You cannot pay to be the answer inside ChatGPT. You can pay to sit beside it. I keep meeting owners who think those are the same purchase. They are not, and the difference decides where your money should go.

Say a homeowner types: my garage door spring snapped, can I fix it myself? ChatGPT writes an answer. Separately, an ads system checks whether that conversation calls for a relevant ad and may place a labelled card below the response. OpenAI describes the ad stack as separate models and code, running in its own service. If that service is slow, the answer still goes out on time without an ad.

The answer and the ad are different products. Hold on to that fact and the rest gets easier: what an advertiser can control, why a keyword list will not run this campaign, why a landing page still has to close the gap, and why a competitor can be cited above your paid card. I will build those layers with the same garage door conversation throughout.

Layer 1: the answer is not the ad inventory

In our garage door conversation, the model answers using what it knows and whatever material it fetches that looks useful: perhaps manufacturer information, a safety guide or a local discussion. Its answer and citations are not awarded to the highest bidder. The ads system works separately from a brief summary of the conversation, using that context for its own retrieval, filtering and auction. OpenAI says the orchestrator does not pick the winner or set the price; a separate auction tool does. Returning no ad is a normal successful outcome.

The distinction is visible on the screen. A paid unit is a sponsored card below the response, set apart from the answer and carrying an advertiser name, title, short description, landing page and image. If the homeowner sees a warning about torsion springs in the answer and a repair company underneath it, the company bought the card. It did not write the warning.

I tell clients to stop asking for sponsored answers here. The plumbing does not offer that purchase. Before thinking about bids or copy, decide which of two outcomes you mean: a card beside the answer, or a citation inside it. They call for different work.

Layer 2: the conversation matches the card

Back to the homeowner with the broken spring. There is no keyword box where you bid on garage door repair near me. Instead, you write context hints at the ad-group level: plain-language descriptions of conversations where your ad could help. A hint might describe a homeowner with a broken torsion spring asking whether DIY is safe and what same-day repair costs.

The system weighs those hints alongside your title and copy, your landing page, the current conversation’s intent, and past-chat signals when personalization is on. Hints help an ad get retrieved; they are not exact-match targets or delivery guarantees. I used to tell clients to start every search campaign with keywords. I was wrong for this surface. Think in situations, not strings.

You still make familiar advertising decisions. You choose the hint, creative, landing page, objective and bid. OpenAI lists Views billed by CPM, Clicks billed by CPC, and Conversions billed per click or impression rather than per conversion. There is a $25 daily minimum, and OpenAI recommends a $3 to $5 starting max CPC in its relevance-weighted second-price auction. None of those controls lets you insert a sentence into the model’s reply. They help determine whether your separate card appears.

The feedback loop is different, too. Reporting is aggregated for privacy: impressions, clicks, spend, CTR, average CPC or CPM, and conversions if your pixel is set. You do not get the conversation or a prompt-level report. You do not get a view by hint. If you expected a search term report to mine after getting matched to a conversation, that missing loop is the first operating constraint to accept.

For the garage door advertiser, this means writing a hint broad enough to describe a real situation and a card specific enough to offer the next step. Then judge what the reporting actually shows. Do not pretend an aggregate click total tells you which exact homeowner question worked.

Layer 3: the landing page has to finish the job

Suppose the homeowner clicks the sponsored card. ChatGPT does not book the repair inside the conversation. The visitor goes to your site with the answer still fresh in mind: torsion springs can injure you, so DIY may be a bad idea. A page that opens with a generic company history makes that visitor work to find the point. A page that clearly shows same-day service, service area, price range and proof meets the situation the card just addressed.

The gap between a click and a customer is easy to underestimate. In a June 2026 test of 3 campaigns, 8 ad groups and 48 ads, 97,131 impressions produced 1,263 clicks at a 1.30% CTR but almost no sign-ups. Two ad groups took half the impressions. Delivery followed conversational volume, not an advertiser’s tidy division of budget across ad groups. More cards on the screen did not automatically mean more conversions.

One report of advertiser discussions quotes a practitioner whose ChatGPT visitors stayed about 5 seconds on average. Others described cheaper local-service clicks; others saw weak leads. Those accounts are not a universal verdict on the channel. They are a reminder of what the mechanism can and cannot do. Matching puts a relevant offer near a moment of interest. The page and offer still have to turn that interest into action. Fix those before raising the bid.

Case 1: your competitor gets cited above your ad

Now suppose the answer links to Apex Garage Door’s spring replacement guide. Your same-day repair card appears beneath it. It feels as though the system gave Apex the valuable part and charged you for the leftovers. The separation from Layer 1 explains it: the citation and the paid placement were not prizes in the same auction.

A study of more than 50,000 commercial prompts found ads in about 26% of answers, yet the advertiser’s own domain was cited in only 3.63% of those placements. The figures do not tell you which individual ad to buy. They show why paying for a card is not a shortcut to being linked in the answer. If you want to understand why ChatGPT cites your competitor, inspect the material it can use for that answer, not just your bid.

I would not pause a relevant ad solely because the citation went elsewhere. The card still gives you a buyable route to the homeowner; Apex’s link does not offer your same-day repair. Work on the citation separately through pages and corroboration the model can fetch. Keep the paid slot doing its job while you work on earning the link.

Case 2: you get cited, so should you stop paying?

Reverse the scene. The answer links to your spring safety guide, and your sponsored card appears below it. Is the card waste? Not necessarily. The citation gives the reader information and a reason to trust your material. The card gives you control over a different message: same-day spring replacement, your service area and a phone number. One does not make the other permanent. A different phrasing of the homeowner’s question may produce a different citation.

They also reach different portions of the audience. Ads are shown to logged-in adults on Free and Go plans and to logged-out traffic, not to Plus, Pro, Business, Enterprise or Edu users. A B2B practitioner quoted by WordStream called that a significant drawback for B2B. If your buyer uses a paid work seat, the card cannot reach that screen, while an organic citation still can.

Measure the two appearances separately. A citation is not an impression you bought, and an ad click is not proof the answer recommended you. Keep both only when each has a job you can name.

Case 3: the budget is small and the sale takes time

Say you sell a $20k commercial door system and can spend $1,500 a month. The same model predicts a harder paid test than the homeowner’s urgent repair. Conversations may cluster unevenly: in that June 2026 test, two ad groups took half the impressions. Early auction prices may look inviting, with anecdotes of $25 CPMs in April falling to about $10 CPM alongside roughly 0.5% CTR. But cheap reach does not shorten a 90-day sale.

The buyer’s question may also have several parts. As I have argued about how ChatGPT handles a buyer’s question, one hint will not cover every concern in a committee purchase. A thin budget spread across too many situations may teach you little. This channel will not work for everyone. For a long-cycle seller with little room to test, I would skip broad paid experimentation and put the money into useful pages that have a chance of being cited first.

That is not an argument against ads. It follows from what an ad buys: a chance to present a card beside one conversation, not ownership of every question a buyer asks before signing.

One screen requires two kinds of work

The layers now add up. Paid placement means writing hints and creative, setting bids and budgets, watching delivery, and making the landing page useful after the click. Organic citation means publishing pages the model can fetch, with information it can use and corroboration beyond the page itself. The two jobs meet on a screen, but they should not disappear into one performance number.

I have watched teams give both jobs to a Google Ads freelancer and wonder why the citation never comes. I have also seen them give both to a content writer and wonder why the card burns cash. What the sales deck calls an integrated search function often means two people forwarding each other reports. Split the measurement even if one team owns the strategy.

Other search surfaces make the distinction easier to see. In Google AI Overviews, Search and Shopping ads can appear above, below and, since October 2024, woven into the overview under a Sponsored label. Perplexity tested sponsored follow-up questions in November 2024, then phased ads out and said in February 2026 that it had no plans to bring them back. ChatGPT’s card sits below a separate answer. The placements differ; in this garage door example, paying for visibility still does not amount to writing the answer.

Why I would run the surfaces from one context

Our homeowner does not have to stay in ChatGPT. She can ask there whether DIY is safe, skim a Google AI Overview about repair cost, then click a Bing ad because that is the browser she uses at work. Each surface has its own placement rules. She carries the same problem across them: a broken spring and a need for someone to fix it.

Run those surfaces in isolation and the failure is familiar from old PPC accounts. One operator buys ChatGPT clicks with a generic hint. Another buys Google queries using old keywords. Nobody owns the pages that might earn a citation. Nobody checks whether the landing page answers the concern that started the conversation. Spend rises; each report explains its own corner of the screen.

I would rather give all three surfaces the same business context while keeping their results distinct. Search Ads at groas brings Google Ads, Bing Ads and ChatGPT Ads under one flat $1,499 monthly fee, with bids, budgets and landing pages worked around the clock and a named strategist accountable for the result. The goals informing a ChatGPT hint can also inform the Google structure; the work on useful pages can support the organic side. One team can ask whether the card earned a useful click and whether the answer earned a citation, without letting either number hide the other.

Run the model on the next question

The homeowner follows up: who can fix a torsion spring today for under $400 near me? Start with the answer side. Pages showing same-day service, service area and spring pricing give the model material relevant to that question. Whether your page is cited depends on what the model uses for its answer. Your bid cannot settle it.

Now take the paid side. The ads system uses a summary of the conversation to retrieve and evaluate a possible card. A hint about urgent spring repair may put your offer in contention; a relevant card can appear below the answer even if another company gets cited. After the click, your page must show the area, hours, price information and a clear path to book. A generic page makes the homeowner hunt for all four.

That is the whole model applied to a new question: earn the answer, buy the slot, and measure the two separately. If the citation names someone else, the paid card can still offer the repair. If the card never appears, a useful page can still be cited. Neither route excuses a weak landing page. The homeowner needs a fixed door, not an explanation of your reporting structure.