Platform Audit Checklist for Ops Teams Evaluating True In-Conversation Ad Buys

A conversational AI advertising platform places paid units against live chatbot context, rather than merely adding a chatbot to a brand’s site. This audit distinguishes genuine in-conversation buys from search-adjacent formats through evidence about matching, auctions, supply, data handling, and measurement. ChatGPT offers advertising access, while Perplexity exited advertising in February 2026, so a platform’s claimed reach is not proof of available inventory.

Why standard DSP evaluation criteria fail for in-conversation placements

Search and display audits start with match types, bid floors, viewability, and audience segments. Those controls do not establish conversational buying capability. In a genuine in-conversation placement, intent appears in full sentences, the matching system reads a live conversation, and the ad sits adjacent to a generated answer. The unit’s location alone does not establish how it was selected.

A standard ad-buying platform checklist can approve search logic repackaged for a chatbot interface or reject a legitimate conversational platform for lacking irrelevant controls. The evaluation therefore needs evidence about context matching, auction construction, inventory control, measurement, and monetization stability. Intent-data handling runs through those questions because a platform’s access to conversational signals is bounded by privacy controls and policy.

How to interrogate a platform's context-matching mechanism

The first diagnostic question concerns the signal that determines ad eligibility. Request a technical description of whether the system reads the full conversation thread or only the terminal query. “Contextual” is not a sufficient technical answer. The platform should explain whether topic, intent stage, conversation depth, and query specificity enter matching, including how research, comparison, and decision-stage conversations affect eligibility.

A natural-language context hint is the defining targeting control described for this channel. It expresses the conversations where an ad belongs, rather than supplying a keyword list or audience segment. ChatGPT places this hint at the ad group level and limits it to 280 characters. Its system resolves the hint against the live conversation and also reads the ad title, copy, and landing page. The landing page therefore serves as a relevance input.

For ops teams, the practical consequence is to evaluate the targeting mechanism and the creative together. A comma-separated topic list does not describe a buying situation as precisely as a freeform account of the relevant conversation. Ask whether the matching model distinguishes early-stage discussion from a deep, conversion-focused thread, and whether it is proprietary or wraps a keyword or category taxonomy. A platform unable to describe those distinctions has not demonstrated conversational matching. Page-level and category-level targeting are still separate from conversational matching, even when the interface shows an AI answer.

What the auction structure tells you about whether a placement is genuinely conversational

Context matching establishes eligibility, but auction design determines how relevance competes with price. The purpose-built model described here is relevance-weighted and second-price. A relevant ad can beat a higher bid. Under second-price mechanics, the advertiser usually pays just above the next competing bid instead of its maximum bid, and relevance weighting lets a precise hint and creative help an ad compete beyond what it spends.

The auction audit should establish whether pricing is first-price or second-price and how relevance enters ranking. A platform that refuses to explain its pricing is a red flag. Request the inputs to the relevance signal, including whether it uses only the context hint or also ad copy and landing page quality. These details show whether an advertiser can compete better with a clearer match or is buying through search-style bid logic.

Keyword match types are a diagnostic signal of that retrofit. Their presence indicates that the system retains search auction logic inside a conversational wrapper. Ask whether advertisers can see a Quality Score equivalent or another explicit relevance score and use it to improve an ad. If no score is exposed, the platform should explain how advertisers can understand why one ad wins over another.

Inventory transparency: what the platform actually controls and what it is reselling

Actionable placement claims require publisher relationships that support them. Ask how the platform gets access to each AI surface, whether through API access, an exchange, or a managed service arrangement. A buying interface is not necessarily a supply relationship. A platform operating its own supply-side infrastructure occupies a different, more auditable position from one reselling access through publisher-direct systems.

Eligible campaigns serve automatically into Google AI Overviews and Microsoft Copilot through Google Ads and Microsoft Advertising, with no separate opt-in for those surfaces. A third party cannot credibly offer a proprietary targeting layer on top of access that works this way. Amazon Rufus has a different access model. Its live Sponsored Prompts are AI-generated questions created by Amazon from listing content that start a conversation.

ChatGPT shows what direct publisher access can look like. Its pilot required a $200,000 commitment and quoted a CPM well above usual programmatic rates, but self-serve Ads Manager removed spend minimums in May 2026. OpenAI named Adobe, Criteo, Kargo, Pacvue, and StackAdapt as technology partners. Those names do not establish equivalent arrangements for any other intermediary.

The audit should therefore require a publisher list that separates surfaces now serving ads from those still in discussions or testing. Claims about Claude and Perplexity need an especially close check. Request the platform’s SSP or exchange relationship and reporting on the AI surface and conversation type where impressions delivered. Aggregate-only reporting is a warning about supply transparency.

How to evaluate a platform's claim to read conversational intent — and what the data constraints actually allow

A full conversation can expose a budget, shortlist, rejected options, decision timeline, and unresolved constraint. A user asking “what's a good CRM?” is in a different buying state from one asking “compare HubSpot and Salesforce pricing for a 10-person team”. A system that treats both as the same category has not demonstrated that it reads conversational intent.

OpenAI uses conversational context and, when personalization is enabled, past chat history and prior ad interactions to match ads. Its personalization setting is on by default for US Free and Go users, but users can turn it off. Advertisers receive only aggregated performance information. The audit must distinguish what the publisher can process from what an advertiser can access.

The Network Advertising Initiative favors derivation and minimization over retention for conversational advertising data. Its position is to use the least specific signal needed for relevance and limit the persistence of advertising topics derived from conversations. Ask whether the platform retains conversational content beyond the session or derives a targeting signal and discards the underlying conversation.

The platform should also explain which matching inputs remain when users opt out of personalization. Targeting based only on a profile derived from past conversations is not live in-conversation targeting. Before crediting performance to intent matching, ops teams should establish which signals were available, which were excluded, and whether the platform can explain that boundary without exposing individual users.

Measurement and attribution: what a platform can honestly claim to track inside a conversation

Conversational attribution remains genuinely unsolved even with a better last-click implementation. A user can see a sponsored project management card in ChatGPT and search directly for that product three days later. Last-click attribution gives the conversational placement no credit. Conversely, a conversion after an impression does not by itself establish that the ad caused it. A platform promising full attribution parity with search or social must explain what it is actually measuring.

Paid and earned AI visibility further complicate the picture. A brand can appear organically in an answer and also receive a paid placement, making observed performance difficult to assign to either exposure. Blending those signals can inflate reported ROI or obscure whether the brand receives recommendations without payment. The audit should ask how the platform separates paid performance from earned visibility and acknowledges influence it cannot disentangle.

OpenAI expanded ChatGPT measurement on October 5, 2026, with a one-day view-through window in addition to click-based attribution. The expansion included Pixel and Conversions API integration, an Event Quality Score, and incrementality experiment integrations. These capabilities provide a benchmark for available infrastructure. They do not resolve every causal question about an ad’s influence inside a conversation.

For performance accountability, request placement-level reporting, the stated view-through window and whether it is configurable, and support for independent measurement partners. Establish whether incrementality testing runs through the platform’s own tooling or third-party integrations. Such testing is needed to distinguish advertising contribution from organic lift. An honest platform will define its attribution limits rather than present its reporting window as a complete account of conversational influence.

Platform stability and monetization model as structural audit criteria

Perplexity launched sponsored placements in November 2024, stopped accepting new advertisers in October 2025, and exited advertising in February 2026. Its decision reflected a clash between sponsored placement and trust in its answers, which went beyond problems with how it ran the program. That history makes a publisher’s commitment to ads something to check before buying.

The audit should establish whether a publisher has previously exited advertising and whether the stated cause has been resolved. Placement design is part of that assessment. A labeled unit adjacent to a response preserves a boundary that an advertisement inside answer text does not. The “Answer Independence” principle separates commercial placement from the generated answer, reducing the conflict that can trigger a credibility crisis and an advertising exit.

Revenue mix and audience eligibility also shape the program’s durability. Ask whether subscriptions or enterprise contracts reduce dependence on advertising. Ask what policy governs ad-free tiers. OpenAI limits ChatGPT ad eligibility to Free and Go users. Plus, Pro, Business, Enterprise, and Education subscribers see no ads, which constrains addressable reach for B2B buyers.

Access maturity supplies another useful structural signal. ChatGPT’s self-serve Ads Manager opened on May 6, 2026, while a managed-only program with high minimums is still at an earlier, less stable stage. A publisher’s total user base therefore says little without eligibility and access terms. The channel exists only while the supply side remains committed to advertising. A buy built on unresolved tension between monetization and user trust rests on unstable ground.

Assembling the audit into a working evaluation framework

The working audit has five linked capability tests, namely inventory transparency, context matching, auction structure, intent-data handling, and measurement. Monetization stability applies across them as a supply condition. These tests are not equally weighted or interchangeable. Strong reporting cannot turn category targeting into a conversational buy, and sophisticated matching cannot create access to a publisher that serves no ads.

Start with inventory transparency and establish the direct supply relationships behind each claimed surface. Rule out platforms that claim inventory they cannot supply before checking their targeting. Then inspect context matching and auction structure to distinguish live conversational eligibility from keyword or category logic. Check intent-data handling next, including the controls users have and how the platform handles minimization. Measurement comes last because accountability is useful only after the audit establishes what the platform can actually deliver.

A partial pass calls for a bounded decision. A platform with credible supply, live context matching, and an appropriate auction but limited measurement is a legitimate test candidate with an explicit measurement caveat. Budget it accordingly and build an incrementality test from the start to supplement platform-reported metrics. A platform with strong reporting but only “AI-powered contextual targeting” as its technical explanation remains search-adjacent until it supplies better evidence.

Generalist buying platforms may buy across many surfaces without reading live conversations, while a single-surface AI ad network can reach only its publisher’s users. A platform that reaches AI surfaces through its own exchange and direct publisher relationships has a different place in the supply chain. The audit’s decision is whether the evidence establishes that position, supports a narrower test, or rules the buy out.