Cross-Surface Budget Allocation for Multi-AI-Platform Campaigns

A campaign spread across ChatGPT, Google's AI Overviews, and Microsoft Copilot is not one channel with three faces. Each surface runs its own inventory structure, its own targeting model, its own audience mix, and its own measurement system, so a single spending ratio applied across all of them will overfund some and starve others by design. Treating "AI advertising" as one line item is the mistake this piece sets out to correct.

Start with what each surface actually is. ChatGPT runs a purpose-built ad system: a labeled sponsored card sits below the answer, never inside it, and reaches only logged-in adults on the Free and Go tiers. Every paid tier, Plus, Pro, Business, Enterprise, and Education, carries no ads at all. Targeting runs through a 280-character natural-language context hint rather than keywords, resolved by a relevance-weighted second-price auction, with CPM, CPC, and oCPC (Conversions) objectives on offer. OpenAI opened a self-serve Ads Manager on May 6, 2026, with no spend minimum, after a beta phase that had required a substantial minimum commitment.

Google's AI Overviews and AI Mode work on a different premise entirely: they are monetized through the Search, Shopping, and Performance Max campaigns an advertiser already runs, with no separate opt-in or opt-out for advertisers, though publishers gained an opt-out through Search Console as of June 2026. There is no segmented reporting for these AI surfaces specifically, even though Google Search Console added dedicated organic impression reports for them. New formats, including Conversational Discovery Ads and Highlighted Answers, are in testing via Google Marketing Live. Microsoft Copilot automatically opts in every eligible campaign and ad type, with no way to opt out, building its ads from assets already in the advertiser's account. Its targeting reads the full arc of a session rather than just the latest query, and it reports clicks, CTR, conversions, CPA, and ROAS, with impressions visible through placement reports, though ROAS models may need adjusting for conversational placements.

Two other players round out the map by their absence. Perplexity launched sponsored answers in November 2024, then stopped accepting new advertisers and wound the program down. Its leadership told the Financial Times that sponsored placement risks making users suspicious of the whole answer. Anthropic has never sold ad placement in Claude and has positioned itself against the practice. Neither is a live buying option today, and both belong in the map anyway, because they define the boundary of what's investable.

None of this divergence happened by accident. Four companies evaluated the same commercial opportunity and reached four different conclusions about how, or whether, to monetize attention inside a conversation. Perplexity's retreat is the clearest evidence that user trust functions as a hard constraint on these surfaces, not a soft marketing consideration to be managed with disclosure language. Any allocation framework built on the assumption that these are versions of the same channel will misread that signal and misallocate the budget that follows from it.

How each surface's targeting model changes what a budget buys

A dollar spent on ChatGPT enters a different machine than a dollar spent on Copilot or on Google's AI surfaces, and the difference shapes what that dollar is actually buying. It determines what that dollar is actually buying. Allocation can't be handled with one CPM or CPA target stretched across surfaces.

ChatGPT has no keyword layer. The advertiser writes a 280-character context hint, a natural-language description of the conversations where the ad belongs, and the system matches that description against the live conversation as it unfolds. That matching doesn't stop at the most recent message. It reads the entire conversational trajectory: what the user asked several exchanges earlier, how the model responded, and where the exchange appears to be heading. That is a fundamentally different unit of targeting than a page of content matched to a search term, and it rewards advertisers who think in terms of conversation shape rather than topic labels. Three signals decide whether an ad performs on ChatGPT: the context hint itself, the ad's title and copy, and the landing page it points to. Getting all three in tight alignment is the single highest-leverage optimization available on the surface.

Copilot's targeting runs on what Microsoft calls "ad voice," a read of the entire session rather than the last query alone, tracking the conversation's direction and intent as it develops. That matters disproportionately for B2B advertisers, because Copilot sits inside Microsoft's productivity tools, where session context can carry professional intent signals that consumer chat surfaces simply don't generate. There's also a budgeting wrinkle specific to Copilot: no separate budget line exists for it. Spend flows automatically from whatever campaigns are already running, so the real decision isn't how many dollars to earmark for Copilot, it's which existing campaigns qualify for Copilot placement and how well their assets are built for conversational relevance.

Google's AI Overviews and AI Mode work through inherited auction logic rather than a conversational targeting layer. Eligible campaigns serve automatically, and the targeting model is the existing Google auction, not a conversational layer. Ads only appear where commercial intent is detected in both the user's query and the content of the AI Overview itself, a dual-relevance test that Google's system evaluates on its own, without advertiser input. Combined with the absence of segmented reporting for these placements, that means an advertiser can't isolate how a budget performs inside the AI surface from how the rest of the campaign performs, which limits how much confidence anyone can place in an AI-specific allocation decision on Google today.

The practical failure mode occurs whenever teams treat these mechanics as interchangeable. A context hint written as a comma-separated list of topics matches poorly on ChatGPT, because the system is reading conversational intent, not a keyword string. Running identical creative across all surfaces misses that Copilot assembles its ads from existing account assets while ChatGPT evaluates the landing page as a targeting signal. Budget planning has to start from these mechanical differences instead of a shared CPM assumption carried over from search or social.

The audience tier constraint as the first allocation filter

Before any performance metric enters the allocation decision, an advertiser has to account for which users each surface can actually reach, because on the most prominent surface, ChatGPT, a majority of the most commercially valuable users are structurally excluded from seeing ads.

The ChatGPT constraint is architectural, reaching only logged-in adults on the Free and Go tiers. Every paid subscriber, Plus, Pro, Business, Enterprise, and Education, sees no ads at all, by OpenAI's own documentation. For consumer brands with broad free-tier audiences, that ceiling matters less. For B2B advertisers, it matters a great deal, because the users excluded from ChatGPT ads skew toward higher-intent, higher-spend professionals, the same people most likely to be paying for a subscription in the first place. Any B2B team allocating budget to ChatGPT needs to weigh that exclusion explicitly before the first dollar moves. Excluding this audience still leaves ChatGPT a viable surface for B2B. It means the allocation has to price in the audience ceiling rather than treating total ChatGPT user volume as a stand-in for the audience a given campaign can actually reach.

Copilot sits on the other side of that constraint. Every Copilot user above 13 (above 18 in some regions) who can see Microsoft ads across its services is reachable with no paid-tier exclusion, though users aged 13-17 see only non-personalized, contextual ads and cannot be reached via audience-based targeting. Because Copilot is embedded directly inside Microsoft's productivity software, the audience it reaches includes professionals in the middle of active work, a session-level intent signal that consumer-facing chat surfaces have no equivalent for.

Google's AI Overviews sit somewhere in between: broad reach, but conditional delivery. Ads inside AI Overviews run across multiple countries with no subscriber-tier exclusion of any kind, but the format excludes sensitive verticals outright, including finance, healthcare, alcohol, gambling, politics, and adult content. And because delivery depends on that dual-relevance test between query intent and AI Overview content, an eligible campaign is not guaranteed to serve even where the audience is reachable. Reach is broad; delivery is conditional.

The context hint, the ad title and copy, and the landing page are the three signals that matter for bid strategy, and precise alignment across all three is the highest-leverage optimization available. Surface budget weight should track the share of each surface's reachable audience that actually matches the campaign's customer profile, weighed against the surface's raw user count. A consumer brand with a wide free-tier footprint has a straightforward case for ChatGPT spend. A B2B software company whose buyers sit almost entirely on paid tiers has a much weaker one, and no amount of bid optimization changes that math. Map the customer profile to the reachable tier on each surface before touching a single bid number. That step comes first, not last, because no performance metric downstream of it can be trusted if the audience underneath it was never reachable to begin with.

Conversational intent signals should drive budget weighting over impressions

Once the reachable audience is mapped, the next question is what that audience is actually revealing inside the conversation, and conversational prompts carry intent signals sharper than anything a keyword or a demographic bucket can produce. A budget framework that weights spend by impression volume instead of intent signal strength will underfund the most valuable inventory on every surface it touches.

Consumers are increasingly using AI chat to work through a purchase decision before they ever open a search engine, and those exchanges generate a layer of intent data that doesn't exist anywhere else in the marketing stack. A user who never types "project management software" into a search box but spends several exchanges describing team coordination problems, missed deadlines, and the friction of remote collaboration is showing exactly the kind of buying signal a keyword system is built to miss. Conversational targeting has to read the problem being described and the trajectory of the conversation itself.

That intent builds in stages inside a single conversation. An opening query establishes the initial ask. Follow-up questions narrow the consideration set. Later exchanges start to signal readiness to act. Research cited in the surface data finds that users go roughly six prompts deep before they leave the conversation for the open internet to convert, making the mid-to-late stretch of the conversation the highest-value ad window. Conversion timing after that point also varies by category: some purchases, flights and electronics among them, close within a short window once the conversation ends, while others run on a longer consideration cycle. Budget and bid timing should track that category-level pattern rather than apply one attribution window to every product line.

A sportswear example makes the funnel-mix point concrete. A meaningful share of prompts in that category are upper-funnel and informational, asking about fit, material, or use case rather than naming a product to buy, while a comparable share are directly transactional. The informational prompts still carry buying signals that a traditional search query would never surface, because the user is describing a need rather than typing a product name. Both phases deserve budget, but they call for different bid levels and different creative, with the informational stage built to earn trust and the transactional stage built to close.

Intent signals cut both ways, though, and allocation has to account for the risk side as well as the opportunity side. When a user names a brand inside a prompt, a competitor's sponsored placement can appear right alongside it, and in a conversational surface that can land harder than the equivalent moment in search, because the user often experiences the exchange as advice rather than as a visible auction for ad space. A disciplined allocation plan sets aside a defensive budget line for these brand-adjacent conversational moments, borrowing a practice long familiar from search bidding but adapting it to a surface where the ad doesn't look like an ad to the person reading it.

The measurement gap that makes cross-surface attribution unreliable today

None of the allocation logic above can be optimized with precision yet, because the measurement infrastructure across these surfaces has not caught up to the spending flowing into them. The attribution problem in AI advertising is structurally different from the attribution problem in search or social, and applying last-click or single-session models to conversational AI campaigns will systematically undercount the channel's contribution and distort subsequent allocation decisions.

The core mismatch is conversational. A user can ask several follow-up questions inside a chat, leave without converting, and complete the purchase later through an entirely different channel. A last-click model then credits that later channel and records zero contribution from the AI surface that actually built the intent. Given that users go roughly six prompts deep before moving to the open internet to convert, the conversion event is almost never in the same session as the conversational moment that generated it, a structural gap that standard attribution windows were never built to bridge.

Each surface adds its own version of the problem. Google's AI Overviews and AI Mode offer no segmented reporting at all, so performance inside the AI surface can't be separated from the rest of the campaign's performance, and advertisers have no way to know what share of their conversions the AI placement actually drove. Copilot offers a fuller metrics suite, impressions, clicks, CTR, conversions, conversion rate, CPC, and ROAS, but those numbers measure what happened at the ad unit in a given session. They don't capture the conversational journey that led a user to that session in the first place. ChatGPT reports CPM and CPC data, but the platform is still early, and its reporting layer has not caught up to the depth mature search and social platforms have built over years.

The measurement gap is set to widen before it narrows. OpenAI is building commerce infrastructure directly into ChatGPT through its Agentic Commerce Protocol and Instant Checkout. Once an ad can execute a booking or a purchase inside the conversation itself, the conversion event happens inside the platform, and the advertiser's own attribution system may never register it at all.

That gap doesn't invalidate cross-surface budgeting. It sets the honest boundary around how much confidence any allocation model can claim today. A framework built from the surface mechanics, the tier constraints, and the intent signals described above gives an advertiser a defensible basis for where to put the next dollar. What it cannot yet give is a clean, closed-loop number proving that dollar worked, and any team presenting cross-surface attribution as settled science is overstating what the current measurement layer can support.