Week 1 · ChatGPT Ads fundamentals — the auction, the copy, the attribution
A free, weekly training series for Indian healthcare marketers learning to run ChatGPT Ads. No fluff, no platform-vendor spin — just the mechanics, worked through the way ICG runs them for clients.
What you'll learn this week
Week 1 is the foundation the rest of this series builds on. By the end of this module you should be able to explain, without notes, how a ChatGPT Ads auction decides which advertiser gets to speak inside a conversation, why the copy that wins that auction reads nothing like a Google headline or a Meta carousel caption, and how conversion gets measured when there is no single "click" to point at. If you have run Google Ads or Meta Ads for a hospital, clinic chain, or diagnostic network before, most of the vocabulary here will feel familiar even where the mechanics differ underneath.
We are assuming you already know the basics of paid media — what a bid is, what an impression is, what a conversion event is — because this series is not a "what is digital marketing" primer. It is a practitioner's walkthrough for someone who is going to be spending real rupees on this channel within the next few weeks and needs to understand what they are buying. If that description does not fit you yet, our insights library has broader introductions to healthcare digital marketing that are a gentler starting point.
This week has three core concepts, a worked example pulled from an anonymised ICG engagement, a short exercise you should actually do before Week 2 lands, and a list of the mistakes we see healthcare marketers make in their first month on this channel. Budget about an hour, most of it on the exercise.
Core concept 1 — the intent-first auction
Every paid channel you have used runs an auction of some kind, but the trigger for that auction is different in each case, and the trigger is the thing to understand first because it determines everything downstream — who you are bidding against, what you are bidding on, and what "relevance" even means in that system.
Google Ads triggers an auction off a typed keyword. The user types a query, Google matches it against advertiser keyword lists using match types and quality signals, and the auction runs in the milliseconds before the results page renders. The keyword is the unit of intent. Meta Ads, by contrast, does not wait for an explicit query at all — it triggers off inferred interest and behavioural signals, assembled from what a user has done across the platform, and the auction runs as part of deciding what fills a scroll position. The unit of intent there is a profile, not a query.
ChatGPT Ads triggers off something closer to Google's model than Meta's, but with an important difference: the trigger is not a keyword match against a typed string, it is an intent classification against a live, multi-turn conversation. A user does not type "best fertility clinic Gurgaon" as an isolated query — they might ask "I've been trying to conceive for two years and my doctor mentioned IVF, what should I be asking about clinics near me," and the assistant has to classify the commercial intent buried inside that longer, more natural sentence before an auction is even eligible to run. This is the single biggest shift healthcare marketers need to internalise: your targeting unit is no longer a keyword list, it is a described intent, and your ad has to earn eligibility against the meaning of what was asked, not the literal words used.
That has a direct consequence for how you plan bids. In Google Ads, you can build a keyword list of a few hundred terms and feel reasonably confident you have covered the space. In ChatGPT Ads, the same underlying intent — "help me find a fertility clinic I can trust" — can arrive in dozens of phrasings that share no keywords in common at all. This is why ICG's approach for clients starts from intent clusters rather than keyword lists: we map the handful of distinct commercial intents a speciality generates (diagnosis anxiety, second-opinion seeking, cost comparison, location-constrained search, insurance-driven search) and build bidding strategy and creative around those clusters, trusting the platform's own classification to handle phrasing variance.
The auction mechanics themselves — a bid, a relevance or quality-style score, an effective rank — are recognisably descended from the search-auction model most performance marketers already know. What is genuinely new is the input the auction is evaluating. Get comfortable with that distinction before Week 2, where we go deep on how to actually structure targeting around it.
Core concept 2 — conversational ad copy is not headline copy
The second thing every healthcare marketer gets wrong in their first campaign is treating ChatGPT Ads copywriting as a shorter, chattier version of Google Ads copywriting. It is not a smaller version of the same skill — it is a genuinely different skill, and the failure mode is copy that reads like an ad dropped into a conversation rather than copy that reads like a natural, useful continuation of it.
Think about the mechanics of where your ad copy actually appears. A Google headline sits on a results page next to nine other results, competing purely on the strength of the words themselves — no context, no prior exchange, just a headline and a description trying to win a glance. Meta ad copy sits inside a feed the user is scrolling past at speed, competing against friends' photos and other native content for a fraction of a second of attention. ChatGPT Ads copy, by contrast, is inserted as a response — or part of a response — within a conversation the user is actively engaged in and reading closely, because they asked a real question and are waiting for a real answer. The attention is not scarce in the same way. What is scarce is trust: if the copy reads as an ad, it breaks the conversational frame the user is in, and that break is far more damaging here than a skipped headline is anywhere else.
Practically, this means the copy that performs is copy that answers the actual question first and identifies itself as an option second — not copy that leads with a brand name and a superlative. "Ichelon-managed clinics typically get you a fertility specialist consultation within 48 hours, with transparent per-cycle costs shared before you visit" reads as useful information a user asked for. "India's #1 fertility clinic network — book now!" reads as an ad interrupting a conversation, and users who are mid-conversation with an assistant they trust are unusually quick to discount that kind of language. It is worth repeating: ASCI guidelines already prohibit unverified superlative claims in Indian healthcare advertising regardless of platform, so this is not just a performance argument, it is a compliance one too — more on that in Week 5.
The other structural difference is length and specificity. Google headlines are constrained to roughly thirty characters and have to compress meaning aggressively. ChatGPT Ads copy has room to be a full, specific sentence or two, and that room should be spent on specificity — a number, a timeframe, a named service — rather than on more adjectives. Vague copy that would blend into a Google results page blends in even more inside a conversation, where the user's bar for "worth reading" is set by the quality of the assistant's own answer immediately above it.
Core concept 3 — attribution without a single click
The third foundational idea is attribution, and it is the one most experienced performance marketers underestimate going in, because it looks familiar and behaves differently.
In Google or Meta, a conversion is usually anchored to a single, discrete event: a click on an ad, followed by a landing-page action within an attribution window. The click is the anchor point everything else hangs off. ChatGPT Ads does not have an equivalent single anchor, because the user's path to action is a conversation, not a click — they might see a sponsored response, ask two more follow-up questions, get a second sponsored mention three turns later, and only then tap through. Which of those touches gets the credit?
The model the platform uses is what this series calls conversation-completion attribution: credit is assigned when the conversational thread reaches a defined outcome state — a click-through to a landing page, a saved or bookmarked reference, a form submission reached through the thread — and the model looks back across the turns of that conversation to weight the touches that contributed, rather than crediting only the last thing the user saw. It is conceptually closer to a multi-touch or data-driven attribution model than to last-click, and marketers who have only ever worked in last-click environments should expect their instincts about "what's working" to need recalibrating in the first few weeks of data.
This matters enormously for how you read early campaign data. If you are checking dashboard results expecting a Google-style click-through-then-convert pattern, you will misjudge which parts of your creative and targeting are actually earning credit. ICG's own GA4 instrumentation treats AI Assistant referrals as a distinct acquisition channel rather than folding it into generic referral traffic, specifically because the shape of the conversion path is different enough to distort blended reporting if you don't separate it out. We will build out full measurement setup in Week 4 — for now, the concept to hold onto is this: attribution here rewards being useful across a conversation, not just being clicked once.
Worked example — a real ICG engagement pattern
Here is how these three concepts played out for an anonymised ICG client — a multi-city diagnostics and speciality-consultation network that piloted ChatGPT Ads alongside its existing Google and Meta spend earlier this year.
The client's instinct going in was to port their existing Google keyword list directly into the new channel — a few hundred terms built around speciality names and city modifiers, the kind of list that had served their Search account well for years. We advised against a straight port and instead spent the first two weeks of the pilot mapping intent clusters the way described in Core Concept 1: not "cardiologist Gurgaon" as a keyword, but "help me understand if my symptoms need a cardiologist or just a check-up" as an intent, alongside four or five sibling variants of the same underlying question. That reframing changed the creative brief entirely — instead of writing ad variants around keyword insertion, the copy team wrote around answering the underlying anxiety first, speciality-network as the answer second.
The copy itself went through two rounds. The first round, written by an in-house team member new to the channel, read almost exactly like their existing Google Ads copy — compressed, superlative-leaning, brand-first. Early impression-to-engagement rates were noticeably softer than the client's Search benchmarks. The second round, rewritten around Core Concept 2's answer-first structure, led with specific, checkable information — typical wait times for a first consultation, whether teleconsultation was available same-day, what a diagnostic panel typically costs before insurance — and engagement improved measurably within the first data-stable week.
On attribution, the client's finance team initially flagged the channel as underperforming because last-click-style reporting showed a thin direct-conversion count compared to Google. Once we pulled conversation-completion data and cross-referenced it against the GA4 AI Assistant channel — which, across ICG's broader client base, shows a key-event rate roughly double organic search and around seventeen times direct traffic — the picture reversed: the channel was contributing meaningfully to multi-turn journeys that completed as leads days later, credit that last-click reporting had simply not been built to see. The lesson the client took away, and the one we'd underline for you here, is that judging this channel by Google-shaped metrics in month one is the fastest way to abandon something that is actually working.
The pattern to take from this: don't port your keyword list, don't port your ad copy voice, and don't port your attribution expectations. Each of the three needs its own first-principles pass, and doing that pass properly in week one saves weeks of second-guessing later.
Exercise for you to do this week
Before Week 2: map three intent clusters for your own speciality or service line
Pick the single speciality or service you would launch ChatGPT Ads for first — the one with the clearest existing demand signal from your Google Ads or GSC data. Then, without opening a keyword-planning tool, write down three distinct underlying intents a prospective patient's family member or a corporate-wellness buyer might have when they'd plausibly ask an AI assistant about it. Not keywords — full sentences, the way a real person would actually phrase the question in a conversation.
- For each intent, write two full-sentence versions of how a real person might phrase it to an assistant — vary the framing, not just the wording, so you're testing whether you've actually captured the underlying need rather than one phrasing of it.
- For each intent, draft one sentence of ad copy that answers the underlying question with a specific, checkable fact — a timeframe, a number, a named service feature — before it mentions your brand at all.
- Flag, in a single line per intent, what a conversation-completion outcome would plausibly look like for that intent — a click-through, a saved reference, a form fill three turns later — so you have a hypothesis to test once Week 4 covers measurement setup.
Keep this document. Weeks 2 and 3 build directly on it, refining the targeting and rewriting the copy drafts you produce here, so the more honestly you do this exercise now, the less rework you'll do in two weeks.
Common failure modes to avoid
Porting a Google keyword list unchanged. It will underperform because it is answering the wrong question — the platform is classifying intent, not matching strings, and a keyword list optimised for string-matching misses the phrasing variance intent classification is built to catch.
Writing brand-first, superlative-led copy. It breaks the conversational trust the channel depends on, and in Indian healthcare advertising it also risks the same ASCI compliance issues that apply everywhere else — covered fully in Week 5.
Judging month-one performance on last-click logic. Conversation-completion attribution surfaces value on a different timeline and through different touches; pulling the wrong report in week two is the single most common reason healthcare marketers prematurely kill a pilot that was actually working.
Hold onto your Week 1 exercise document — you'll need it open when Week 2 covers targeting structure in depth.
