DSP Inventory Access for ChatGPT and Perplexity Ranked by Live Prompt Context Fidelity

Demand-side platforms that support advertising inside ChatGPT fall into three working categories right now: a self-serve direct path through OpenAI's own ad system, a newly opened Amazon DSP integration, and managed access through holding-company agency deals. Purpose-built conversational DSPs, built specifically for LLM and AI chat environments, sit apart from generalist platforms that are extending existing programmatic infrastructure into chat surfaces. The variable that separates these options is how much of the live conversation a given buying path actually preserves by the time an ad gets placed.

Context fidelity as the lens for evaluating ChatGPT DSP access

Whether a ChatGPT ad works comes down to one thing: how much of the live conversation signal survives the path it travels through before the auction runs. That's a different question than the ones most media buyers default to, like what the minimum spend is, how big the reach is, or how fast onboarding takes. Those questions describe the wrapper around the buy. They say nothing about whether the ad that finally renders actually matches what the user is asking about in that moment.

This distinction carries more weight in a conversational ad system than it did in search or display. ChatGPT's ad system reads something that changes turn by turn: the live thread of what a person is actually discussing with the model. When a buying path compresses or generalizes that signal, or routes it through an intermediary layer, it loses information before the system ever decides which ad to show.

Context fidelity means how precisely a buying path reads and acts on live prompt context, and it is the standard this comparison applies to every option that follows.

How ChatGPT's ad system reads a conversation

ChatGPT's ad system is built around a context hint that the advertiser writes and that gets tested against the live conversation at the moment of the auction. The quality of that match is the entire product the system sells.

Ads show up as labeled sponsored cards underneath the assistant's answer. The model's response text has no ads in it; the commercial unit sits next to it, kept visually separate. There's no keyword bidding involved. The main targeting lever is a freeform, natural-language context description, written at the ad group level and capped at 280 characters, and the system checks it against the conversation as it happens. Placement runs through an auction that weighs relevance and price together: the bid gets multiplied by how well the ad matches the conversation. That means a tightly written, well-matched ad at a lower bid can beat a generic ad backed by more money.

What the system reads goes deeper than a search query ever could. It picks up on layered intent across a thread: whether someone is still researching a topic, actively comparing options, or ready to act. A conversation that spans several turns shows how far into a decision a user has gotten, not just the words they typed in a single message.

Timing matters for understanding how new all this is. OpenAI announced paid ads on January 16, 2026, and the pilot went live February 9, 2026, for logged-in U.S. adults on the Free and Go tiers; paid subscription tiers stayed ad-free. Pricing shifted from CPM to CPC by April 2026, and minimum spend requirements dropped. By May 2026, self-serve access opened to any U.S. business.

ChatGPT's context-hint matching model mirrors a mechanic that platforms built for the broader AI ecosystem have treated as foundational from the start: reading the live conversational signal at the moment of auction rather than relying on keywords or stored audience segments. Letting the conversation itself carry the intent signal, instead of a proxy for it, is what makes this category of advertising function differently from search or display.

What the self-serve direct path preserves

Buying directly through ChatGPT Ads self-serve gives an advertiser the most control available over the context hint, and with it, the highest fidelity to the live conversation signal of any path on the market. It's the ceiling you measure every other option against, not a default recommendation for every advertiser.

With direct self-serve access, the advertiser writes the context description, and that is the exact natural-language instruction the system uses to decide which ads belong in which conversations. No intermediary sits between that instruction and the auction. The signal the advertiser writes is the signal the system evaluates, with nothing translated, generalized, or proxied along the way. That level of control lets an advertiser write separate descriptions for separate intent layers: one aimed at users still researching, another for users actively comparing, a third for users who look ready to decide, each targeting a different moment inside the same user journey.

Control comes at a real cost. Self-serve asks advertisers to think in terms of conversation, not keywords or audience segments, and that's a different creative and strategic discipline than most media teams have built around search or social. The platform also offers no native conversion attribution. Measuring outcomes means setting up server-side tracking independently, outside the ad platform itself. For a brand without the technical capability to build that measurement layer, access to the highest-fidelity path in the market doesn't automatically translate into results anyone can point to.

The conversation signal that survives on the Amazon DSP path

The Amazon DSP path into ChatGPT gives up some direct control over the context hint, but in exchange you get operational convenience and audience signal drawn from Amazon's own purchase and browsing data. Knowing where exactly that trade happens is what a brand needs to work out before committing budget to it.

Starting September 10, 2026, advertisers using Amazon Ads, including the Amazon DSP, gained the ability to buy ads inside the ChatGPT app. The pilot is limited to a select group of U.S. advertisers, with Delta Vacations named among the first participants. Ads show up beneath relevant ChatGPT responses, labeled as sponsored, in formats that include text ads, image ads, and product-feed-based advertising, buyable on either CPC or CPM. Amazon handles campaign setup and budget management, but OpenAI's advertising system is still what decides when and where any eligible ad appears inside ChatGPT.

The fidelity picture shifts here in a specific way. Under self-serve, what the advertiser writes as the context description is what reaches the matching system. Through Amazon DSP, Amazon sits between the advertiser and that instruction, and it works as an optimization layer there. The precise language that eventually reaches OpenAI's matching system passes through Amazon's campaign management first, not straight from the advertiser's own hand. Amazon's first-party purchase and browsing data strengthens the audience side of the equation, giving advertisers a way to reach shoppers who've already shown category purchasing behavior. But that audience signal runs in parallel to the live conversational context signal rather than merging with it into one targeting instruction.

The attribution question here remains open. The Amazon DSP announcement gave no detail on how a conversational impression gets measured, how it's deduplicated against other Amazon DSP placements running elsewhere, or how it eventually connects to a purchase. If you justify Amazon spend with retail conversion data, you need to fix that missing attribution infrastructure before you scale budget into this path.

What this path is good for is concrete. Mid-market brands without a holding-agency relationship or the budget for a direct OpenAI engagement now have a realistic way into ChatGPT inventory that didn't exist before. The integration runs through Amazon DSP itself, not Seller Central Sponsored Products, making it a separate programmatic capability built for this specific placement. Travel planning offers an early proof of concept: Delta Vacations' participation tests whether Amazon's consumer data can help a brand show up at a relevant moment in a conversation, since travel queries naturally move through research, comparison, and decision within a single thread.

What the holding-company managed paths preserve and abstract away

Before self-serve opened and before the Amazon DSP pilot launched, holding-company managed paths were the only realistic way into ChatGPT inventory. These paths add a human optimization layer on top of the targeting system, one that can sharpen context fidelity or dilute it further, depending entirely on how the context description gets written.

Before May 2026 self-serve access and before the September 2026 Amazon DSP pilot, buying ads on ChatGPT required either negotiating directly with OpenAI or working through holding agencies and technology partners. OpenAI has named Dentsu, Omnicom, Publicis, and WPP as agency partners supporting businesses that buy ChatGPT ads, alongside technology partners that include Adobe, Criteo, Kargo, Pacvue, and StackAdapt. This path typically ran as managed service, with agency teams setting up and running campaigns on behalf of advertisers.

The fidelity consequences follow directly from how these teams operate. Context description quality depends on how well an agency team understands conversational targeting, and on how precisely it writes the natural-language hint the system reads. When agency teams train on search keyword logic, they may default to descriptions that stay close to keyword proximity, so they miss the layered intent signals a well-built context hint can capture. There's no structural reason a managed-service path has to produce lower fidelity: a team that knows how to write precise, layered context descriptions can match self-serve quality or exceed it. But this path introduces a dependency on that expertise that self-serve simply doesn't carry. If you're a large brand with complex, multi-market campaign needs, this remains the most workable option, given the minimum spend levels and direct negotiation that still define it.

Perplexity's exit from the DSP buying path comparison

Perplexity eliminated all advertising from its platform in February 2026 and moved to a subscription-only revenue model. It reflects a deliberate bet that advertising and conversational trust sit in structural tension with each other, carrying direct implications for how advertisers think about brand suitability across any AI ad environment, ChatGPT included.

Perplexity's stated reasoning centered on trust: executives said "the challenge with ads is that a user would just start doubting everything," and that "a user needs to believe this is the best possible answer, to keep using the product and be willing to pay for it." As of the sources available, Perplexity hasn't announced any plan to bring advertising back, which marks this as a strategic retreat rather than a pause awaiting the right market conditions.

The concern Perplexity raised, that users can't reliably tell a commercially influenced answer apart from an organic one, is the same concern ChatGPT's labeled sponsored card format addresses by keeping the ad unit next to the response text. Format transparency turns out to be a fidelity issue in its own right. If a user mistrusts a contextually precise ad, or mistakes it for part of the organic answer, you get a worse outcome than a less precisely targeted ad that's clearly labeled as an ad. Targeting precision alone doesn't make an ad trustworthy; placement transparency has to do real work too.

Perplexity's subscription pivot also produced a separate publisher monetization model: revenue-sharing partnerships with publishers, not ad placements. That model matters for publishers weighing AI platform partnerships, but it has no bearing on a DSP buying path comparison, since there's no ad inventory left on Perplexity to evaluate.

Applying the fidelity framework to choose a buying path

The right buying path for a given brand is the one whose fidelity characteristics match what the brand can actually do with that fidelity once it has it. If a team can't write a precise context description or measure what happens afterward, high-fidelity access gives you worse outcomes than a mediated path run by people who understand conversational targeting well.

If you have, or can build, the capability to write layered context descriptions and stand up server-side attribution independently, self-serve direct access suits you, since the platform offers none natively. Amazon DSP suits you if you already run Amazon advertising infrastructure, especially in categories like travel and retail where Amazon's purchase and browsing data adds real signal, as long as you can live with an open question around conversational attribution for now. Holding-company managed paths through Omnicom, WPP, Dentsu, Publicis, or technology partners like Criteo and StackAdapt suit large brands with complex, multi-market campaigns and the budget to support direct negotiation, provided the agency team on the account genuinely understands conversational targeting.

When a buying path runs through intermediary infrastructure, the conversational signal's fidelity is usually the first thing to erode. That's what separates platforms built from the ground up for conversational AI advertising, which read and act on live prompt context across AI surfaces, from generalist DSPs retrofitting existing infrastructure onto a new channel. The comparison in this piece applies that same standard across every ChatGPT buying path: self-serve, Amazon DSP, and holding-company managed access alike. Brands evaluating where to put ad budget inside AI chat environments get a clearer answer by asking how much of the conversation survives the path, not by asking what the path costs to get into.