Cost-Per-Citation: Pricing Earned Media Against AI Ad CPMs

TL;DR

→  The number everyone quotes — “earned media costs £X per AI citation” set against a ~$25 ad CPM — is a category error. It prices a durable asset as if it were a flow cost, like quoting the cost-per-mile of a car before you know how far you’ll drive it.

→  Earned and paid have different cost shapes: earned is high fixed cost with near-zero marginal cost (average cost falls with every exposure); paid is near-zero fixed cost with constant marginal cost (flat). They cross exactly once.

→  Compare on cost per outcome, not CPM. The CPM denominator is rigged against earned, because the citation sits inside the trusted answer while the ad sits beside it — they convert at completely different rates.

→  The crossover is set by the citation’s expected lifetime-breadth, risk-adjusted with a citation half-life. Below it, rent; above it, build — and the build margin widens the longer the citation lives.

The benchmark everyone quotes is the wrong number

A cottage industry of benchmarks has appeared to answer a natural question: what does it cost to get cited in an AI answer, and is that cheaper than buying a sponsored placement? The benchmarks usually land on a figure — some cost-per-citation in pounds — and set it beside ChatGPT’s roughly $25 CPM to declare a winner. The figure is worse than imprecise; it is measuring the wrong kind of thing. A CPM is a flow price: you pay it again for every thousand impressions, forever. An earned citation is a stock — a one-time cost that then produces exposures at no further charge for as long as the citation survives. Comparing the two directly is like comparing the price of a taxi per mile to the sticker price of a car, and concluding the car is expensive because it costs more than one ride.

It is worth being precise about the paid figure the benchmarks reach for, because it is real and moving. ChatGPT’s sponsored inventory opened on a rate-card CPM near $60 with six-figure managed minimums, then fell as self-serve buying and a Criteo partnership arrived, settling toward roughly $25 in observed spend and dipping lower in thinner categories. That is a legitimate flow price, and its very concreteness is the trap: it is quotable to the penny, refreshes on every impression, and invites a like-for-like number on the earned side that cannot honestly exist. The earned figure quoted back at it is a fixed cost wearing the costume of a rate. The two look like the same kind of thing and are not, and every downstream mistake grows from treating them as though they were.

The car analogy is exact, and it is the whole article in miniature. The sticker price of a car tells you nothing about its cost per mile until you divide by the miles you will drive. Drive it once and it is the most expensive journey imaginable; drive it 100,000 miles and it is nearly free per mile. A backlink or earned mention that produces a citation works the same way: its true unit cost is the earn cost divided by all the future answer-exposures it will generate, and that number is not knowable from the earn cost alone. Anyone quoting a single cost-per-citation has quoted a car’s sticker price and called it a fare.

So the first move in pricing earned against paid is to stop asking “what does a citation cost?” and start asking “what does a citation cost per unit of what it produces, over its life?” That reframing does real work, because once both channels are expressed as a cost per unit of output, they stop being two prices to compare and become two cost curves with different shapes — and curves with different shapes cross, which means the answer is never “earned is cheaper” or “paid is cheaper” in general. It is one or the other depending on exactly where you sit.

Two costs, two shapes: flat rent versus a declining-average-cost asset

Paid placement has almost no fixed cost and a constant marginal cost. Every additional thousand impressions costs the same ~$25 whether it is your first or your millionth; the average cost per exposure is a flat line. That flatness is paid media’s great virtue — it is predictable and instantly available — and its permanent ceiling, because the cost never falls no matter how much you spend.

Earned media is the mirror image. Getting cited carries a high fixed cost — the research, the digital PR and outreach, the specialist time — and then a marginal cost close to zero, because once the citation exists, each additional answer it appears in costs you nothing. The average cost per exposure therefore falls with every exposure the citation earns: steep at first, when the fixed cost is spread over a handful of appearances, then flattening toward zero as the citation accumulates a long tail of firings. This is a classic declining-average-cost asset, the economics of a fixed investment amortised over rising output.

The shape has a consequence worth stating outright: earned media exhibits economies of scale in exposure that paid media structurally cannot. Every extra query the citation fires on, every extra month it survives, spreads the same fixed cost over more output and lowers the average. Paid has no such mechanism — buying more impressions buys them at the same unit price, never a lower one — so it can never enjoy the falling average cost that is earned media’s whole financial case. That is why the channels cannot be ranked in the abstract. They are not two points on one curve; they are two different curves, and which is cheaper is a question about how much output you expect, not about the channels themselves.

Put a flat line and a falling curve on the same axes and they intersect once. To the left of the crossing — low exposure volume — the earned curve sits above the paid line, because the fixed cost has not been spread over enough output to pay for itself; paid is genuinely cheaper there. To the right, the earned curve drops below the paid line and keeps falling, so earned is cheaper and the gap widens without limit. Nothing about this is a matter of taste or brand preference; it is the arithmetic of two cost structures. The only question that matters is which side of the crossing your citation lands on — and that is decided not by the sticker prices but by how many amortising exposures the citation will produce. This is a pricing lens, distinct from the timing question of whether a citation can arrive before your window closes; the two can be run independently and, as we will see, they agree at the boundary.

Even cost-per-exposure is rigged against earned

The obvious way to put both on one denominator is cost per thousand answer-exposures — an effective CPM for each channel. For paid that is just the quoted CPM. For earned it is the fully-loaded earn cost divided by lifetime exposures, in thousands. Compute it and something uncomfortable often appears: on a pure exposure basis, earned media can look more expensive than ads at realistic volumes, because ad inventory is genuinely cheap to print and a single earned campaign is not.

But exposure parity is the wrong test, and treating it as the test is the second great error after the sticker-price one. An impression is not an impression. The ad renders beside the answer, clearly labelled as sponsored, and is skipped by most readers — early observed click-through on ChatGPT’s sponsored slots runs around 0.68%, against roughly 6.66% for a traditional search result. The citation, by contrast, is inside the answer: it is the thing the model is recommending, carrying the engine’s borrowed authority. A user who reaches you through the recommendation arrives pre-sold in a way no adjacent ad impression can match — Criteo’s data has large-language-model referrals converting around 1.5 times better than other channels, and some measurements put AI-referral conversion near 14% against roughly 3% for ordinary organic clicks.

The gap compounds through the funnel rather than appearing once. A sponsored impression that is skipped never enters the funnel at all, so the ad’s low click-through is not a one-off haircut but the first of several multiplicative losses. The citation’s onward engagement starts from the privileged position of being the answer the user asked for, so a far larger share of its exposures become visits, and those visits carry the intent of someone acting on a recommendation rather than someone who tapped a labelled promotion. Multiply those differences down the chain and a single cited exposure can be worth many times a sponsored one — precisely the value a cost-per-exposure comparison discards before the comparison even begins.

So a cost-per-exposure comparison silently assumes the two exposures are worth the same, when the entire point of being cited rather than advertised is that they are not. Pricing earned against paid on CPM is choosing the one denominator that discards earned media’s only structural advantage. The honest denominator has to be cost per outcome — a lead, a signup, a sale — which forces each channel’s conversion rate into the maths instead of assuming it away.

The honest comparison: cost per outcome, not cost per impression

The instrument that makes the comparison fair is an effective cost-per-outcome for each channel — call it the citation eCPA. It is built by walking each channel from its native cost down to a common unit: a qualified outcome.

The paid path is short and every step is observable. Start at the CPM: $25 per thousand impressions is $0.025 per impression. Divide by the click-through rate — take 0.68% — and you get a cost per click of about $3.68, which lands squarely inside the $3–$5 range operators actually report. Divide again by the landing-page conversion rate; at a middle-of-the-range 3% (First Page Sage puts sponsored-answer conversion between 1.1% and 6.0% by vertical) the cost per outcome is roughly $123. Every term in that chain is a number you can look up.

The earned path has the same shape but one uncertain term. The numerator is the fully-loaded cost to earn the citation. The denominator is the lifetime number of outcomes it produces: lifetime answer-exposures, multiplied by the citation’s onward-engagement rate (far higher than an ad’s, because it is the recommendation), multiplied by the conversion rate of that high-intent traffic. Written as a formula, earned cost per outcome equals earn cost ÷ (lifetime exposures × position engagement rate × conversion rate). The paid side has no lifetime term because it has no memory — you rebuild the whole cost every time. The earned side lives or dies on that lifetime term, which is exactly where the risk sits, and exactly what the next section prices.

One more asymmetry cuts against the paid side’s apparent tidiness. The paid chain looks fully observable, but its measured conversions are unreliable in a way the CPM hides: sponsored-answer reporting still carries no third-party verification, lags, and — by the platform’s own accounting — misses a large share of conversions that occur outside the click window. The £123 is therefore a flattering figure with real error bars of its own. The earned side wears its uncertainty on its sleeve, in the lifetime term; the paid side buries comparable uncertainty inside a number that merely looks precise. Neither channel’s cost per outcome is certain — only one of them pretends to be, and the pretence is itself a reason not to over-weight it.

Notice what this does to channel selection. Because the earned numerator is fixed and its denominator grows with the citation’s reach and longevity, earned cost per outcome falls as the citation does more work, while paid cost per outcome is pinned wherever the ad platform’s CPM, CTR and conversion leave it. A sponsored placement and an earned citation are not two prices for the same thing; they are a rental rate and an amortisation schedule, and the schedule wins as soon as the asset is used enough.

The crossover, in one table

Holding the earn cost fixed at an illustrative £18,000 and the paid cost-per-outcome fixed at roughly £123, the only variable left is how many lifetime cited-exposures the citation produces once its uncertain life is discounted. That single number decides everything, and it sorts categories into rent, break-even, and build.

The table is just the two cost shapes made concrete. In a thin, volatile category — a narrow set of queries, a citation that churns out of the index within a couple of model updates — the fixed earn cost never amortises and paid is the rational choice. At the crossover the two are indifferent on price, so the decision falls to other factors (control, speed, or the timing test). In a broad, durable category the earned curve has dropped well below the flat paid line, and in a flagship evergreen it keeps falling until earned is a fraction of the paid cost per outcome. The same £18,000 buys a terrible deal or a spectacular one depending entirely on the denominator — which is why the denominator, not the price tag, is the thing to forecast.

Placing your own category on the table needs only two of your numbers and one of the platform’s. Your earn cost and your best estimate of risk-adjusted lifetime exposures give the earned column; the CPM run through your own funnel gives the paid column. The crossover sits wherever earn cost ÷ (lifetime exposures × engagement × conversion) equals the paid cost per outcome, which rearranges into a single break-even exposure count: earn cost ÷ (paid cost per outcome × engagement × conversion). Any category you expect to clear that many amortising exposures is a build; any that falls short is a rent. It is one division, and it replaces a decision that otherwise runs on instinct and vendor anecdote.

Risk-adjusting the earned side: the citation half-life

The paid cost per outcome is contractually certain — you know the CPM before you spend. The earned cost per outcome is an expectation over an uncertain lifetime, and honest pricing has to discount it for that uncertainty rather than quote the optimistic case. The clean way to do this is a citation half-life: the number of months after which, on current evidence, roughly half of a citation’s answer-exposures have decayed away as models retrain, competitors earn corroboration, and query patterns drift.

Estimating a half-life is not pure guesswork if you instrument it. Sample the share of your target answers in which the citation appears on a fixed cadence, and watch how that presence decays across model refreshes; even a few months of that series yields an empirical half-life far better than a rule of thumb. Categories then sort themselves: a citation anchored to a proprietary dataset or genuine first-hand experience decays slowly and earns a long half-life, while one resting on an easily-replicated claim is out-corroborated fast and earns a short one. Because the half-life is measurable, it is also improvable — building on harder-to-displace inputs is, in eCPA terms, simply a way of lengthening the denominator and lowering the price.

A half-life converts a hopeful “this will run for years” into a decayed sum of exposures that is far smaller than a naive lifetime × monthly-rate multiplication, and it is that decayed sum that belongs in the eCPA denominator. Categories differ sharply here: a citation grounded in a durable, hard-to-displace source — a proprietary dataset, a piece of genuine first-hand experience — has a long half-life; one resting on a cheap, easily-out-earned mention has a short one. The discipline is to estimate the half-life before committing the earn budget, and to treat a short expected half-life as a signal to rent rather than build, since the fixed cost will never amortise.

Two things shorten the half-life in ways teams miss. The first is buying citations on cheap, spammy inputs that later have to be cleaned up — the earn cost then quietly includes a disavow and remediation tax that the sticker figure never showed. The second is downstream leakage: a citation whose value is intercepted before it converts has a lower effective denominator even if the citation itself persists, so the same churn logic that governs the citation’s life also governs whether its exposures actually become outcomes. Pricing and defence are the same balance sheet viewed from two ends.

What actually belongs in the numerator

An eCPA is only as honest as the earn cost on top of the fraction, and this is where most internal numbers quietly cheat — in both directions. Understating the numerator makes earned look free; overstating it makes the whole channel look uncompetitive. Getting it right means loading in the costs that do not appear on an invoice.

The fully-loaded earn cost includes the specialist and analyst time that produced the asset, not just the external spend; the research or data-collection cost behind anything genuinely citable; the outreach and distribution effort, whether that runs through niche editorial placements or a wider press push; and an allocation for the campaigns that earned nothing, because a realistic cost-per-citation has to carry the misses as well as the hits. Against those, two credits belong on the other side of the ledger: the citation almost always produces conventional search and referral value beyond the AI answer, and the same asset frequently earns several citations across related queries, so the fixed cost is shared across every one of them rather than charged to the first. A quick way to sanity-check the numerator is to benchmark it against what comparable assets cost rivals to produce, which a competitor earned-media analysis will approximate.

The campaigns-that-missed allocation is the line teams most often drop, and it can dominate the numerator. If only one in four data-study pitches earns a durable citation, the true earn cost of a citation is roughly four times the cost of the one that landed — the three misses are part of what the hit cost. A channel with a 25% hit rate and a £4,500 per-campaign cost carries an £18,000 real cost-per-citation, which is exactly the figure the worked example below uses. Leaving the misses out makes earned look artificially cheap and sets a budget that cannot survive a normal hit rate; folding them in is what turns the eCPA into a number you can plan against rather than one you will overspend behind.

Market also moves the numerator more than the CPM side, because earning is labour and labour is priced locally. The same citable asset costs materially less to produce through teams in India and South Asia than through a London agency, and outreach economics differ again across international markets. Ad CPMs vary far less by geography than earned-media labour does, which means the crossover point itself shifts by market — the same category can sit in “build” for a team with cheap production and “rent” for one paying premium rates. Pricing earned against paid is therefore partly a question of where you earn, not only what you earn.

A worked example

The company is invented and the figures are illustrative; the arithmetic is real and reproducible. A B2B analytics vendor — call it QuorField — commissions an original benchmark study and a digital-PR push to promote it, at a fully-loaded cost of £18,000. The study earns a citation that fires across about a dozen categorical queries in its space, generating roughly 9,000 cited-answer exposures a month. Its naive cost-per-citation looks like the whole £18,000 — terrifying next to a $25 CPM, and enough to talk a nervous finance team out of ever doing it again.

Now price it properly. QuorField estimates a citation half-life of about eight months for a study-backed source, which discounts the raw lifetime into roughly 70,000 cited-answer exposures before the citation decays to insignificance. Because it is the recommended source rather than an ad beside the answer, its onward-engagement rate is far above a sponsored slot’s — take a conservative 8% against the ad’s 0.68% — and the high-intent traffic converts at around 4%. That is 70,000 × 0.08 × 0.04 ≈ 224 qualified outcomes over the citation’s life. Divide £18,000 by 224 and the earned cost per outcome is about £80.

The paid comparison is the ~£123 per outcome computed earlier. So at QuorField’s volume, earned wins outright — and the one-time £18,000 keeps producing after it is spent, whereas matching those 224 outcomes with ads would cost roughly £27,600 and would have to be paid again next quarter to repeat. The crossover makes the boundary explicit: with an £18,000 earn cost, earned only beats paid once the citation clears about 146 lifetime outcomes, which is roughly 46,000 lifetime exposures. QuorField’s 70,000 clears it comfortably; a thinner category yielding 25,000 would not, and there the honest recommendation is to rent a sponsored placement instead — the same answer the timing test gives for a short window, arrived at independently through price. When the pricing lens and the timing lens agree that a category is a rental, that is as close to certainty as this discipline offers.

The timeline is worth seeing, because it is where the asset logic becomes visible. In month one, with only ~9,000 exposures banked, QuorField’s citation has produced perhaps 29 outcomes and its running cost per outcome sits above £600 — far worse than the ad. By month four the accumulated exposures have dragged the average down through the £123 paid line; by month eight, at the half-life, it reaches the ~£80 quoted above and is still falling. A finance team that judged the campaign at month two would have killed a winning asset for looking like a losing one. That is the single most common way good earned media gets defunded, and it is a direct consequence of reading a stock as if it were a flow — the same error as the sticker price, now wearing a stopwatch.

Where the pricing lens and the timing lens meet

There is a separate rent-or-build test that turns on time rather than money: can an earned citation arrive before the window of value closes? A firm now appears to hold two rent-or-build tests that might disagree. In practice they rarely do, and where they differ the disagreement is informative rather than contradictory. The pricing lens asks whether a citation will amortise — whether it earns enough lifetime outcomes to beat flat rent. The timing lens asks whether it will arrive in time to matter at all.

A thin, fast-churning category fails both at once: too few amortising exposures to justify the fixed cost, and too short a window to earn the citation before the moment passes. A broad evergreen passes both. The genuine tension shows up only in the corners — a durable category with an urgent, closing window, where price says build and time says rent. There the resolution is not to pick one lens over the other but to do both: rent to cover the window now, and build in parallel for the amortising future, because the two lenses are pricing two different things and both bills come due. When money and time return the same verdict, act on it without hesitation; when they diverge, you have found a place that needs a rental and an investment at once, not a choice between them.

The objection this argument has to survive

The strongest challenge is not that the maths is wrong but that its central input is unknowable. It runs: citation lifetimes are so volatile — a model update can evaporate your presence overnight — that any half-life estimate is fiction, and an amortisation built on a fictional denominator is false comfort. On this view the entire “earned is a declining-cost asset” framing is a story we tell to justify spend we cannot actually price; the only number you truly control is the ad CPM, so treat earned as a sunk-cost bet and paid as the one channel you can manage. That objection deserves to be taken at full strength, because its premise is correct: earned lifetimes genuinely cannot be forecast precisely.

But unforecastability is an argument for a discount, not for abandoning the model — and the discount rarely reverses the conclusion except exactly where you would expect. Feed the pessimism straight into the half-life: assume a short one, discount the exposures hard, and recompute. Across most broad categories the earned cost per outcome still comes in under the flat paid line, for the simple reason that paid’s certainty is also its trap — you pay for every exposure, every time, with no accumulation, so a channel you “control” is one whose cost you can never reduce. The uncertainty discount narrows earned’s advantage; it does not usually flip it. And where it does flip it — thin, fast-churning categories with genuinely short half-lives — is precisely the rent zone the crossover already identifies and the timing test already flags. So the objection does not break the framework; it calibrates it, and the honest response is not to stop amortising but to amortise conservatively and let the rent zone be as large as the evidence says it should be. A model that tells you when not to trust it is more useful than a sticker price that never admits doubt at all.

There is a subtler version of the objection that deserves an answer: that paid’s flat cost is not merely a price but a hedge — optionality you can switch on and off as conditions change, which carries real value in a volatile market. That is true, and it is exactly why the framework keeps a rent zone rather than declaring earned universally superior. But optionality is a reason to rent at the margin, not a licence to misprice the core. A firm that lets the value of flexibility talk it out of ever building the durable asset pays the rental rate in perpetuity for the privilege of never committing — the most expensive form of caution there is. Price the asset honestly first; then buy optionality knowingly, in the categories whose volatility actually warrants it, rather than letting a general preference for flexibility quietly veto every investment.

What to do Monday

Pricing earned against paid comes down to refusing the CPM comparison and doing the cost-per-outcome one instead, with the risk written in rather than wished away.

• Never quote a bare cost-per-citation. It is a sticker price with no mileage. Insist on a denominator — outcomes over the citation’s life — before the number means anything.
• Convert both channels to cost per outcome. Walk paid down CPM → CTR → conversion; walk earned down earn cost ÷ (lifetime exposures × engagement × conversion). Comparing on CPM discards earned’s only real edge.
• Estimate a citation half-life first, and be pessimistic. Discount the exposures before they enter the denominator; treat a short half-life as a signal to rent, and route those decisions the way you would route any featured-snippet or owned-surface play where persistence is the whole question.
• Compute your crossover and place each category on it. Below the crossover exposure volume, rent; above it, build; at it, let timing or control break the tie.
• Load the numerator honestly. Include specialist time, the cost of campaigns that missed, and any remediation of cheap links that a low-quality shortcut would later force; credit the non-AI value and the multiple citations one asset earns.

The reassuring conclusion for anyone watching ad budgets balloon is that the citation is not competing with the ad on the ad’s terms. A sponsored slot is rent, priced per exposure and payable forever; an earned citation is a capital asset whose cost per outcome falls the longer it lives and the more queries it serves. On a CPM the asset can look expensive; on the only denominator that pays the bills — cost per outcome, amortised and risk-adjusted — it usually wins, and wins by more every month it survives. Price it that way, and the right link-building strategy stops looking like a cost to be justified against cheaper ads and starts looking like the balance-sheet decision it actually is. The teams that get this wrong are the ones still quoting sticker prices as fares; the ones that get the fundamentals of the discipline right price the mileage, and buy the car. For the underlying benchmarks, keep a live view of the current link and citation statistics, and pressure-test any tool’s cost claims against the actual capabilities of your stack.