Trafficking and Tagging Standards for In-Conversation Paid Placements on ChatGPT and Perplexity

ChatGPT is the only large conversational AI advertising platform with live, self-serve paid placements available at scale as of October 2026. Perplexity has exited advertising, while another vendor’s chatbot serves ads from that vendor’s existing advertising campaigns. Buying these placements means entering conversations on a platform. Effective trafficking therefore requires ChatGPT-specific creative, context hints, tagging and safety controls, with measurement designed around the limits of conversational attribution.

Which platforms let brands buy ads inside live AI chatbot conversations

OpenAI began testing ChatGPT ads in February 2026 and added a self-serve Ads Manager plus CPC bidding on May 5, 2026. For practitioners, ChatGPT now provides the distinct buying workflow this guide addresses. Another vendor’s chatbot serves ads automatically from that vendor’s existing advertising campaigns, rather than requiring a separate conversational trafficking process. The available options are not interchangeable buying surfaces.

Perplexity stopped taking new advertisers in October 2025 and ended advertising in February 2026. It cannot be trafficked as of October 2026, so practitioners should remove it from paid media plans. Its exit also explains the market’s concentration. After running the first at-scale conversational AI ad experiment, Perplexity concluded that advertising created a trust problem it could not solve and pivoted to subscriptions. The split between ad-supported and subscription-led services makes user trust an operating constraint.

How ChatGPT's ad system works architecturally — and why it is different from keyword search

ChatGPT matches ads using the context and intent of the current conversation, the landing page, ad copy, image and context hints from the advertiser. When a user enables ad personalization, select signals from the broader ChatGPT experience can also inform matching. A search phrase is not the organizing unit. The system instead interprets needs expressed in natural language, including budget, features already considered and stage of the buying journey.

That distinction gives practitioners a basis for structuring ad groups around four intent layers. Informational intent, where the user is researching, fits awareness objectives, while comparative intent calls for creative that shows how the options compare. Transactional intent fits direct response, and problem-solving intent calls for messaging that shows how to solve the user’s problem. Each layer describes a different job for the ad. Ad groups should be built around those different jobs instead of copying the keyword structure of a search account.

The auction combines bid with a relevance score and uses a second-price mechanism. Reach planning also has a firm audience boundary because ads appear only to Free and Go users. Plus, Pro, Business, Enterprise and Edu subscribers are ad-free. Advertisers buy a separate sponsored placement while the answer remains independent, a distinction that also governs the safety controls discussed below.

Creative specifications: what ChatGPT Ads Manager actually accepts as of October 2026

The confirmed placement is a sponsored unit below the answer. Its title accepts up to 50 characters, and its description accepts up to 100 characters. OpenAI recommends roughly 16 to 24 characters for the title and roughly 32 to 48 characters for the description. Out-of-spec creative faces rejection rather than automatic resizing, so the shorter recommended lengths should not be confused with the maximum permitted lengths.

The minimum budget is $25 per day in USD (EUR 15, GBP 15, INR 725), averaged over a week. Campaign objectives include Views (CPM), Clicks (CPC), and Conversions, with the objective shaping the applicable specifications and bid floors. Views optimizes for visibility, while Clicks optimizes for traffic. Objective selection is part of trafficking. The presence of a Conversions objective should not, by itself, be treated as evidence of a complete purchase-attribution system.

The landing page must be publicly reachable and open to the OAI-AdsBot crawler. It is also a targeting input. OpenAI reads the page to assess relevance before serving the ad, so weak alignment between the page and the intended conversation can reduce match quality and suppress delivery despite a competitive bid. Practitioners should allow OAI-AdsBot through robots.txt and other bot-blocking rules before launch. That pre-flight task supports both eligibility and the matching process that connects the finished unit to a conversation.

Conversation context hints: how targeting works without keywords

Context hints describe the intent of conversations where an ad should appear. A useful brief captures the user’s need, constraints and decision stage so that the ad copy and landing page address the same situation. Commercial frames including “best for,” “vs,” “alternatives,” “pricing,” “security,” “migration,” and “integrations” provide starting points. Their value lies in the decisions they express, rather than in treating each phrase as an exact-match trigger.

Specificity makes those descriptions operationally useful across different categories. For B2B campaigns, the depth of a user’s technical questions about a product category can show purchase intent on its own, without demographic signals. Context hints should capture that depth. The same principle applies to e-commerce. “Running shoes for marathon training in hot weather” describes a more precise need than “running shoes,” and the hint should preserve those constraints.

Practitioners should maintain a curated prompt bank of high-intent conversation frames as the working basis for context hints and ad groups. This organizes the work much as a search keyword list does, without assuming that ads are matched by keyword. Eligibility still constrains the conversations that can be targeted. Finance and health briefs must explicitly exclude vulnerable-user situations, because category approval does not grant unrestricted contextual permission.

Context hints operate with other available targeting inputs. Country applies to every campaign, while standard campaigns support states, cities, markets and postal codes where available. Platform options include Android app, Android web, desktop web, iOS app and iOS web, with custom audiences built from email or phone number lists. Audience size also affects usable targeting. Custom audience inclusion and bid adjustments require a minimum number of matched users, and OpenAI recommends a substantially larger pool for reliable delivery.

Bidding mechanics and what the current cost data actually tells practitioners

ChatGPT’s real-time, relevance-weighted auction makes cost dependent on more than the bid. Reports from the pilot show a wide CPM range, and those figures give a rough guide to costs. Actual costs vary by account, category, geography, objective, eligibility and auction conditions. The evidence is still early. With only a few months of self-serve-scale activity behind the channel, comparisons with Google Search or Meta provide directional context.

Bidding decisions should start with what the campaign aims to achieve. Views uses CPM pricing and optimizes for impression delivery, making it appropriate when the goal is presence at a relevant conversational moment. Clicks uses CPC pricing and optimizes for traffic, so it fits direct-response campaigns that can measure what happens on the landing page. This choice should precede the tracking setup. An awareness campaign and a traffic campaign need different evidence of success, even when they use the same creative and context hints.

Conversation-depth bidding has been reported as a forthcoming feature. It should not anchor current campaign setup. Practitioners can instead connect the available objective to the outcome they can actually observe, while treating early cost reports as provisional. That restraint is especially important when click costs appear precise but the channel’s contribution to later purchases remains difficult to attribute.

Tracking parameters and the tagging setup ChatGPT requires

UTM parameters persist on clicks from ChatGPT ads, but ensuring that OpenAI’s oppref click parameter survives redirects to preserve source attribution is a launch requirement. Link shorteners, affiliate tracking layers and other redirect chains can strip the parameter before the visitor reaches the destination. Practitioners should check the full redirect path before launch. A tagged starting URL is insufficient if the destination loses the information needed to identify the click.

The standard tagging structure should populate utm_source, utm_medium, utm_campaign, utm_content and utm_term. However, the content and term fields should describe the conversation layer. There are no keywords to report here. Those fields can distinguish the campaign’s intended context, provided practitioners do not confuse their own labels with evidence of the actual conversation that triggered delivery.

Along with UTMs, practitioners should plan to record conversation depth, primary intent signal and conversation timing when those data are available. Depth describes the number of exchanges before a click, while timing describes how long the user engaged before clicking. The primary intent signal concerns the phrases or questions associated with the match. These are measurement requirements. Adding custom parameters to a URL does not establish that Ads Manager exposes the underlying conversation data or can populate those parameters.

OpenAI’s advertising policies have changed several times since launch, so an older setup may no longer meet the requirements. Practitioners should check OpenAI’s current documentation directly. The practical recommendation is to bookmark the policy page and assign a review owner. Changes can affect required parameters, permitted creative and eligible categories, linking tagging maintenance to the wider trafficking process.

The attribution gap: what ChatGPT ads cannot yet measure and how to work around it

Conversational discovery does not always leave a trackable click path. A user may encounter a brand while researching in ChatGPT and convert weeks later through a direct visit or branded search. Standard last-click attribution cannot connect that discovery to the purchase, and multi-touch models may also miss the conversational session. This is a structural measurement gap. It reflects a mismatch between how conversations shape decisions and how attribution systems record identifiable visits and interactions.

The early rollout description from April 2026 reported impressions and clicks, without conversion tracking, demographic insights or purchase attribution. That description predates the broader self-serve rollout. Reporting capabilities may have expanded, so practitioners should inspect the current Ads Manager for live fields rather than treat the April account as the present specification. Even richer reporting would not automatically resolve unclicked discovery.

Work to measure these effects should start when the campaign launches. Branded search lift monitoring can show whether branded search volume rises during and after campaign flights, while direct traffic segmentation compares visits in those periods with the pre-flight baselines. Both are imperfect proxies for assisted discovery. A dedicated landing page for each campaign helps show what users do later when information about the click source is lost through redirects or paths that span more than one session. A post-conversion survey asking “How did you first hear about us?” adds self-reported discovery evidence where automated attribution is thin.

The strategic orientation for early adopters is to build a measurement framework that preserves distinctions between observed clicks, downstream behavior and inferred influence. Conversation-layer data can support context-hint refinement where it is genuinely available. The benefit is proprietary learning under unsettled measurement conventions.

Ad safety, brand suitability, and what responsible trafficking inside a conversational interface requires

OpenAI states that paid placements remain separate from ChatGPT’s answers and that the ad system and chat model run on separate infrastructure. Advertisers cannot influence what the model says. Users can also disable ad personalization, reducing the signals available for matching. OpenAI excludes accounts identified as belonging to users under 18 from ad delivery. Practitioners should document that platform-enforced audience exclusion in campaign records rather than treat it as a control they must implement manually.

Structural separation does not remove the need for contextual judgment. Even if its category is approved, a health insurer may not place an ad while a user describes symptoms or asks for a personal diagnosis. Financial advertisers likewise may not target conversations about financial vulnerability or distress. Context hints must clearly rule out those conversations. A misplaced ad in a sensitive exchange can feel more personal than a mismatched banner because it intrudes on a conversation in which the user has disclosed a need.

Safety settings require approval before launch and subsequent review of context hints and negative exclusions against current policy quarterly at minimum, and more frequently while the platform remains in active development. OpenAI enforces clear “Sponsored” labeling, but relevance and restraint remain essential to suitable placement. Perplexity’s exit illustrates why that distinction is consequential. Getting the specifications, hints, tracking and safety controls right helps keep ads from being rejected and earns the right to make good use of a channel that users trust.