Earned vs Paid AI Citations: The 2027 Visibility Budget Split

TL;DR  “Earned vs paid AI citations” sounds like the old organic-vs-PPC split, so most 2026 coverage recommends a ratio — 70/30, 60/40 — and tells you to optimise it. That imports a substitution assumption the architecture forbids. OpenAI’s

“answer independence” runs ads on a separate system from the answer, and across 25M+ cited links only 0.3% of what AI answers cite is paid — against 84% earned. Earned and paid no longer compete for one slot: the citation makes the recommendation and the ad, sitting beneath it, can only harvest the demand the answer created — or fight the demand it conceded to a rival.

So this is not a budget you split. It is a dependency you sequence. Paid is a multiplier on your earned position, positive where you’re already cited and negative where you’re not. This piece gives you two instruments — the Earned/Paid Interaction Matrix and the Visibility Budget Waterfall — to price and order the spend correctly.

The phrase “budget split” is doing your thinking for you

Ads arrived inside the answer on 9 February 2026, when OpenAI switched on Sponsored Recommendations for free and ChatGPT Go users. Within roughly six weeks the format was posting a ~$100M annualised run-rate; Microsoft shipped Sponsored Answers in Copilot; Google kept Gemini’s app ad-free while quietly routing Shopping and Performance Max inventory into its AI surfaces; and Perplexity, having pioneered in-chat ads in 2024, walked away from them entirely. For the first time, showing up in an AI answer had two doors: earn the citation, or pay for the placement.

The industry reached for the nearest analogy and called it a budget split — organic versus PPC, all over again. Nearly every guide published since February proposes a ratio and tells you to test and reallocate it, exactly as you would between SEO and paid search. That framing is comfortable, familiar, and wrong in a way that will cost real money in 2027. It carries a hidden assumption — that earned and paid are substitutes, two rivalrous routes to the same click, so a pound moved from one to the other buys roughly the same visibility by a different means. On a classic search results page that was true. Inside an AI answer it is structurally false, and the whole point of this article is to show why, and what to do instead.

What “answer independence” actually is to a marketer

OpenAI sells answer independence as a promise to the user: the model gives its real response first, and any ad sits underneath as a clearly labelled, visually separated suggestion that does not change the answer. Fidji Simo has been blunt about it — ads will not influence the answers ChatGPT gives you — and the ad selection is contextual, reading the topic of the conversation rather than bidding on a keyword.

Read the same fact from the buy side and it stops being a courtesy and becomes a wall. Answer independence means the recommendation and the advertisement run on separate systems. The one position that mattered — being named in the answer — is the one position you cannot buy. Everything purchasable sits beneath it. And the data confirms the wall is real, not rhetorical: Muck Rack’s May 2026 analysis of more than 25 million cited links across ChatGPT, Claude and Gemini found that 84% of AI citations come from earned media and just 0.3% from paid or advertorial content. A brand’s own site accounts for only 5–10% of what answers reference. The channel you can buy is almost entirely absent from the channel that does the recommending, by design.

This is the same signal the citation-ranking research has been pointing at for a year. Ahrefs’ study of 75,000 brands puts branded web mentions at a 0.664 correlation with AI visibility against 0.218 for raw backlinks; Seer Interactive found brands with third-party mentions appeared in about 75% of AI answers versus 1% for brands without. Citations are won by corroboration, not purchased by placement — which is why the earned side of this ledger behaves nothing like a media buy and everything like digital PR and earned coverage.

The mechanism behind that 0.3% figure is worth stating plainly, because it is the reason the wall will not come down as the ad market matures. Retrieval-augmented generation works by pulling documents at answer time and cross-referencing claims across independent sources before it commits to a recommendation. A sponsored card carries no such corroboration weight — it is metadata attached to the response, not evidence retrieved into it. A brand claim on your own site is, to the retrieval layer, unverified; the same claim reported independently by a journalist or a category analyst becomes citable, because it has cleared exactly the independence test the model is built to apply. That is why a brand’s own website supplies only 5–10% of what answers reference, and why the University of Toronto work found a systematic, overwhelming preference for earned coverage over owned and social content. The recommendation layer is engineered to distrust anything that looks bought — and an advertisement is the most bought thing on the page.

Watch the pricing trajectory and you can see OpenAI itself treating ads as a separate product on a separate track. The format launched managed-only in February 2026 at a $60 CPM with $200,000–$250,000 minimums, sold through the big holding-company agencies; by March, Criteo had joined as an ad-tech partner and observed CPMs had fallen to roughly $25 (as low as $15 in places), with the minimum dropping to $50,000 and a $3–$5 CPC pilot appearing; by May there was no-minimum self-serve. That is a channel maturing along the classic media-buying curve — opening expensive and exclusive, then commoditising toward performance pricing. None of that motion touches the answer. The ad market can get cheaper, broader and more sophisticated indefinitely and still cite nothing, because it was never wired into the recommendation in the first place. Falling CPMs make the Compounding cell cheaper to exploit; they do nothing for the Contested cell except lower the price of the headwind.

You cannot split a budget between two things that don’t trade off

A budget split is a meaningful operation only when the channels are substitutes — when a pound spent on one reduces how much you need of the other, and both compete for the same outcome. On a Google results page they did: a paid click and an organic click both landed the same visit to the same page, so moving budget between them traded off cleanly, and you optimised the mix.

Inside an AI answer the two channels do not compete for the same slot at all. The earned citation determines what the answer says — which brands are named, in what order, with what framing. The paid card determines what sits beneath the answer once it has been said. One shapes the recommendation; the other harvests whatever demand the recommendation released. They are not rivals dividing a slot — they are two layers of the same response, doing different jobs, and a pound moved from earned to paid does not buy the same visibility by another route. It buys a different thing on a different layer, and frequently a worse thing. Spend that treats them as rivals manages to under-earn and over-pay at the same time. The correct object is not the ratio between them; it is the dependency between them.

The inversion: paid used to hedge weak organic — now it multiplies strong organic

Here is the reframe that changes the spend. In classic search, paid was a hedge against weak organic: the worse you ranked, the more you needed the ad, because the ad was how you reached users your organic position could not. The relationship between the two was inverse — paid filled the gap organic left.

Inside the answer the relationship flips to positive and multiplicative. A sponsored card is worth most exactly where you are already cited: the neutral answer names you, the user reads an independent recommendation, and your card beneath it harvests that primed intent with corroboration already done. The same card is worth least — often negative — exactly where a competitor holds the citation, because now the card is arguing with the recommendation directly above it, and the AI referral it competes against converts several times better than a click that arrives contradicted. Paid stopped being a substitute for earned and became a function of it. You do not buy ads to cover the queries you are losing. You buy them to compound the queries you are winning.

The size of the effect is not marginal. AI referrals already convert far above classic organic traffic — industry measurements put AI-sourced visits around 14% conversion against roughly 3% for ordinary organic search, because the user arrives having been recommended rather than having sifted a list themselves. A sponsored card sitting beneath a citation inherits a slice of that primed, pre-sold intent: the answer has done the persuading, and the card catches a user already leaning toward the category and, often, toward you. Strip the citation away and the same card inherits the opposite — a user who has just been steered elsewhere. The card did not change. The ground it landed on did. That is the practical meaning of “multiplier”: the identical creative, at the identical CPM, returns a different amount depending entirely on the earned position above it.

That single inversion reverses the instinct almost every marketer will bring to this channel. The reflex — “we’re not showing up for this query, so let’s buy our way in” — is precisely the move that lands your money on the one cell where a pound buys a headwind. It is the natural thing to do, it feels like initiative, and it is subsidised by every dashboard that reports click-through without reporting what the click was worth. The rest of this piece is a way to see that before you spend, not after — to make the earned position visible as the thing that sets the price of everything beneath it.

Instrument 1 — the Earned/Paid Interaction Matrix

Take any commercial query and locate yourself on two axes: whether you are cited in the answer (yes/no) and whether you are running an ad beneath it (yes/no). The two axes are not independent — that is the whole point — and the four cells have sharply different economics.

The matrix asserts interaction, not allocation. Two of the four cells are traps that look like progress: “Contested” feels like buying visibility but is buying a headwind, and “Organic hold” tempts teams into defending a free position with paid it does not need. Read across the diagonal and the rule falls out on its own — earned position sets the ceiling on what paid can be worth.

The reason this is a matrix and not a slider is that the vertical axis governs the value of the horizontal one. On a slider you assume every position is worth roughly the same and you slide the money toward whichever costs less per click; the two ends are interchangeable. Here they are not. Moving from “no ad” to “ad” changes your return by a completely different amount depending on which row you are in — strongly positive in the cited row, flat-to-negative in the uncited one. A single average “paid ROI” computed across both rows is a statistical fiction: it blends a channel that compounds with a channel that fights itself and reports the meaningless midpoint. Most teams that conclude “ChatGPT ads don’t work for us” have in fact run mostly Contested inventory and averaged it against a thin sliver of Compounding, producing a number that condemns the whole channel for the sin of one cell. Segment by row first, and the channel stops looking broken and starts looking conditional — which is what it is.

Instrument 2 — the Visibility Budget Waterfall

If the channels are not parallel, the question stops being “how do I divide the budget?” and becomes “in what order does each pound become useful?” Because the layers are serial, money spent out of sequence is money wasted: a paid pound with no earned citation beneath it falls straight into the Contested cell. The waterfall makes the sequence explicit and non-negotiable, because each step is the denominator of the next.

1. Step 0 — Citability floor. If the answer engine cannot retrieve, parse and preview you, nothing downstream can fire. This is the accessibility and machine-readable layer the Zyppy meta-analysis ranks highest (URL accessibility 9.5/10, preview control 9.2/10). It is a prerequisite, not a strategy — but skip it and every pound above it is spent on a brand the model can’t quote.
2. Step 1 — Earned citation presence (the denominator). Third-party corroboration that gets you into the answer itself: independent coverage, comparison inclusion, proprietary data others cite. Since 84% of what answers cite is earned and 0.3% is paid, this is the only lever that moves the recommendation. Everything paid multiplies against the position you build here. Work it like earned link and mention building, not like a media buy.
3. Step 2 — Paid harvest. Only now does paid earn its keep — and only on the queries where Step 1 already puts you in the answer (the Compounding cell). Each pound here lands on primed, corroborated intent. This is where sponsorship-style paid placement finally pays, because it is buying amplification of a position you already hold.
4. Step 3 — Paid contest (bounded exception). Time-boxed raids on high-value queries you don’t yet own, each with a written kill-rule and a named owner. Priced as interception, never as a citation substitute. If it becomes a standing line item, you have quietly moved back to treating paid as a hedge — the exact mistake the inversion warns against.

Stated as a rule: you do not divide the money, you sequence it, and the sequence is fixed because each step is the denominator of the next. A team that reports its “earned/paid split” as a KPI is measuring a rivalry that does not exist; a team that reports citation presence and cost-per-cited-query is measuring the thing that actually governs the return. If you want the raw inputs for that denominator, the 2026 citation and AI-visibility statistics are the numbers to build the model on.

Calling Step 1 “the denominator” is not a metaphor reached for effect — it is the actual arithmetic of the channel. If earned citation presence is near zero for a query, then the return on any paid spend against that query is a small number multiplied by roughly nothing, and the product is roughly nothing regardless of how much you pour in. Raise the earned position and you raise the coefficient every downstream pound multiplies against, so the same paid budget suddenly returns because it finally has something to scale. This is why front-loading paid — the thing the budget-split frame actively encourages when it tells you to “get into AI ads early” — is close to the worst possible sequencing. You are pushing the numerator up a curve whose denominator you have not built yet. The counter-intuitive discipline the waterfall enforces is patience with the paid layer: the correct first move in a paid-AI programme is often to spend nothing on paid at all, and route the entire budget through Steps 0 and 1 until there is an earned position worth harvesting.

The measurement asymmetry that quietly breaks the “optimise the split” loop

The split-and-optimise frame assumes you can measure both channels on the same basis and reallocate toward whichever performs. You cannot, because the two sides are not merely different — they are measured on incompatible bases, and one of them is barely a basis at all. OpenAI currently exposes only impressions and clicks for Sponsored Recommendations: no conversion tracking, no attribution model, no ROAS. You can see that a card was shown and clicked; you cannot see whether it moved a single deal. The earned side is measured differently again — by citation presence and share of the answer — which is a stock, not a stream, and lags by weeks.

So the feedback loop the “test and reallocate” model depends on has nowhere to close. You can’t compare cost-per-conversion across two channels when one channel reports no conversions, and you can’t optimise a ratio whose two terms are denominated in different, non-comparable units. This is a large enough problem that it earns its own treatment later in this cluster; for now the takeaway is narrower and sharper: the measurement asymmetry is a second, independent reason the budget-split model fails. It fails on structure (non-substitution) and it fails on measurement (no shared basis), and either one alone is fatal to “just optimise the mix.” If your dashboards currently blend the two into a single efficiency number, treat that number the way you would treat a backlink metric that no longer predicts anything — real, precise, and pointed at the wrong outcome.

There is a subtler trap inside the measurement gap. Because the paid card reports clicks and the earned citation reports nothing directly clickable, a naive dashboard will always make paid look like the channel that “works” — it is the only one producing a tidy, attributable number — while the earned side, which is doing the actual recommending, shows up as unattributed “direct” or “organic” traffic that arrives already convinced. Teams then reallocate toward the channel with the better-looking numbers, which is the channel with the more legible numbers, which is not the same as the channel with the better outcomes. Over a couple of quarters this quietly starves the earned denominator to feed a paid layer that only worked because the denominator existed. The instrument writes the strategy: measure clicks and you will fund clicks, even as the thing producing your pipeline is a citation you never credited. The fix is not a cleverer attribution model bolted onto the ad platform — that model does not exist yet — but a deliberate refusal to let the one legible number in the system dictate the whole allocation.

Where the split behaves differently — by engine and by vertical

The structural argument holds everywhere, but the shape of the paid layer varies enough to change tactics per surface. Four facts worth holding in view:

• ChatGPT (~61% of AI search traffic) is the live paid market: contextual cards, CPMs that fell from a $60 / $200K-minimum managed launch to ~$25 observed (as low as $15 via Criteo) and a $3–$5 CPC pilot. Crucially, the ad-free tiers (Plus, Pro, Business, Enterprise, Edu) skew toward senior, employer-funded buyers — so on high-consideration B2B queries, the audience that sees your ad is systematically thinner on decision-makers than the audience that reads the citation.
• Perplexity exited advertising entirely in February 2026. There is no paid door at all — the only way into a Perplexity answer is organic citation — which makes it the pure case of the thesis and a useful stress test of any strategy that secretly relies on being able to buy presence.
• Google keeps the Gemini app ad-free by policy while routing Shopping and Performance Max inventory into AI Overviews and AI Mode, so “no Gemini ads” and “ads in Google’s AI answers” are both true at once. Microsoft’s Copilot runs Sponsored Answers and claims strong conversion lift, though independent verification is thin.
• Vertical density swings hard: 76.4% of shopping queries return a sponsored placement; finance and insurance hit 85.9% ad density while healthcare sits at 42.3% (and regulated/health/politics queries are largely ad-suppressed). The denser the auction, the more expensive the Contested cell — which is exactly where the split-frame pushes you to spend. Density also varies by market, so
• international and cross-market campaigns should map the paid layer per country before assuming a single global ratio; the earned denominator travels, the paid surface does not.

The strongest objection — and where it actually lands

The best case against everything above: “When a rival owns the citation and I genuinely cannot dislodge them, the sponsored card is my only route to that user. There, paid does substitute for the earned slot I can’t win.” This is real, and it is exactly the Contested cell — but calling it a substitute mis-prices it three ways.

First, it converts against a headwind: the click arrives having just read an answer that recommended someone else, and it competes with an AI referral that converts several times better than a contradicted one. Second, it is the most expensive real estate in the channel — contested commercial verticals carry the highest ad density, so you bid into the densest auction to fight the strongest current. Third, you cannot even confirm it worked, because there is no conversion data to close the loop. So paid-as-interception survives — as a bounded, time-boxed raid with a kill-rule, run while you build the earned position — but it does not survive as a standing substitute for citation. The thesis narrows; it does not break. You can rent your way into the frame. You cannot rent your way into the recommendation, and the rent comes due every month the citation still isn’t yours. Treating a permanent Contested buy as “coverage” is the paid-media equivalent of propping up a position with tactics that carry standing risk — defensible as a raid, ruinous as a residence.

The audience-skew point deserves its own beat, because it quietly amplifies every weakness of the Contested cell in exactly the market where teams are most tempted to buy their way in. Ads in ChatGPT serve to free and ChatGPT Go users; the paid tiers — Plus, Pro, Business, Enterprise and Education — are ad-free. In B2B, the people on employer-funded seats are disproportionately the senior buyers and economic decision-makers, and they are precisely the cohort that will never see your sponsored card. They will, however, see the citation, because the citation is on the answer everyone gets. So on a high-consideration B2B query, the Contested buy is paying a premium CPM to argue with the answer in front of the junior slice of the buying committee while the senior slice reads an uncontested recommendation for your competitor. The substitution isn’t just weak on average — it is weakest against the exact people whose opinion decides the deal. The earned citation is the only asset that reaches the whole committee, and no amount of paid spend backfills the seats the format structurally excludes.

A worked example: the £30k that bought a headwind

A UK B2B SaaS company in the project-management category read the “AI has two lanes, split roughly 70/30” advice and moved about a third of its acquisition budget into ChatGPT Sponsored Recommendations, aimed at its core commercial queries — “best project management software for agencies” and neighbours — where the neutral answer consistently named three incumbents and not them. The card’s click-through looked healthy (the number the platform shows). Pipeline stayed flat (the number nobody could see, because there is no conversion tracking). Every click was arriving pre-contradicted: the user had just read an answer recommending a competitor, then clicked an ad that disagreed with it.

The Interaction Matrix named the problem in one pass: they were buying the Contested cell across the board and mistaking it for “buying visibility.” The fix followed the waterfall, in order. They pulled paid off every query where they weren’t cited. They redirected that spend into the earned denominator — a small set of independent category comparisons and one proprietary benchmark dataset that gave journalists and the models something concrete to quote — the same corroboration logic behind placing citable assets through niche edits and editorial mentions. As those queries began to cite them, they re-introduced paid only there (the Compounding cell), where the card now harvested demand the answer had created. And they kept exactly one Contested raid — a single flagship query, time-boxed, with a written kill-rule and a named owner accountable for calling it.

The sequence was the whole result. The earned position moved first — citation presence on a handful of priority queries within two quarters — and only then did the identical paid pound start returning, because it finally had a denominator to multiply against. They hadn’t been overspending on ads. They’d been spending them one step too early, on the one cell where a pound buys a headwind. Running the programme now needs someone who can hold both layers at once — closer to a modern link and citation specialist than a paid-media buyer, and closer still to whoever owns the foundational question of what earns a mention at all.

What to do Monday — and how you’d know if this is wrong

• Split your top commercial queries by the only axis that matters: are you CITED in the answer, yes or no? That map — not your earned/paid budget ratio — is your real position.
• Overlay current paid. Any spend on a “not cited” query is sitting in the Contested cell. Flag it as a headwind buy, not a visibility buy.
• Move Contested spend two ways: into earned corroboration on those queries, and into paid harvest on the queries you already own.
• Keep at most one or two Contested raids, each time-boxed, each with a written kill-rule and a named owner who answers for it.
• Retire the “split %” KPI. Report citation presence and cost-per-cited-query instead. And if a query starts throwing citation risk — a hallucinated or hostile answer — handle it on the earned layer, the way you’d approach a manual-action or reputation recovery, not by buying a card beneath it.

The thesis is falsifiable, which is the point. If, through 2027, brands with no earned citation on a query reliably matched the conversion of cited brands purely by buying the sponsored card — if paid presence substituted cleanly for citation presence — the non-substitution argument would be wrong and the budget-split crowd would be right. Every current signal points the other way: paid is 0.3% of what answers cite, the ad-serving audience skews away from the ad-free senior buyers who see only the citation, and there is no attribution to even close the loop. If your own data ever shows the substitution holding, trust your data over this article — but instrument for it before you assume it. If you don’t yet have the tooling to measure citation presence at query level, the AI-aware visibility tools are where to start, and the mechanics of holding an answer position connect directly to winning the featured-answer slot in classic search.

This is the opening article of the paid-AI-answer economy, and the rest of the cluster is one structural fact turned over from every side: the recommendation is earned and un-buyable, and everything you can purchase sits beneath it and multiplies against it. From there follow the questions worth their own treatments — when a sponsored placement genuinely beats an earned one and when it can’t; whether answer independence survives contact with advertiser pressure; how to defend organic citation share against sponsored competitors; how to price earned media against AI ad CPMs; how to attribute a sponsored recommendation with no conversion data; what Perplexity’s ad-free bet means for everyone else; and how the whole thing composes into a blended earned-paid-owned portfolio for 2027. Each is an elaboration of the same wall. Get the wall right — sequence the money instead of splitting it — and the rest of the cluster is tactics. Get it wrong, and you will spend 2027 buying headwinds at premium CPMs and calling it a strategy. When you cross into a new market, revisit the sequence from Step 0, because citation dynamics in India, South Asia and other emerging AI-search markets reset the denominator before any paid layer is worth pricing. If a query still refuses to cite you after the earned work is done, that is a signal to audit and clean rather than to buy — the same discipline as a considered disavow and link-cleanup pass, applied to the corroboration graph instead of the backlink profile.

The stakes underneath all of this are larger than one channel’s efficiency. The organisations that internalise the non-substitution rule early will spend 2027 compounding: earning citations that are un-buyable and therefore un-copyable, then harvesting them cheaply as CPMs fall. The organisations that keep the budget-split mental model will spend 2027 renting — buying Contested inventory at premium prices, watching click-through look healthy and pipeline stay flat, and concluding either that AI ads don’t work or that they simply need to bid higher. Both conclusions are wrong, and both are expensive. The gap between the two groups will not read as a difference in ad budgets; it will read as a difference in who gets recommended, which is the only visibility that was ever hard to buy. Answer independence did not remove advertising from AI search. It relocated the one position that matters to a place money cannot reach, and handed a durable advantage to whoever is willing to earn it while everyone else is still trying to purchase it.