Pre-Flight QA Checklist for AI Ad Campaigns

Real-time prompt intent analysis means reading a live conversation, not a search query, to decide which ad belongs next to it. A natural-language context hint gets matched against the full trajectory of a dialogue in a relevance-weighted auction, and if you get that wrong, your conversational AI campaign will likely underperform. Here you will find the pre-flight checks you need before the first impression serves, grouped by where legacy QA habits break first.

Why legacy QA processes cannot cover conversational AI advertising

A search campaign's pre-flight list is short and static: keywords, match types, bid caps, landing page load speed. Every item on that list describes a fixed input feeding a deterministic system, and a QA reviewer can check each one in isolation because none of them depend on what happens next in the user's session.

Conversational AI advertising breaks that model at the root. The targeting surface is a live, multi-turn dialogue rather than a query typed into a box, so the system serving the ad has to read what the user asked several exchanges ago, how the assistant responded, and where the conversation looks like it's heading before it can decide an ad belongs there. The primary targeting control reflects that shift: instead of a keyword list, advertisers write a 280-character natural-language context hint, and teams that treat it like a comma-separated topic list get weak matching, consistently and predictably, because a topic list describes a subject while the system needs a situation. Placement itself works on different physics too. A relevance-weighted, second-price auction means a precise advertiser with a small budget can beat a vague advertiser with a large one, and no legacy checklist has a line for that, because no legacy channel rewards precision over spend so directly. Layer onto that a set of inventory rules that differ sharply by platform: ChatGPT ads reach only logged-in adults on the Free and Go tiers, while every paid tier stays ad-free, and that constraint changes who a B2B campaign actually reaches no matter how good the targeting is.

None of this is a matter of degree. It calls for a checklist built around entirely new categories of checks: platform eligibility, context hint construction, creative fit for a conversational slot, disclosure rules still being written, and measurement that accounts for a structurally different attribution gap. Each of those gets its own section below.

Platform eligibility and inventory rules to confirm before anything else

Before any question about creative quality, the first question a campaign manager has to answer is whether the campaign can run on a given platform at all, and in front of whom. The answers differ enough across platforms that assuming eligibility rather than confirming it is the fastest way to invalidate every check that follows.

ChatGPT ads reach only logged-in adult users on the Free and Go tiers; Plus, Pro, Business, Enterprise, and Education all remain ad-free per OpenAI's own documentation. Self-serve access opened May 5, 2026, with spend minimums removed, campaign budgets starting at a low daily minimum, and a choice between a Reach objective billed on CPM or a Clicks objective billed on CPC. If you run a B2B play aimed at heavy ChatGPT users, check what tier those users sit on, because if your audience has moved to a paid plan, your campaign may reach almost nobody, no matter how well it's built. Microsoft Copilot works on the opposite principle: every eligible campaign and ad type is automatically opted in, advertisers cannot opt out, no specific placement is guaranteed, and ads get assembled from assets already sitting in the account. Audience eligibility runs to every Copilot user above 13 (above 18 in some regions), but users between 13 and 18 see only non-personalized ads. Claude sells no ad placement at all, which makes it a non-starter as a paid surface regardless of what a brand's audience looks like there.

Vertical exclusions add further restrictions on top of these eligibility rules. Finance and healthcare campaigns face restrictions on both ChatGPT and Google AI Overviews, and because those restrictions don't match each other, you can't assume clearance on one platform carries over to the other. For any brand already running active campaigns through Google or Microsoft, there's a more immediate task than planning new ones: those campaigns may already be surfacing inside AI-generated answers without anyone having decided that on purpose, and auditing what's already live matters more at this stage than planning what comes next.

How to audit context hints before they go live

The context hint is the highest-leverage input in the entire campaign, and it's also the one most teams get wrong first, because it looks like a keyword field and isn't one. A keyword records what string a person typed. A context hint has to describe what that person is trying to accomplish and whether an offer is actually useful to them at that moment, which makes writing one a creative skill as much as a targeting decision.

The difference shows up clearly in practice. A weak hint reads like a repurposed keyword list: "project management software, team collaboration, task tracking, enterprise tools." That string describes a product category, and a system built to match conversational situations has little to grab onto in it. A stronger hint describes the conversation itself: "A user comparing project management tools because their current system can't scale past 50 people, weighing cost against migration effort." That version gives the matching system a situation, a stage of decision-making, and a constraint, which is the kind of signal the auction can actually weigh against a live dialogue. A user deep in a conversation about scaling pain might never type the product category name at all, so a hint built around the underlying problem catches that conversation anyway, while a hint built around the category name alone misses it.

Four checks belong in a pre-flight review here. First, read the hint back and ask whether it describes a situation, a problem, a stage of decision-making, rather than a topic; if it reads like a list of nouns separated by commas, it needs to be rewritten before anything else happens. Second, check the hint against the ad copy and landing page and ask whether someone mid-conversation, in the exact situation the hint describes, would find the ad useful rather than intrusive, because that's the standard the auction is optimizing toward. Third, confirm the hint is calibrated to the right depth of conversation: conversational targeting can reach users four to six turns deep into a research exchange, and a hint written only for an opening query will miss everything that happens after it. Fourth, check that the hint sits within the 280-character cap and carries no keyword-list formatting, because that formatting tells the system it was built for a different kind of targeting. Because placement runs on a relevance-weighted, second-price auction, you can't just ask whether the bid is high enough; you have to ask whether the hint is precise enough to win on relevance even against a bigger budget. Conversational AI advertising platforms that read the full trajectory of a dialogue in real time, rather than matching against a single query, have made this difference concrete: a hint written to describe a user's underlying problem rather than the topic they happened to type is what unlocks accurate matching in this channel, and a poorly built hint undermines even a campaign with strong creative and solid budget behind it.

Creative and copy checks specific to conversational ad formats

Creative in this channel has to work next to an answer the user already trusts, and that changes the bar it has to clear. A ChatGPT ad renders as a labeled sponsored card below the assistant's response, and it's never woven into the answer text itself, because that separation is a deliberate design decision, not an accident of layout. It means the creative has to earn attention in a slot where the user's primary task already feels complete, which tolerates generic or interruptive copy far less than a search ad or a social placement does. Copy that mirrors the informational register of an AI response, measured and direct rather than promotional, tends to perform better in that context than copy built on the conventions of banner advertising.

The auction evaluates creative alongside the context hint and the landing page as a single package, so a pre-flight review has to check all three together for coherence, not treat the landing page as something you look at only after the click. If the hint targets someone partway through a research conversation about a scaling problem, but the copy assumes no prior context and the landing page pitches the product in generic terms, that mismatch lowers the relevance score the auction weighs and makes the campaign less competitive no matter what the bid is. A pre-flight check should confirm the tone reads as a natural continuation of a helpful exchange, that the copy reflects the specific conversational situation the hint describes, and that the landing page substantively matches that same conversational context. It should also confirm the creative doesn't touch excluded topics even tangentially, because contextual matching can trigger category filters an advertiser never anticipated.

The creative layer is the bottleneck in LLM advertising stacks, and volume, on-brand consistency, and pre-launch coherence checks across hint, copy, and landing page are what separate campaigns that win the auction from campaigns that lose it on relevance before bid even enters the picture.

Disclosure and labeling requirements that vary by platform and jurisdiction

Disclosure rules for conversational AI ads are not standardized across platforms, and they move faster than the legal and brand-safety review processes most advertisers already have in place, so you can't assume your existing compliance sign-off covers this channel. A ChatGPT ad carries a platform-enforced "sponsored" label on its card, but that enforcement covers the label itself, not the copy underneath it, so advertisers still need to verify their own creative doesn't blur the line between the ad and the assistant's actual answer. Google AI Overviews ads carry a "Sponsored" label meant to signal transparency to the user, and a pre-flight check should confirm that label is present and rendering correctly in preview rather than assuming it will appear as expected at launch.

Vertical exclusions carry a disclosure dimension of their own. Google AI Overviews excludes sensitive verticals, including finance and healthcare, from ad eligibility within the AI Overviews surface itself, though ads placed above or below the Overviews remain eligible across all verticals, and campaigns in categories adjacent to finance or healthcare should confirm they don't trip those filters by accident. Legal review needs to look specifically at whether creative is phrased in a way that could read as part of the assistant's own response, not as a separate sponsored message, because that kind of phrasing carries both platform policy risk and regulatory risk that a general ad-claims review won't necessarily catch. And because disclosure rules for AI-generated or AI-adjacent content are being written at different speeds in different markets, a campaign running across several geographies needs a jurisdiction-by-jurisdiction check rather than one global compliance sign-off covering all of them at once.

Measurement setup and attribution checks before the first impression serves

Measurement infrastructure for this channel has to be built differently from search or social before launch, not patched afterward, because the attribution gap native to conversational AI advertising will distort performance data starting with the very first impression served. A user can engage in a multi-turn research conversation about a problem, leave the chat without clicking anything, and convert later through an entirely different touchpoint, and last-click attribution is structurally unequipped to trace that path back to the original exchange. Last-click models were built for a world where the click and the credit sit close together in time; in a channel where the ad serves next to the fourth or fifth turn of a dialogue, the gap between exposure and conversion routinely grows wide enough that last-click assigns credit to whatever touchpoint happened to come last, which is frequently not the conversational exchange that actually drove the decision.

Each AI platform enforces its own reach and eligibility rules on top of this attribution gap: ChatGPT ads reach only Free and Go tier users, Microsoft Copilot automatically includes every eligible campaign without an opt-out, and Perplexity is currently closed to new advertisers. So if a campaign spans multiple AI surfaces, you have to translate platform-specific rules into one coherent measurement and buying strategy rather than treat each platform as a separate effort, because a fragmented approach tends to miss reach on one platform while it wastes spend on placements that were never eligible to begin with. A pre-flight measurement check should confirm that conversion windows are set wide enough to catch delayed conversions after a conversational exposure, that attribution models account for exposure that happens mid-dialogue rather than at a single click event, and that reporting separates performance by platform, since each one defines reach and eligibility differently. None of that fixes the conversation gap outright, but building for it at setup is what keeps the resulting data honest from the first impression onward.