impression share lost to rank climbing, your own brand terms getting dearer because somebody is bidding on your name.
• Funnel decay. Impression cost flat, click rate falling as frequency rises, or conversion falling after a pricing, page or onboarding change. Your deploy log carries the date.
• Input repricing. Everything moves at once, including channels with no auction in them: airtime, bulk SMS, USSD sessions, agent commissions, courier rates. Or the currency moved. You buy impressions in dollars and earn in shillings, so the cost in the currency you bank rose while nothing else did.
• Mix drift. Components flat, blend different. Organic fell and paid took a bigger share, or you opened a worse geography. The blended number moves when nothing inside it does.

Three cheap tests separate them. Compare a segment your rival contests against one they ignore. Overlay their announcements, the platform’s release notes and your own deploys on the cost chart and see which line moves first. Check whether paid, affiliate and even your recruiter’s quotes moved together, which points at the category rather than at you.

If it is still open, run a holdout, because the dashboard cannot settle it. Across fifteen large advertising experiments at Facebook, covering 500 million user-experiment observations and 1.6 billion impressions, observational methods built on the data advertisers normally hold often failed to reproduce the experimental effects, even after conditioning on extensive demographic and behavioural variables.⁷ Attribution panels are not an instrument for this question. A two-week geographic holdout is.

A number you cannot split into contested and uncontested is not a diagnosis.

Two series worth holding side by side. Meta reported impressions up 12% in 2025 and average price per ad up 9%, after 10% in 2024, and put the rise down to an increase in advertising demand.⁸ Jumia reported sales and advertising expense per order up 9.2%, to 83 cents from 76, while calling its campaigns targeted and performance-driven and holding spend flat at 2.4% of merchandise value.⁹ One number from the seller of attention, one from an African buyer of it, both up about nine points.

That is a base rate, not a cause. If your own acquisition cost rose by single digits last year, the null hypothesis is not that your funnel decayed.

Prediction-market price formation, the signal lens

Assumes: an observed price carries information, noise, and the footprint of large bidders acting for non-informational reasons.

Fits because: signal extraction scored 0.8, and attribution is the founder’s whole problem.

Breaks when: the series is too thin to carry a signal. Below a few hundred conversions a week, a 30% swing sits inside ordinary variance and means nothing.

Counteracts: treating one bad fortnight as a verdict on the team.

May reinforce: over-testing, and stalling on a call that an uncontested control segment would have settled in a week.

GEER: what you can still move when the price is not yours

Five channels carry this exposure: measurement, the bid, the segment, your own price, and the channel class. Work up from the reversible.

1. Split the cost series by channel, geography and term, ninety days back. One hour.
2. Pull impression share lost to rank, overlap rate and outranking share. Write the rival’s name on the page. One hour.
3. Build the three-line date overlay. Their announcements, the platform’s changes, your deploys. Half a day.
4. Cap the contested bid at your value bound and let the slot go. Free, reversible the same afternoon.
5. Move that budget to terms and geographies they do not contest. A day of setup.
6. Raise your own price. If the clearing price of attention rose nine points, yours can move too, and price is the only lever that changes contribution the same week.
7. Run the geographic holdout. Two weeks, one market, some nerve.
8. Open one non-auction channel. A distributor agreement, an agent tier, a reseller margin, a signed LPO pipeline. Negotiated prices do not move because somebody closed a round.
9. Shift spend to reactivating your own file. It is the one inventory nobody can outbid you for. A month of focus.
10. Match the bid. Most expensive, least reversible, last for a reason.

No-lever flag: if the contested slot is your only working channel and no uncontested segment exists, this is not an acquisition-cost problem. It is a single-counterparty distribution problem the funded rival just exposed. Cost out a second channel and put it in this quarter’s plan.

RADAR: the fortnight after the announcement

DO NOW, by T+3 days. Reversible, and correct under every scenario.

1. Decompose the last ninety days into impression cost, click rate, landing conversion and close rate, split by geography.
2. Pull the auction-position report and name the bidder who moved.
3. Build the date overlay and mark which line moved first.
4. Write your value bound as an actual number, and cap the contested bid at it.

HEDGE, by T+14. Cheap insurance against a wrong diagnosis.

1. Start the two-week holdout in one uncontested market.
2. Open one negotiated channel with a small committed budget.
3. Re-quote your three largest non-auction acquisition inputs, agent commission, reseller margin, media slot, and lock a six-month rate before the rival gets to them.
4. Test a price rise on the segment with the lowest churn risk.

DEFER AND TRIGGER. Irreversible, so pre-commit the observable now.

1. Defer: matching the bid at scale, a brand campaign funded from the same line, a new market entry, any growth hire priced off the old acquisition cost.
2. Trigger to match: the rival holds the contested slot for eight consecutive weeks, your uncontested segments are exhausted, and payback on a matched bid still lands inside your cash runway. Then match in one geography, for one quarter, with the stop date written first.
3. Counter-trigger: at T+28, impression cost is up while click and conversion rates are flat in both sets. Stop the creative work. It was never the creative.

From the other side of the table. DO NOW: ask for the acquisition series split into contested and uncontested, never blended. HEDGE: date the drift against the competitor’s announcements rather than the company’s releases, because the two imply opposite decisions. DEFER: your markdown. A capital-efficient company in a category you still believe in is cheapest in the quarter its acquisition cost looks worst for reasons that were never its own.

CHAIN: what the filings say a funded entrant does to a category

The comparison set is not other companies in your sector. It is any market where a bidder with an exogenous budget entered a repeated auction against incumbents pricing off unit economics. Ride-hailing, food delivery, online grocery, mobile-money agent recruitment and telecom subscriber acquisition all sit in it.

The recorded default is matching. Uber told its own investors it would generally match a rival’s incentives even at a negative margin.⁴ That is the base rate for behaviour, and it is why the state lasts: matching funds it from both sides. Seven years on, the same competitive language is still in the annual report.⁵

Present conditions cut both ways. Capital is dearer than in 2021, so rounds are smaller, ceilings lower and states shorter. Against that, platforms keep folding bidding into automated objectives, which removes the slot-level control the first model depends on, and Meta’s disclosure shows price per ad rising while impressions grew, so added supply is not relieving the price.⁸

Now net out what would have happened anyway. Some of the rise is category drift and some is seasonal. Deduct the uncontested segment’s movement from the contested segment’s, and what remains is the portion with a name attached. It is usually smaller than the founder’s first estimate and larger than the board’s.

Matrix-break flag. All three models assume a stable mechanism with a visible bid. If platforms complete the shift to fully automated bidding, where you submit a target and a budget and the machine bids, the bid below you stops being observable and the position reports degrade into estimates. The holdout then stops being the last test on the list and becomes the only instrument you have.

Where this ensemble goes blind

The ensemble is silent on three things, and one of them is the thing you most want to know.

Whether their bid is disciplined or panicked. The models price the effect, never the intent. A well-run funded rival and a desperate one produce an identical impression cost this month and completely different ones in nine.

Your own contribution to the price. These auctions are quality-weighted, so part of what you pay is your relevance. The first model treats that as fixed. It is the one input inside the mechanism you actually own.

Whether the customer got more valuable. If the market genuinely learned something, the higher clearing price is correct and your ceiling should rise. Read this piece as an attack and you will underbid a market that just improved.

Underneath all three sits the governance problem. A diagnosis that clears the funnel is an instruction to stop work people have already started, on a channel they are measured on. Sunk effort has gravity, so the finding has to arrive with a decision attached or it gets absorbed and ignored.

So here is the decision. By T+3, split the acquisition series into contested and uncontested and put one number on each. If the uncontested cost is flat and the contested cost is up, the funnel is fine, the creative work stops this week, and the contested bid gets capped at your written value bound. That call holds whatever the three gaps above turn out to say.

Sources and notes

1. Edelman, B., Ostrovsky, M., and Schwarz, M. “Internet Advertising and the Generalized Second-Price Auction: Selling Billions of Dollars Worth of Keywords.” American Economic Review 97(1), 2007, 242-259. Author copy: full text. The mechanic quoted is in Section I: “an advertiser in position i pays a price per click equal to the bid of an advertiser in position (i + 1) plus a minimum increment (typically $0.01).”
2. Varian, H. R. “Position Auctions.” International Journal of Industrial Organization 25(6), 2007, 1163-1178. Author copy: full text. The profit expression for position s and the statement that in equilibrium each agent’s bid is bounded above and below by a combination involving the bid of the agent below him are in Section 2 and Section 2.1.
3. Ostrovsky, M., and Schwarz, M. “Reserve Prices in Internet Advertising Auctions: A Field Experiment.” Journal of Political Economy, 2016. Author copy: full text. The Bulow and Klemperer (1996) result, that adding just one more bidder with a zero reserve price is always preferable to setting the optimal reserve price, is quoted in the introduction; the Yahoo field experiment result on revenues follows in the same section.
4. Uber Technologies, Inc., Form S-1 registration statement, filed 11 April 2019. Filing. The matching language is in the business section: “we will generally choose to match these incentives, even if it results in a negative margin,” alongside the conditional that if competitors reduce their incentives, “we believe that we would not be required to invest as heavily in incentives.” Note: sec.gov serves 403 to a spoofed browser user agent and 200 to an honestly self-identifying one, per SEC access policy. A normal browser reads it fine.
5. Uber Technologies, Inc., Form 10-K for the fiscal year ended 31 December 2025, filed 13 February 2026. Filing. Risk factors: “well-capitalized competitors in nearly every major geographic region,” and “Many of our competitors are well-capitalized and offer discounted services, Driver incentives, consumer discounts and promotions.” Same user-agent note as above.
6. Wolfers, J., and Zitzewitz, E. “Prediction Markets.” NBER Working Paper 10504, 2004. Full text. Under “Can Event Markets Be Easily Manipulated?”: “There have been several known attempts at manipulation of these markets, but none of them had much of a discernible effect on prices, except during a short transition phase.”
7. Gordon, B. R., Zettelmeyer, F., Bhargava, N., and Chapsky, D. “A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook.” Marketing Science, 2019. Author copy: full text. The 15 experiments, 500 million user-experiment observations and 1.6 billion ad impressions, and the finding that observational methods often fail to reproduce the experimental effects, are all in the abstract.
8. Meta Platforms, Inc., Form 10-K for the fiscal year ended 31 December 2025, filed 29 January 2026. Filing. “In 2025, ad impressions delivered increased by 12%, as compared with an increase of 11% in 2024… In 2025, the average price per ad increased by 9%, as compared with an increase of 10% in 2024… The increase in average price per ad in 2025 was driven by an increase in advertising demand.”
9. Jumia Technologies AG, Form 20-F for the fiscal year ended 31 December 2025, filed 24 February 2026. Filing. “Sales and advertising expense increased by 12.1% from $17.3 million in 2024 to $19.4 million in 2025… Sales and advertising expense per Order increased by 9.2% to $0.83 in 2025, compared to $0.76 in 2024. As a percentage of GMV, Sales and advertising expense remained constant at 2.4% in both 2024 and 2025.”