Most providers selling ChatGPT ads management will not tell you this upfront: there is no search term report to mine. An ad inside ChatGPT is matched to a conversation, not a typed keyword. That one difference decides which levers a manager can touch.

Say you sell cushioned everyday running shoes for beginners training for their first 5K. On Google, you would buy a phrase, inspect the searches that triggered your ad, and cut the waste. In ChatGPT, an orchestrator model reads the multi-turn conversation, turns relevant parts into queries, then runs a separate auction for a sponsored card below the answer. You never see the chat that triggered your ad. OpenAI does not share conversations and gives advertisers aggregates such as impressions, clicks, spend, and conversions. I used to tell clients the search term report was where the money was made. Here, there is no report. Management has to happen somewhere else.

First principle: the conversation is the query

On Google, the query is one line. In ChatGPT, it may be the last five minutes of talk. If someone spends four turns explaining that they run a small agency and need free guest accounts for clients, then asks about project software, the earlier details can shape which ads are considered. Earlier turns matter to matching, not just the last prompt.

Our shoe buyer does not have to type best running shoes. She might say she has sore knees, runs twice a week after work, and wants to finish her first 5K. The system interprets that history, retrieves eligible ads, and runs a relevance-weighted second-price auction for a sponsored card below the answer. The sponsored unit does not feed back into the answer text.

That is the first constraint on the job. You cannot pick a keyword and demand the corresponding conversation. You can make the ad easier to recognize as a fit for that buyer, then measure what happens after she clicks.

Second principle: the advertiser supplies clues, not a keyword list

The system needs material to reason over. You supply a title, copy, image, and landing page, plus up to 2,000 context hints per ad group describing fit. You also set location and platform controls, any custom audiences, a budget, a bid, and a conversion event. “Cushioned everyday running shoes for beginners training for a first 5K” says more about the intended buyer than “running shoes”.

Those inputs do different jobs:

  • Creative and page tell the buyer what you offer and give the system relevance signals.
  • Context hints describe the situations in which the ad fits.
  • Eligibility and spend controls determine where the ad can compete and what you are willing to pay.

A hint is not an exact-match keyword wearing a longer coat. Hints guide relevance but do not guarantee delivery. For the shoe seller, “beginner planning a first 5K who wants a cushioned daily trainer” describes a buyer. A stack of slight variations on “running shoes” mostly describes a category.

Write hints like buyer descriptions. If the system must infer who the shoe is for, do not make it infer from two generic words.

Third principle: feedback arrives after the click, not from the chat

A manager can see impressions, clicks, spend, CTR, average CPC or CPM, and conversions if tracking works. The manager cannot inspect the conversation behind an impression because OpenAI does not share it with advertisers. There is no search terms tab to sort on Monday morning, no auction insights panel, and no impression share column to argue over.

I spent years starting Mondays with a search term report and a batch of negatives. That habit does not transfer. If our beginner shoe ad gets clicks but few purchases, the manager can inspect the ad group, promise, landing page, and purchase path. The manager cannot point to a list of triggering chats and announce that marathon runners caused the problem.

The steering wheel is your inputs plus the post-click trail. A good manager changes one with a reason and watches the other for evidence that the change helped.

The levers a manager can actually use

Context hints and eligibility: define fit, then rule out bad fit

Targeting starts with eligibility. You can set geography down to state, DMA, or zip, with exclusions; choose platform and schedule; and upload custom audiences if you have them. Lists can use emails, E.164-format phone numbers, or Google ad IDs in CSV or TXT files up to 500MB, hashed with SHA-256. Inclusion or bid adjustment requires 25,000 matched users, with 100,000 recommended; any size list can exclude. Ad-group bid multipliers run from 0.1x to 10x, and exclusion beats inclusion.

Then describe the fit. For our shoe seller: beginner training for a first 5K; evening runner who wants a cushioned everyday shoe; shopper looking for cushioning rather than carbon plates. Those are descriptions of the same plausible buyer, not bids on three secret keywords. If an ad group attracts clicks without purchases, revise its hints alongside the creative and page. The report will not tell you which individual conversation went wrong.

On the exclusion side, Negative Phrases let eligible advertisers exclude unsuitable conversational contexts. DoubleVerify and IAS are building independent suitability checks without seeing private chats. An exclusion says where you refuse to show; it does not force placement somewhere better.

Review hints and exclusions weekly. Change geo and audiences when the business calls for it, not because an account needs something to happen on Tuesday.

Creative and offer: make the small card earn its click

The card sits below the answer. It does not borrow authority from the answer text, so its promise has to stand on its own. For the shoe seller, I would test a title that names the buyer and offer plainly: a cushioned daily trainer for new runners, with a first 5K plan included. Clever copy can wait until the basic offer is clear.

Change one promise at a time and read CTR against conversion rate. A higher CTR is not a win if the extra clicks do not buy. If the ad promises a beginner shoe while the page leads with products for every kind of runner, the mismatch starts before anyone reaches the cart.

Rewrite creative when click quality is wrong, not when you are tired of reading it.

Landing page and tracking: check where the money goes

OpenAI counts the landing page alongside copy, hints, and conversation when judging fit. A generic page for all running shoes gives the system less to work with than a page built around the beginner 5K offer. The headline can repeat that promise; the first screen can address cushioning and sizing; the path to add to cart or start the plan can be clear.

Then check measurement before judging the ad. In one practitioner test, $1,000 to $1,200 spent on agency terms brought 92 clicks at about $13 average CPC, roughly 1% CTR, and zero conversions. A UTM bug also sent GA4 attribution into direct traffic. Those are two problems to separate: the offer may not have converted, and the tracking could not cleanly tell the story.

If the pixel or Conversions API event fires incorrectly, bid decisions rest on bad counts. Get a working click-to-purchase trail before calling an ad group a success or failure.

Budget and bids: pay for a test you can read

Money affects eligibility and pace; it does not buy the exact conversation you want. Objectives include Views billed on CPM, Clicks billed on CPC, and Conversions using oCPC or oCPM, where billing is still per click or impression rather than per conversion. The auction is relevance-weighted and second-price. The US minimum is $25 per day averaged over a Sunday-to-Saturday week, with spend of up to twice that amount on a single day.

OpenAI suggests starting around $3 to $5 max CPC. Early observed CPCs range from $3 to $22; a pilot $60 CPM carried a $200,000 minimum. In one practitioner test, a $3 bid was flagged as potentially unable to deliver, while 92 clicks cost about $13 each on average. A starting bid is not a promise about the price of a click.

For our shoe seller, I would start with one beginner ad group, a tight daily cap, a Clicks objective, and a max CPC the business can afford if only one in fifty clicks buys. Then I would watch pacing by day, not hour. I would change the bid when delivery stalls or CPA drifts, not because Tuesday feels slow.

Set a budget you can hold long enough to read the post-click numbers. The draft test window is three weeks, not three anxious afternoons.

What is missing from the bench

If you come from Google Ads, you will reach for familiar controls. The documented ChatGPT ad controls do not include keywords, match types, negative keyword lists, age and gender targeting, interest buckets, lookalikes, or site retargeting. I kept looking for the negative-match version of a context hint my first week. It was not there.

Negative Phrases are not that missing match type. They exclude unsuitable conversational contexts for eligible advertisers. They do not reveal the chats you appeared beside, and they do not secure the conversations you prefer. Likewise, because you cannot see the chat or consult auction insights and impression share, you cannot diagnose lost impression share to rank or budget the way you would on Search.

If a provider promises keyword-level control inside ChatGPT, they are describing a different product. Ask which actual setting they intend to change.

Why this changes the manager’s week

On Google Ads, the weekly routine was visible: pull the search term report, add negatives, split an ad group, adjust bids. I could show you fifty line items and call it management. That routine is gone here. A quarterly call about the opportunity, followed by advice to test context hints, is not running an account. It is narrating one.

Real management looks smaller and more repetitive. Rewrite hints when an ad group gets clicks without purchases. Tighten the shoe page so it delivers the 5K promise in the card. Check that the purchase event still fires after a site update. Move budget when one ad group produces two checkouts from forty clicks and another produces none from sixty. Because conversation data does not come back, that post-click trail carries the decision.

None of those moves proves which sentence in a private chat made someone click. They do give the manager an edit, a reason for it, and a result to check. Ask to see that chain in the change log. A slide about relevance is not a substitute.

Five questions that expose a monthly strategy call

Ask these on the sales call and listen for edits, dates, and counts. A consultant may have useful ideas. A manager should also be able to show what changed in an account and what happened afterward.

  1. Can you show me a week of changes from a live ChatGPT account? Look for rewritten hints, a Negative Phrase, a creative or page change, or a bid or budget move, with the reason recorded. A strategy deck does not answer the question.
  2. Which contexts would you exclude for my business, and how? The answer should distinguish Negative Phrases from geo and audience exclusions. Anyone who says an exclusion buys preferred placement has misunderstood the lever.
  3. How do you connect a click to money without seeing the chat? Ask about UTM naming, pixel and Conversions API checks, and a purchase or qualified-pipeline report tied to the ad group. CTR and CPC alone do not pay the bill.
  4. What have you rewritten because click quality was wrong? Ask for a before-and-after hint or headline aimed at a specific buyer slice, such as beginner 5K shoppers rather than the wider running-shoe category. Our look at how conversation context affects ChatGPT ad matching explains why this distinction matters.
  5. Who owns the number after the click? Name the person accountable for CPA or ROAS and ask when they would pause spend if the page stops converting. If the provider sells advice rather than edits, compare the work honestly with the course, consultant, and done-for-you models. Price 90 days of advice against 90 days of account work.

Run the model on a different business

Say you own a dental clinic buying emergency appointments. In week one, you launch two hint groups: cracked tooth after hours and routine cleaning. By Friday, the cleaning group has 40 clicks and no booked exams; the emergency group has 25 clicks and three bookings.

Start with what you know. The emergency group has produced bookings; the cleaning group has not. You do not know which private conversations produced the cleaning clicks. A manager checks the booking event, inspects each group’s promise and landing page, and rewrites the cleaning hints to describe routine exams more precisely rather than claiming to have found a bad search term. The manager can exclude job-seeker and DIY contexts where those placements do not fit, make the emergency page’s after-hours availability and booking button clear, and move half the budget toward the group producing appointments. Then the manager logs the edits and watches whether bookings follow.

That is the same model as the shoe account: describe the fit, set eligibility, make the card and page agree, verify the conversion, and allocate spend against the result. It produces work you can inspect without pretending anyone has read the patient’s chat.

If a provider cannot show you that kind of week, they are not managing. They are observing. groas runs this as done-for-you Search Ads covering Google, Bing, and ChatGPT for one flat fee, with hints, creative, pages, bids, and tracking owned by one team accountable for cost per acquisition instead of a monthly memo. Ask for the log before you sign. Pay for edits that move revenue, not slides that describe relevance.