The Coming Multi-Engine Ad Auction: A 2027 AI Media Plan

TL;DR

•  “The multi-engine ad auction” smuggles in a false assumption — that the engines will converge into one marketplace you bid across from a single console. The evidence points the other way: toward divergence into incompatible auctions.

•  Google extends its keyword-and-Performance-Max auction into AI Overviews and AI Mode; ChatGPT runs a contextual auction with no keywords, demographics or remarketing; Copilot prices through Microsoft Advertising; Perplexity and Claude run no auction at all.

•  They will not merge, because each auction’s primitive is downstream of the platform’s own data endowment, which is not shareable. No common currency, no common targeting, no common measurement.

•  The Auction-Primitive Map reads each engine’s bidding primitive and shows where your intent is under- or over-priced. Fragmentation is the opportunity: when capital cannot flow freely between incompatible auctions, mispricing persists — and mispricing is where advantage lives.

•  The Cross-Engine Arbitrage Test turns that into a posture per engine: concentrate, maintain, or earn-only. The 2027 plan hunts the underpriced engine; it does not spray budget evenly across all of them.

•  The one currency common to every engine is the earned citation — un-buyable on the ad engines, the only route in on the ad-free ones. The paid layer fragments; the earned layer is the only unified auction, and you bid there with content, not cash.

The phrase hides a false assumption

“The coming multi-engine ad auction” is a comfortable phrase, and comfort is usually where a bad assumption is hiding. When most planners say it, they picture convergence: a near future in which ChatGPT, Google’s AI surfaces, Copilot and the rest settle into a single marketplace you buy across from one console, the way programmatic display consolidated a thousand ad networks into a handful of demand-side platforms. One bid, many engines, a tidy cross-channel dashboard. That is the mental model behind almost every “prepare for the multi-assistant future” post published this year.

Here is the counterparty claim, stated at full strength so we can test it rather than caricature it: “By 2027 the AI ad surfaces will behave like one auction — intermediaries and holding companies will knit them together, you will plan across them from a single workflow, and budget will flow to whichever engine returns the most, just as it does across today’s programmatic estate.” It is a reasonable extrapolation from how every prior ad channel matured. And it is very likely wrong, because the AI engines are not fragments of one market waiting to be consolidated. They are structurally different auctions that do not share a bidding language — and, in two cases, run no auction at all.

This article builds the media plan for the world that is actually arriving: not one auction, but several incompatible ones plus an ad-free remainder. It is a fair reading — the convergence case has real, named evidence behind it, and we will give that its full due before the end. But the plan that wins in 2027 is not “bid across all engines.” It is “find the engine where your intent is underpriced, concentrate there, and fund the one asset that pays across all of them.” For the foundations of how paid and earned interact inside an answer, our work on AI Overviews and backlinks sets the ground; here we are one level up, at the level of the media plan.

Five engines, five auctions — or none

Lay the surfaces out by how you actually buy them, and the incompatibility is immediate. This is not five doors into one room; it is five rooms with different locks, and two with no door for sale.

Google folds its AI placements into the machinery advertisers already run. Ads reach AI Overviews and AI Mode through Performance Max, Shopping, broad-match Search and AI Max for Search, per Search Influence’s mid-2026 breakdown — near-zero setup, because the campaigns already exist and smart bidding (Target CPA, Target ROAS) reprices in milliseconds when an Overview reshapes available impressions. The bidding primitive is the mature Google one: keywords, audiences, first-party signals, PMax goals.

ChatGPT runs something structurally different. It does not sell keywords, demographics or remarketing lists; you supply context hints and short descriptions, and the system matches contextually, per Lapis’s 2026 assistant guide. Since OpenAI opened its self-serve Ads Manager to all US businesses in May 2026, you can set a CPM or CPC bid — but the thing that bid buys against is conversational context, not a keyword or an audience list. As one advertiser guide put it, ChatGPT sits between Google and a native channel, capturing more intent than social (users describe their problem in words) and more consideration depth than search (conversations span turns), yet it is two decades behind Google on auction transparency, attribution, audience controls and fraud infrastructure (Greg Hal, May 2026).

Copilot is a third primitive again: ads serve through Microsoft Advertising, the only mature LLM ad channel before ChatGPT, priced on Microsoft’s Bing-and-LinkedIn auction and now extended via AI Max for Search, with Microsoft reporting +69% CTR and +76% conversion on lower-funnel Copilot ad types versus traditional Bing (Greg Hal). And then the remainder: Perplexity runs no ad auction at all — it exited advertising in February 2026 for a subscription model — and Claude is deliberately ad-free, a principle Anthropic underlined with Super Bowl advertising criticising in-assistant ads (Lapis). Meta AI is still “mulling” ad formats. So the “multi-engine auction” is, on inspection, three incompatible auctions and two engines with no auction to enter. Any plan that treats them as one surface is planning for a market that does not exist.

And the asymmetry of scale sharpens the point rather than softening it. ChatGPT alone reaches over 900 million weekly users, most on ad-eligible free tiers, so its opaque contextual auction sits in front of an audience larger than most established channels — mispriced inventory at enormous scale. Yet on that same surface, as on Google’s and Copilot’s, the ad is architecturally separated from what the model recommends: you can bid for the box beside the answer, never for the citation inside it. That separation is not a footnote to the auction landscape; it is the reason the paid and earned layers behave so differently across every engine, and why a media plan has to account for both rather than collapsing them into a single “AI spend” line.

Why they will not converge into one auction

It is tempting to read this incompatibility as mere immaturity — a rough early phase that standardisation will smooth into a single marketplace. That reading underestimates how deep the divergence goes, because each auction’s primitive is not an arbitrary design choice. It is downstream of the platform’s own data endowment, and data endowments do not merge.

Google can auction keywords and audiences because it owns two decades of query logs, a first-party identity graph and the PMax model trained on them. ChatGPT cannot auction those things — it has no comparable ad-targeting history — so it auctions the one asset it does own: live conversational context. Microsoft prices on Bing search intent fused with LinkedIn’s professional graph, an endowment neither of the others has. Each platform’s bidding primitive is the monetisable shape of the data it happens to hold, and because no platform will hand its proprietary data to a shared exchange, the primitives cannot be made fungible. A keyword bid and a context-hint bid are not two dialects of one language; they price different substrates. There is no exchange rate between them.

That is a durable reason, not a transitional one. Programmatic display converged because the underlying unit — a cookie-identified impression on an open web page — was common across networks, so an exchange could clear it. The AI engines share no such unit; the impression on ChatGPT is a moment inside a private conversation the platform will not externalise, and the impression on Google is a placement inside Google’s own answer. Absent a common unit, there is nothing for a unified auction to clear. The engines can be bought in parallel — more on the intermediaries who enable that later — but parallel buying across incompatible auctions is not the same thing as one auction, and the difference is the whole game.

The window matters because the mispricing is temporary in magnitude even where it is durable in structure. OpenAI’s own investor projections, scooped by Axios in April 2026, run $2.5B in ad revenue for 2026 to $11B in 2027, $25B in 2028 and $100B by 2030 (Truist’s analyst note is more conservative, pegging 2026 under $1B; the channel was running near $109M a month mid-year). A surface scaling that fast will not stay thinly-contested forever — demand is arriving to meet the inventory. The durable fact is that the auctions stay incompatible; the perishable fact is that any given engine’s underpricing closes as advertisers pile in. That combination is precisely why a 2027 plan is an active hunt rather than a fixed allocation: the incompatibility that lets mispricing exist is permanent, but each specific mispriced pocket has a shelf life. The same pressure is reshaping how people reach these answers in the first place — the shift toward agentic AI browsers moves more discovery inside the assistant, which is what makes each engine’s auction, and each engine’s citation, worth contesting now rather than later.

The Auction-Primitive Map

The map reads each engine on four dimensions and then draws the conclusion the first three force. What you actually bid — the currency you set. The targeting substrate that bid buys against. How well you can measure the result. And, off those three, how fully priced the surface is — which is where the fifth column lives: the location of any edge. A mature, liquid, well-measured auction is fully priced and yields no edge; a thin, opaque, or under-demanded one holds mispricing; a surface with no auction cannot be bought at all, only earned.

Read across each row before comparing down the columns — the point is that the columns never line up, which is exactly why no single bid can span them.

Nothing in the middle three columns is shared. A bid means a different thing in every row; the substrate it buys is different; the confidence you can have in the result is different. That is the visual proof that the “one console, one bid” future is not arriving on these foundations — and it hands you the more useful question in its place: given that the surfaces are priced differently, which one is currently underpricing the intent you happen to sell?

Fragmentation is where the money is

The instinct to mourn fragmentation is backwards. In a unified, liquid auction — the programmatic ideal — competition equalises returns: capital floods any surface offering above-market return until the return is bid back down to par. That is efficient, and efficiency is precisely what leaves no edge for anyone. Fragmentation into incompatible auctions does the opposite: because capital cannot flow freely between surfaces — different creative, different targeting substrate, different measurement, often different teams — returns can stay persistently unequal. An engine can underprice your intent for quarters at a time, because the friction that would arbitrage the gap away is exactly the friction the incompatibility creates.

So the fragmentation everyone treats as a headache is the source of whatever edge exists. The 2027 media plan is a mispricing hunt, not a coverage exercise. Two structural pockets are visible right now. ChatGPT underprices consideration-depth intent because its auction is opaque and young and most advertisers cannot yet measure it, so demand lags the inventory’s real worth — the classic profile of a surface that is cheap because it is hard, not because it is bad. Copilot underprices business and enterprise intent because advertiser demand has not caught up to a channel Microsoft has quietly run since 2023, which is what a claimed +76% lower-funnel conversion advantage against Bing implies. Google’s AI surfaces, by contrast, are the mature, liquid, fully-priced case — you should expect to pay fair value there and find no edge, only the cost of being present. The map is not telling you to spread evenly; it is telling you where the same pound buys more.

This is also why porting one engine’s playbook to the next destroys the edge you came for. The reason ChatGPT is underpriced is that its context-driven auction resists the keyword-and-audience habits that make Google efficient; run it like Google and you both misfire the creative and compete away the very inefficiency you were exploiting. Each pocket stays profitable only as long as it is worked on its own terms.

There is direct evidence the mispricing is real rather than theoretical. Criteo, reporting on its ChatGPT integration through 2026, found that traffic referred from LLM platforms converts at close to twice the rate of traditional search in several retail categories — up from the roughly 1.5× it first measured — and characterised the budgets flowing in as additive rather than reallocated from other channels. Additive budgets on a surface converting at up to 2× search is the fingerprint of intent selling below its worth: advertisers are finding new return, not merely shuffling spend. That is the arbitrage, visible in the data. But it also carries a warning — the pace at which you press a mispriced pocket matters, because moving too slowly cedes it to the competitors arriving weekly, and moving too fast on an unmeasured surface scales a bet. The discipline of pacing your earning velocity against what you can actually verify applies to paid arbitrage just as much as to earned links: press hard where you can measure, test-and-hold where you cannot.

The Cross-Engine Arbitrage Test

A mispricing hunt needs a decision rule, or it becomes a hunch. The Cross-Engine Arbitrage Test asks four questions of each engine and resolves to one of three postures. The questions run in order; a failure early stops you before you spend.

• Is the auction mispriced? Thin advertiser demand, opacity, or a channel most buyers have not reached yet all signal that intent may be selling below its worth. A mature, liquid, heavily-contested auction is fairly priced by definition — no edge to hunt.
• Does its primitive fit your intent? A context-only auction rewards considered, describable, multi-turn purchases and punishes narrow navigational ones; a keyword auction is the reverse. Underpriced intent you cannot actually target through the engine’s primitive is not an opportunity — it is a mismatch.
• Can you measure it well enough to know you are winning? An underpriced surface you cannot measure is a bet, not an arbitrage. Where deterministic tracking fails — which is most of the AI surfaces — you need an incrementality read before you scale, or you will mistake a mispriced win for a lucky one.
• Is there an auction at all? Two of the five engines have none. On those, no bid exists to place; the only entry is the earned citation, and paid budget aimed there is aimed at a door that is not for sale.

The postures: CONCENTRATE where the auction is mispriced, its primitive fits your intent, and you can measure it — this is where disproportionate budget earns disproportionate return. MAINTAIN where the auction is fairly priced but being present is table stakes — pay fair value, do not over-invest chasing an edge that competition has already closed. EARN-ONLY where there is no auction, or where its primitive fights your intent — skip paid entirely and win the surface through the citation. Measurement across all three leans on lift rather than last-click, for reasons our entity-authority measurement work sets out; deterministic attribution is the one thing none of these auctions reliably gives you.

The only unified auction is the earned one

Step back from the paid surfaces and one currency is common to every engine on the map — the one column the incompatibility does not touch. On the ad-funded engines you cannot buy your way into the answer: Google, ChatGPT and Copilot all separate ads from what the model actually recommends, so the citation inside the answer is earned, never auctioned. On the ad-free engines the citation is the only route in at all. Which means the earned citation is the single asset that pays across all five engines regardless of their auctions — and it is priced the same way everywhere: on merit, not on a bid.

That reframes the whole media plan. The paid layer is fragmented into incompatible auctions you must work one at a time; the earned layer is the only unified auction there is, and you bid in it with content and corroboration rather than cash. One investment — becoming the answer’s cited source — clears on Google’s surface, ChatGPT’s, Copilot’s, Perplexity’s and Claude’s simultaneously, because all five draw the substance of their answers from the same earned material. Our companion analysis of how business models shape each platform’s incentives showed why the earned citation survives however the ad question resolves; the point here is sharper still: it is the only line item on the plan that spans every engine at once. The mechanics of winning it — the digital-PR and newsjacking that earns the coverage, the technical groundwork that makes your pages retrievable to every engine’s crawler — are a single spend against five surfaces.

This is the resolution to the fragmentation problem, not a way around it. You do not defeat five incompatible auctions by mastering all five; you fund the one auction that spans them first, then treat the paid engines as engine-specific supplements bought where they are underpriced. The earned citation is the floor the whole plan stands on; paid is what you add on top, selectively, where the competitive citation and link picture shows an underpriced pocket worth pressing.

The scale of the earned layer is not a rounding error against paid — it is the main event, which is what makes funding it first a matter of arithmetic rather than principle. Muck Rack’s May 2026 analysis of more than 25 million citations across ChatGPT, Claude and Gemini found that 84% of AI citations are earned and just 0.3% are paid or advertorial. The engines assemble their answers overwhelmingly from earned material; the ad auctions decorate the margins around those answers. A plan that pours budget into the 0.3% while starving the 84% has inverted the map of where AI visibility actually comes from. The paid auctions are real and worth working where they are underpriced — but they are the supplement, and the earned citation is the base.

What actually goes in the 2027 plan

Assemble the pieces into the plan itself, and it inverts the usual template. The usual template leads with a paid budget split across the ad engines and treats earned as a nice-to-have. The plan the map produces leads with the earned citation and treats paid as targeted arbitrage.

The earned citation is funded first, because it is the cross-engine constant — every pound of it works on all five surfaces, and it is the only spend that lowers what you must pay the paid auctions for presence (you do not bid for a slot you already occupy on merit). Paid budget then concentrates on the underpriced engines the arbitrage test flags — not spread evenly, but weighted hard toward the mispriced pocket where your intent fits the primitive. Google’s AI surfaces get a maintain allocation: present, competitive, not over-fed. Perplexity and Claude get no paid line at all — only earned effort. And crucially, creative is built per primitive: the context descriptions ChatGPT wants are not the keyword-and-headline assets Google wants, and porting one to the other is how teams both waste production and forfeit the edge. Measurement is per-engine and mostly incremental, because none of these auctions hands you trustworthy deterministic attribution.

What the earned line actually funds is worth making concrete, because “earn the citation” is easy to nod at and hard to resource. It is the commissioned research an engine will quote, the expert commentary sourced through journalist platforms like Connectively and its successors, and the interactive tools and calculators that earn links and get named as the reference answer at once. These are not five separate production efforts for five engines; they are one body of citable material that every engine’s retrieval draws on. That is the leverage: the paid auctions demand bespoke creative per primitive and give you nothing back when you stop paying, while the earned assets are built once and clear on every surface for as long as they remain the best available source. Underfunding this line to buy more auction inventory is the most common and most expensive error in an AI media plan — it pays rent on five surfaces to avoid buying the one asset that would lower the rent on all of them.

The plan, in one line: earn the citation everywhere, buy the underpriced auction hard, hold the priced auction steady, skip the auctions that do not exist, and never run one engine’s creative on another. The engines that reward this earliest are the international and vertical ones where advertiser demand is thinnest — the European-market surfaces in particular, where Criteo only extended ChatGPT access mid-2026 and competition has barely arrived.

A media plan, worked

Make it concrete. Verrine is an invented but specific case: a UK project-management SaaS, considered B2B purchase, with a 2027 AI media budget of £600,000 and a finance director who wants the allocation on one page. Run the primitive map and the arbitrage test across the five engines.

Google AI Mode. Mature keyword-and-PMax auction, fully priced; “project management software” is among the most contested B2B terms there is. The test: priced, primitive fits, measurable — but no edge. Verdict: MAINTAIN. Verrine keeps its existing PMax and Search campaigns flowing into AI Overviews (near-zero setup) at roughly £180,000 — present and competitive, not over-fed.

Copilot. Microsoft’s auction, under-demanded relative to inventory, Bing-plus-LinkedIn substrate that fits a B2B buyer precisely, and mature enough to measure. The test passes on all four. Verdict: CONCENTRATE — £200,000, the single largest paid line, because a claimed +76% lower-funnel advantage against a thin-competition auction is the clearest mispricing on the board for this buyer.

ChatGPT. Context-only auction, underpriced consideration depth, and Verrine’s multi-turn “help me choose a PM tool for a 40-person agency” intent fits the primitive well — but measurement is weak. Verdict: CONCENTRATE, capped. £120,000 as a test-and-scale line with a geo-holdout read before any increase, because the mispricing is real but unmeasured spend is a bet, not an arbitrage.

Perplexity and Claude. No auction on either. Verdict: EARN-ONLY. No paid line; both are won, if at all, through the citation — and Perplexity’s high-intent research audience makes the earned effort well worth it.

The overriding line comes off the top: £100,000 to earned citation — the studies, contributed expertise and retrievable source pages that get Verrine named inside the answer on all five engines at once, funded before any auction because it is the only spend that spans them and the only route onto the two ad-free surfaces. Total: £100k earned + £200k Copilot + £120k ChatGPT + £180k Google. Now watch the naive plan misfire. A team following the standard template splits the £600k evenly across the three ad engines (£200k each), ports its Google keyword creative into ChatGPT’s context slot where it does not fit, over-pays Google where the auction is fully priced, under-invests Copilot where it is cheapest, and funds no earned citation at all — so it pays full auction price on every engine for a presence it could have earned, and buys nothing on Perplexity or Claude because there is nothing to buy. Timeline: month one, both plans look busy and similar. By month six, Verrine’s plan is compounding an earned presence across all five surfaces while paying a discount on the two underpriced auctions; the naive plan is paying par on Google, missing Copilot’s edge, wasting its ChatGPT budget on mismatched creative, and invisible on the ad-free engines. The gap is not effort. It is reading the auctions as one when they are five.

When the argument is wrong

The strongest objection is not hypothetical — it is already underway and has a name. Criteo became OpenAI’s first advertising-technology partner in March 2026; by June, more than 2,000 brands were running ChatGPT ads through its platform, access had extended to the UK, Japan and South Korea, and holding companies and independent agencies (Tinuiti among them) were plugging ChatGPT into their existing media-planning workflows, per Criteo’s own updates and PPC Land. Criteo activates over $4 billion in annual media spend across 17,000 advertisers. State the objection at full strength: “The intermediation layer you wave away is being built right now. Brokers and agencies will sit above the incompatible auctions and let you plan across all of them from one workflow. That is convergence in all but name, and your fragmentation thesis is a snapshot history will erase.” The evidence is real and the trend is genuine.

Concede it fully — and then draw the distinction that survives it. A broker over incompatible auctions is an activation layer, not a unified auction. Criteo lets you place bids in ChatGPT from a familiar console and fold the results into a cross-channel report. What it does not do — what it cannot do while the platforms keep their data private — is create a single auction in which one bid competes across Google, ChatGPT and Copilot on a common currency with common price discovery. Under the console, the auctions stay separate: Google’s keyword clearing, Microsoft’s Bing-LinkedIn clearing, ChatGPT’s contextual clearing, each pricing a different substrate the others cannot see. The broker reduces operational friction — one workflow instead of three — but it does not touch the economic friction, and it is the economic friction that lets returns stay unequal. Mispricing persists beneath the console precisely because the console cannot equalise what it can only translate.

So intermediation does not refute the plan; it delivers it. A single workflow that reaches every engine is exactly the tool you want for a concentrate-where-underpriced strategy — it lowers the cost of pressing the mispriced pocket without pretending the pockets have vanished. The advertiser who reads Criteo as “convergence, so spread evenly” gives back the edge; the one who reads it as “cheaper access to still-separate auctions” keeps it. The secondary objection — “the standard advice is just to run all the engines in parallel, which is simpler” — fails on the same seam: parallel presence is not arbitrage. Running all engines at an even split is the coverage exercise the map argues against; it pays par everywhere and captures the mispricing nowhere. The falsifier would be genuine convergence — the platforms pooling their targeting data into a shared exchange with one clearing price. Nothing about how any of them monetises its proprietary data suggests that is coming; every one of them is building a walled auction, not contributing to an open one.

The 2027 plan on one page

Five moves, and none of them depends on the engines converging — or on their staying apart.

• Map the primitives, not the logos. For each engine write what you actually bid, the substrate it buys, and how well you can measure it. The incompatibility across those columns is the plan’s starting fact.
• Hunt the underpriced engine. Concentrate paid budget where the auction is thin or opaque and its primitive fits your intent — Copilot for business intent, ChatGPT for considered multi-turn purchases today. Do not spread evenly; even spreading pays par everywhere.
• Fund the earned citation first. It is the only currency common to all five engines and the only route onto the ad-free two. It also lowers what you pay the auctions for presence you can instead earn.
• Build creative per primitive. Context descriptions for ChatGPT, keyword-and-audience assets for Google, Microsoft-audience creative for Copilot. Porting one engine’s assets to another forfeits both the fit and the edge.
• Measure by lift, per engine. None of these auctions gives trustworthy deterministic attribution. An underpriced surface you cannot measure is a bet; a holdout read turns it into an arbitrage you can scale with confidence.

The multi-engine ad auction that most planners are bracing for — one marketplace, one bid, one dashboard — is not the one arriving. What is arriving is a set of incompatible auctions plus an ad-free remainder, and that is better news than convergence would be: fragmentation is what keeps returns unequal long enough to exploit. Read the primitives, concentrate where the price is wrong, and fund the earned citation that spans all of them. For the wider system this plan sits inside, the core strategies hub, the foundational what-is-link-building explainer, the current best-tools shortlist and the running 2026 link-building statistics carry the connected threads.