The Blended Earned-Paid-Owned AI Presence: A Portfolio Model for 2027

TL;DR

— “Portfolio” is the right word and the wrong finance. Earned, paid and owned AI presence are not three uncorrelated assets you diversify a budget across — they are complementary inputs to one production function, so total output is capped by your weakest factor, not the sum of the three.

— Owned is the gate (machine-readability and entity resolution decide whether you are eligible to appear at all); earned is the compounding core (the only input that clears every engine at once and cannot be counter-bid); paid is a short-duration lever that multiplies whatever the other two already produce and collapses to near-zero on a weak base.

— The instrument is the AI Presence Production Function; the operating rule is the Binding-Constraint Test — the next pound goes to your weakest factor, never the most comfortable one.

— Media planners pour budget into paid precisely because it is buyable, measurable and familiar — which is exactly when it is usually not the binding constraint. That is the most expensive reflex in AI marketing in 2027.

— Stop reporting an earned/paid/owned “split %”. Report cost-per-cited-query per engine and which single factor is binding this quarter.

1. The portfolio metaphor is right — the finance behind it is wrong

Ask a media planner to model earned, paid and owned AI presence as a portfolio and you can watch the mental furniture arrive: three assets, one budget, an allocation across them, and a rebalance each quarter toward some efficient split — 60% earned, 30% owned, 10% paid, adjusted as the numbers move. It feels rigorous because it borrows the vocabulary of Modern Portfolio Theory. It is also the wrong theory for the problem, and the error is not cosmetic. It sends money to the wrong place with confidence.

Modern Portfolio Theory earns its keep in one specific situation: when you hold assets whose returns are imperfectly correlated. You diversify because when one asset falls another tends to rise, so the blended return is smoother than any single holding. Diversification is a variance-reduction machine, and it only works because the assets are, to some degree, independent bets on different states of the world.

Earned, paid and owned AI presence are not independent bets. They are not three ways to win the same coin-flip. They are three inputs that have to combine before any output exists at all — closer to flour, water and heat than to three stocks. You do not diversify across flour and heat to reduce the variance of a loaf; you need enough of each, in order, or you get nothing. Treating the three AI-presence layers as substitutable holdings to spread a budget across — the instinct MPT trains — is the precise mistake this article exists to kill. The correct model is not an allocation across substitutes. It is a production function over complements. The rest of this piece builds that function, tests it, prices it against a real budget, and then tries to break it.

This matters now because 2026 handed operators all three layers as genuinely purchasable, buildable things for the first time. OpenAI put paid placements inside ChatGPT on 9 February 2026; the machine-readable and agentic-eligibility work that constitutes “owned” became concrete through protocols and structured feeds; and the earned tactics that actually move an answer kept compounding as they always have. Three real levers, one budget, and — for most brands — a default instinct to model them exactly wrong.

The cost of the wrong model is specific, not abstract. An allocation mindset treats a shortfall in one layer as something you can offset by over-funding another: earned is weak this quarter, so we lean into paid to compensate. Under a substitutes model that arithmetic is sound — you are trading along a return curve. Under the real, complementary structure it is a category error, because the weak layer is not a shortfall you can subtract around; it is a multiplier sitting near zero that drags the whole product down with it. The brand that “compensates” for weak earned by tripling paid does not buy back the lost citations. It buys a larger rented sign to stand next to a competitor’s recommendation. The money is spent, the dashboard shows impressions, and the answer still names someone else. That failure is invisible to anyone still holding the portfolio in their head as substitutable holdings — which is most of the market, and the opportunity.

2. Define the three layers by role, not by channel

The layers are usually defined by where the work happens: owned = your site, earned = other people’s sites, paid = the ad account. That taxonomy is useless for allocation because it tells you nothing about what each layer does inside an AI answer. Define them instead by their role in producing a citation.

Owned = the gate

Owned is the machine-readable substrate that decides whether you are eligible to be considered at all: whether an engine can parse your content, resolve it to your entity, and pull you into the candidate set it reasons over. This is the technical, machine-readable foundation plus consistent entity signals, plus feed and protocol eligibility (ACP via a payment provider, UCP via your commerce platform, MCP as the data layer an agent reads). It is closer to binary than it looks: below a threshold of readability and entity-resolvability, you are simply not in the set, and nothing spent on the other two layers can retrieve you.

Crucially, owned readability is not the same as classic ranking. Being parseable and resolvable is a separate qualification from ranking in blue links — which is why so much AI visibility accrues to pages that do not rank well, and why how agentic browsers read a site is now its own discipline rather than a footnote to SEO.

Earned = the compounding core

Earned is the third-party corroboration — citations, brand mentions, reviews, listicle inclusion, original data other people quote — that decides which eligible entity actually gets named. It is the only one of the three inputs that clears every engine at once: the same body of earned authority makes you more citable in ChatGPT, Gemini, Claude, Perplexity and AI Overviews simultaneously, because they are all resolving the same underlying reputation. It compounds, it cannot be bought directly, and a rival cannot out-bid you for it. These are the factors that decide which brand an engine names, and they resolve to entity authority the engine can actually resolve.

The evidence that earned is the selection-mover, not a nice-to-have, is now specific. Ahrefs’ study of 75,000 brands found brand-mention frequency correlated with AI-answer visibility at 0.664 versus 0.218 for backlink count — mentions roughly three times as predictive. Muck Rack’s May 2026 “What Is AI Reading?” analysis of 25 million-plus citations found about 84% of AI citations were earned and 0.3% paid. SE Ranking, measuring citation counts against referring domains, found sites with up to 2,500 referring domains averaged 1.6–1.8 ChatGPT citations per category prompt, while sites above 350,000 averaged 8.4. Earned authority is the load-bearing wall.

Paid = the short-duration lever

Paid is amplification you rent. OpenAI’s Sponsored Recommendations (9 February 2026), Microsoft’s Sponsored Answers in Copilot, and the retailer-owned assistant placements that arrived through 2026 all let you multiply exposure you already have — for as long as you pay. The economics are now legible: ChatGPT ads opened at a $60 CPM that settled to roughly $25 observed, with CPCs of about $3–$5, a self-serve Ads Manager since 5 May 2026, and clearly-labelled placement below the answer under OpenAI’s “answer independence” rule. The channel is real — about $100M annualised run-rate within six weeks of launch — and it is a rental. Stop paying and the exposure evaporates the same day.

That rental character is the whole point. Paid does not enter the answer and it does not compound; it sits beside the answer and scales the traffic that the answer’s naming decision already earned you. On a strong earned-plus-owned base it is a genuine multiplier. On a weak base it multiplies almost nothing.

3. The instrument: the AI Presence Production Function

Write presence as a function of the three inputs — Presence = f(Owned, Earned, Paid) — and the modelling choice that matters is whether the inputs are substitutes (more of one compensates for less of another, as in an allocation) or complements (each is required, and output is governed by the scarcest one). The behaviour of AI citations says complements. The function is closer to Presence ≈ min(Owned-gate) × Earned-core × Paid-lever than to a weighted sum — a gate, a base, and a multiplier — which means your output is capped by whichever factor is weakest, not lifted by whichever budget line is largest.

The distinction is not academic hair-splitting between two curves; it changes what the model tells you to do with the next pound. A weighted-sum model — Presence = 0.5·Earned + 0.3·Owned + 0.2·Paid — has a fatal property: because every term adds independently, you can always raise the total by topping up whichever factor is easiest to buy. It will happily recommend more paid on top of an already-strong paid position, because another unit of paid always adds its weighted contribution. A complements model refuses that move. If owned is the minimum, additional paid multiplies by a gate that is still near zero and the total does not budge. The two models agree only when all three factors are already balanced; the moment one lags — which is the normal state of any real brand — they give opposite instructions. One says “top up the easy factor”; the other says “fix the scarce one”. Only the second is right, because only the second matches how citations actually behave.

Make the intuition concrete. Picture presence as water through three valves in series: owned, then earned, then paid. Flow is set by the most-closed valve, not the average opening. Open the paid valve fully while the owned valve is nearly shut and almost nothing flows — you have maximised the wrong valve. This is why “how much are we spending in total” is the wrong question and “which valve is most closed” is the right one. Optimise the tightest point and the same budget produces multiples of the output; optimise the total and you can pour money in indefinitely while the bottleneck holds the flow flat.

The table names each layer by its role in that function, what it gates or multiplies, how it behaves when the other two are absent, whether it can be bought, and how it behaves over time.

Read down the “Buyable?” column and the trap announces itself: the only fully-buyable, instantly-measurable layer is the lever, not the gate or the core. A budget process optimises what it can measure and purchase, so it drifts toward paid — the layer least able to lift a weak base.

4. Why the factors multiply instead of add

The claim that the three are complements rather than substitutes is not a metaphor to accept on faith. Walk the pairs and the multiplicative structure falls out mechanically.

Paid without owned. You buy a Sponsored Recommendation, but your content is thin, your entity is ambiguous, and the engine cannot resolve “you” cleanly. The ad renders; the organic answer beside it names a competitor the model can actually corroborate. You have paid to place a rented sign next to someone else’s citation. Paid × (owned ≈ 0) ≈ 0.

Paid without earned. Owned is fine — you are eligible, parseable, resolvable — but no one corroborates you, so the organic answer still names the incumbent. Your sponsored slot drives a trickle of clicks at a CTR that trade measurement puts near 0.68% for ChatGPT ads, against roughly 6.66% for traditional search. You are buying the low-trust seat beside the high-trust recommendation. The decisive touch — the model naming a brand in the answer — is the one you did not buy.

Earned without owned. You have genuine third-party authority, but your machine-readable substrate is broken: inconsistent entity signals, unparseable pages, no feed eligibility. The engine reads the corroboration and cannot reliably attach it to you; the citation leaks to a better-resolved neighbour. Authority you cannot make legible is authority you do not get credited for.

The revealing pairing is the fourth one — owned and earned present, paid absent — because it is the only combination that does not collapse. Eligible, well-corroborated, and running no ads, you still get named in the organic answer; you simply capture less of the downstream traffic than you could if you also rented amplification on the highest-intent prompts. That is not a failure state; it is the baseline the other two layers exist to protect and extend. It is also the tell that paid is a lever and not a leg: remove paid and presence degrades gracefully to “cited but under-amplified”, whereas remove owned or earned and presence collapses to “absent”. A factor whose removal degrades output is a multiplier; a factor whose removal zeroes output is a gate or a core. Paid is the only one of the three you can drop without disappearing — which is precisely why it should be the last one funded, not the first.

Each of the collapsing pairings goes toward zero because one factor is missing. That is the formal content of “non-substitutable”: you cannot buy your way out of a weak factor by over-funding a strong one, because the weak factor is a multiplier near zero, not a subtractable shortfall. This also explains a fact that confuses allocation-minded planners — that only a minority of AI-Overview-cited pages, on the order of the 17–38% that also rank in the classic top 10, clear both bars. Ranking and citation are different qualifications, produced by different factors, and having one does not supply the other. The lesson generalises: presence in an AI answer is a conjunction of conditions, and a conjunction fails at its weakest term.

5. The operating rule: the Binding-Constraint Test

If output is capped by the weakest factor, then the only spend that raises output is spend on the binding factor — the current minimum. Money poured into a non-binding factor does not move presence; it buys a taller wall on the side of the house that was already tall enough. So before any budget line is approved, run three checks in order and find where the constraint actually sits.

The three checks (run in order)

1. Owned / eligibility. Take your ten highest-value category prompts. On each engine, is your entity even parseable, resolvable and in the candidate set? If you are absent from the set on two of the major engines, owned is binding — no earned or paid spend returns anything until the gate clears. Diagnose readability, entity consolidation and feed eligibility first. Tooling to check this is now standard in the link-building and GEO tool stack.

2. Earned / corroboration. If you are eligible but the answer still names a rival, earned is binding. The engine can see you and chooses someone better corroborated. The fix is the un-buyable core: original data others quote, third-party citations, reviews, the reputational work behind what link building actually is. Paid spend here amplifies an absence.

3. Paid / window. Only if owned and earned both clear: is there a specific, time-boxed window — a launch, a category whose earned incumbents are weak, a surface where you own selection — where renting amplification beats waiting for earned to compound? If yes, paid is the lever to pull, sized to the window. If no, paid is discretionary and should lose the budget contest to the binding factor.

The checks are cheap to run and worth running literally rather than by intuition. Build a fixed set of ten to twenty category prompts a real buyer would type — “best expense-management software for UK SMEs”, “how do I automate mileage claims” — and put each to ChatGPT, Gemini, Claude, Perplexity and AI Overviews on a schedule. Record three things per prompt per engine: are you present in the answer at all (eligibility), are you named or merely a rival (corroboration), and is any presence organic or sponsored (window). That grid is the whole diagnosis. A brand absent from most cells has an owned or earned problem no ad budget will touch; a brand present-but-unnamed has an earned problem; a brand named organically across engines has earned the right to consider paid on the margins. The grid also travels — the same prompt set, translated, exposes where the binding constraint differs by market, which is why per-market diagnosis beats a single global budget.

The rule that falls out is uncomfortable because it contradicts the reflex: the next pound goes to your weakest factor, never your most comfortable one. The comfortable factor is almost always paid — it is the one with an account, a dashboard, a rep on the phone and a number that moves this week. Planners fund it precisely when it is non-binding, because it is legible, and legibility feels like control. A brand that fails the candidate-set check for “expense-management software” on two engines and responds by raising its ChatGPT ad budget has bought a bigger multiplier on a gate that is still shut.

6. Worked example: Cranmore, £360k, 2027

Cranmore is a UK B2B expense-management SaaS, roughly £12M ARR, with a £360,000 AI-presence budget for 2027 and a board that has read that finance and insurance queries carry the heaviest sponsored-ad density of any vertical — about 85.9% per OtterlyAI, against 42.3% in healthcare. The instinct in the room is obvious: our category is where the ads are, so buy the ads.

The Binding-Constraint Test says otherwise. Running Cranmore’s ten priority prompts across the major engines, the diagnosis at month 0 is that Cranmore is absent from the candidate set on two of three engines: inconsistent entity signals across its site and third-party profiles, no llms-oriented fact pages, no feed or protocol eligibility. Owned is binding. Earned is thin but not the constraint yet, because until the gate clears, earned cannot express itself. Paid is non-binding by definition — a multiplier on an entity the engine cannot yet resolve.

Two plans, same £360,000:

Month 1 flatters the naïve plan. Its sponsored placements are live, a handful of sponsored citations appear, and the dashboard shows motion. The constraint-driven plan is quieter — it is fixing a gate and seeding data that has not yet been quoted. If Cranmore’s board judges on month-1 optics, the naïve plan wins the room, which is exactly how the expensive reflex survives contact with reality.

Month 3 is where the diagnosis starts to show through the optics. The constraint-driven plan has cleared the gate on all three engines; Cranmore’s entity now resolves cleanly and its UK expense-benchmark dataset — a single commissioned survey turned into a public, dated, quotable number — has picked up its first dozen third-party citations, each one un-bought and each one clearing every engine at once. The naïve plan’s citation count is higher on paper, but almost all of it is sponsored, which means it is renting: pause the spend for a fortnight to test durability and the count falls back toward its organic floor, which is still near zero on two engines. The two plans now diverge not in how much presence they show but in what kind — one owns an appreciating asset, the other is servicing a lease.

Month 6 inverts it. The naïve plan is still absent from the organic candidate set on two engines — it never cleared the gate — so its paid citations vanish at every spend pause and its cost-per-cited-query stays high and non-compounding; it has spent £290,000 renting multipliers on an entity the engine still cannot resolve. The constraint-driven plan has cleared the gate, its expense-benchmark data is being cited across engines with no spend attached, and paid is being tapered because earned is now doing the naming. The two plans did not differ in effort or budget. They differed in which factor they fed first. One fed the binding constraint; the other fed the comfortable one. Original data of the kind Cranmore commissioned is doubly efficient here because it earns editorial links and AI citations from the same asset — the reason an interactive tool or calculator that earns links at scale and expert-sourcing platforms like the HARO successors Connectively, Featured and Qwoted sit in the earned column, not the paid one.

7. Rebalancing done right: toward the bottleneck, not toward a split

Here is where the portfolio metaphor pays back the reader who followed it in — by inverting. A financial portfolio rebalances toward a target allocation: drift from 60/40 and you sell the winner to buy the laggard, restoring the split. An AI-presence portfolio does the opposite. You rebalance toward the migrating binding constraint, and the correct allocation is never fixed because the bottleneck moves as you clear it.

Cranmore shows the migration. At month 0 the constraint was owned, so owned got funded out of proportion to any “split”. Once the gate cleared, the constraint moved to earned — and earned became the right place for the next pound, not because a target said 50% but because it was now the minimum factor. As earned compounds, the velocity at which earned authority accumulates shrinks paid’s window of value: the more the organic answer already names you, the less a rented slot beside it adds. Paid’s correct budget falls as earned rises, which is the exact opposite of what a fixed 10% paid line would hold it at.

The migration has a typical shape worth naming, because most brands travel it in the same order. Owned binds first, because eligibility is a precondition and most brands start ineligible on at least one engine. Once the gate clears, earned binds for the longest stretch, because corroboration compounds slowly and cannot be rushed with money — this is the phase that rewards patience and punishes the instinct to declare the owned fix “done” and pour the freed budget into ads. Only when earned is genuinely competitive does paid become the binding lever, and even then only inside specific windows. A brand that funds in that order — gate, then core, then lever — spends every pound on the factor that was actually holding it back at the time. A brand that funds in reverse, chasing the measurable lever first, spends its largest early cheques on multipliers with nothing yet to multiply.

This is why the “earned/paid/owned split %” is the wrong KPI, and why it deserves to be retired outright. A split is a snapshot of an allocation that should never be stable. Report two things instead: cost-per-cited-query per engine (is presence getting cheaper as earned compounds?) and which single factor is binding this quarter (where does the next pound go?). Those two numbers tell you what a split never can — whether you are feeding the bottleneck or decorating the tall wall.

8. Where this breaks: the greenfield substitution window

The strongest objection to a strict complements model is not that it is wrong but that it is too rigid, and there is a real case to answer. In a brand-new category with no earned incumbents — nobody has authority yet, because the category barely exists — paid genuinely does substitute for earned in the short run. If the organic answer has no well-corroborated brand to name, a Sponsored Recommendation can be the most visible option in the response, and it can move real demand while earned is still zero for everyone. And owned is a gradient, not a clean gate: partial machine-readability yields partial eligibility, not a binary in/out. Over some range, then, the factors are substitutes, and the production function overstates the rigidity. That objection is correct on its own terms, and it deserves to be conceded fully rather than waved away.

It is conceded — and then bounded, because the bounds are where the model earns its keep. Three limits hold.

First, the substitution window is time-boxed and closes on contact with any rival’s earned authority. The moment one competitor’s corroboration compounds, the organic answer has a brand to name, and paid drops back to a multiplier beside it. Greenfield substitution is a transient state, not a strategy — it is a bounded rental window that any maturing category eventually closes. Timely, newsworthy earned coverage — the kind of newsjacking that wins citations fast — is often how the first mover slams that window shut on everyone still renting.

Second, the substitution is one-way and decaying. Paid can stand in for absent earned for a while; earned never needs paid to be cited. Muck Rack’s ~84%-earned / 0.3%-paid split is the asymmetry made numerical. A model where A can temporarily cover for B but B never needs A, and A’s coverage decays, is a complements model with a transient — not a substitutes model.

That asymmetry has a pricing consequence the greenfield objection actually sharpens rather than softens. If paid substitution only works while a category has no earned incumbent, then the value of renting is highest at the exact moment it is most perishable — and the correct response is not “spend more because it works now” but “spend fast and convert hard”, turning rented attention into earned corroboration before a competitor does. Read that way, the strongest objection to the complements model becomes an instruction inside it: rent aggressively in true greenfield, but only to build the earned core that will make the rental unnecessary. Rent the window and never convert it, and you have chosen a slower, costlier route to the same weak base.

Third, even in the purest greenfield, paid still requires owned. You cannot enter the sponsored candidate set — or the feed a retailer-owned assistant reads — without the machine-readable substrate. The gate survives the objection untouched; only the earned/paid boundary softens, and only briefly. So the production function holds in the limit that matters: any category that matures, which is all of them. The exception is real, bounded, decaying, and already accounted for. Bounding the claim this way makes it stronger, not weaker — it tells you exactly when to break your own rule and rent, and for precisely how long. Even in technical, developer-heavy greenfields, the durable move is to convert the rented attention into earned standing fast — turning a launch on a technical community like Hacker News into corroboration that outlives the spend.

9. The 2027 operating model

Collapse the whole argument into a standing procedure a team can run every quarter.

Run the Binding-Constraint Test before any budget line. Never approve “increase paid” or “increase owned” in the abstract; first locate the minimum factor across your priority prompts and engines, and let that decide where the next pound goes.

Fund owned to threshold first. It is a gate, it is usually the cheapest factor to clear, and clearing it unblocks the return on everything else. Below threshold, every other pound is wasted.

Fund earned as the compounding core. It is the only spend that clears every engine at once, it cannot be counter-bid by a richer rival, and it is the input that actually moves the naming decision. In a multi-engine world where ChatGPT drove about 61% of AI-search referrals against Gemini’s 24.8% in Q1 2026, earned is the single unit that pays off across all of them at once — and it compounds per market, which is why earned authority built for European markets is banked, not rented.

Treat paid as a sized lever on live windows. Rent amplification when there is a specific window and a base worth amplifying; taper it as earned compounds. Never let it substitute for a weak factor, and never let its legibility win it budget it has not earned on the merits.

Measure cost-per-cited-query and the binding factor, not a split. Keep the numbers current against the wider 2026 link-building and AI-citation statistics, and let the binding factor — not a target allocation — set the next move.

The brands that win the AI answer in 2027 will not be the ones that spend the most across three buckets. Diversifying a budget across earned, paid and owned as if they were substitutable holdings is the confident, well-modelled way to underperform. The winners will be the ones who understand that presence is a product of complements, who find their binding constraint honestly, and who never — however comfortable it feels — spend the next pound into a factor that was not the one holding them back.