There is no keyword to bid on in ChatGPT. If you start with a Google Ads export, you are building around a match decision this ad system does not make.

Say you run a home services company spending around $20k a month on Google. Someone types emergency water heater replacement cost, your keyword matches, and your ad enters the auction. In ChatGPT, that same buyer might ask about a strange smell, a broken heater, a tenant’s deadline, and only then the cost of replacement. The ad is matched to the conversation, not to one clean query. Everything else follows from that.

First, understand what a Google keyword controls

A Google search ad starts with a string the buyer typed. That query matches a keyword under a match type. Eligible ads enter an auction, where Ad Rank determines position. You set the keyword, the bid, and much of what affects Quality Score. You can inspect the search terms you receive and decide which ones to pursue or block.

Take our water heater buyer. They type tankless water heater installation cost Chicago. An exact-match keyword on that phrase may enter the auction; a broad-match keyword on water heater install may enter too. You can see the search term, choose how much you are willing to pay, and send the click to a page about tankless installs in Chicago.

That is the useful mental model for Search: one query, a match decision, then a landing-page choice. It is also the model you need to put down before buying ads in a chat.

A conversation changes what “intent” means

Now follow the same buyer through ChatGPT:

  1. “My basement smells like gas.”
  2. “How do I know if my water heater is dead?”
  3. “I rent this place out. Can my tenant wait till Monday?”
  4. “What does tankless replacement run in Chicago?”

If you look only at turn four, you see a price shopper. Across all four turns, you see a landlord dealing with a potentially urgent situation, deciding whether to repair or replace, and trying to understand a deadline. Intent lives in the accumulation, not just the last sentence. The gas concern also calls for safety guidance before anyone pitches an installation.

That is why exact match has no direct equivalent here. ChatGPT targeting uses natural-language descriptions of conversations where a product fits, rather than keyword targeting. You write what a deck might call context hints. I call them problem descriptions. You can supply up to 2,000 per ad group, but OpenAI says they are not exact-match keywords and do not guarantee delivery in specific conversations.

A hint describes a situation. The system decides whether a conversation fits it. I spent years treating a careful exact-match build as a form of control. For this surface, that control does not exist at the string level. Write the situation you want to serve, not the phrase you hope someone types.

The answer and the ad take separate routes

The matching process has a split worth understanding. A dedicated orchestrator model uses relevant parts of the conversation to find ads without requiring the same wording. It checks eligibility and fit, then sends candidates to a separate auction tool. The model writing the answer is not the model choosing the ad, and the ad system does not feed back into the answer. If an ad arrives too late, the answer still goes out and that ad does not appear.

The placement reflects that separation. Ads tested since February 9, 2026 appear on Free and Go plans in more than 60 countries, below the response and labeled sponsored. The buy uses a relevance-weighted second-price auction with a $25-per-day US minimum. Winning does not guarantee an impression: another fit check follows, and a winner can miss the delivery window.

That matters when a campaign reports few impressions. Low delivery can reflect weak relevance, timing, or both; the report will not neatly separate them. An auction win is not the same thing as a delivered ad. Do not respond to every low-impression hint by simply raising its bid.

Clay figure staring at an empty keyword box as conversation bubbles gather behind it

Replace the keyword export with problem descriptions

Suppose your top Google converter is tankless water heater installation cost Chicago at $48 a click. That phrase is valuable in Search because it catches a buyer who has already narrowed the job. It is a poor starting asset for a conversation in which the buyer is still working out whether the heater is broken, whether replacement makes sense, and what a tenant needs now.

Instead of exporting 500 keywords, I would start with 30 to 50 descriptions a service rep could recognize:

  • A landlord needs to deal with a failed water heater in a rental before a tenant’s deadline.
  • A homeowner is weighing a repair against a tankless upgrade.
  • Someone reports a gas smell near a water heater and needs to understand the immediate safety issue before discussing replacement.

Those are not three ways to say water heater install. They are three different situations. Each calls for a different next step, so each deserves its own ad angle and page.

There is a useful parallel on Microsoft’s Copilot surfaces, though it is not the same system. Microsoft says exact-match keywords take priority over Performance Max, while a conversation is unlikely to match an exact keyword verbatim. Broad match and Performance Max are therefore more likely to serve there. ChatGPT takes the departure further: it drops keyword targeting entirely.

Your Search term reports still have a job. So do call recordings. Look past the phrases and identify the circumstances behind them: who has the problem, what they have already tried, and what decision remains. The strings are evidence about customer problems, not a campaign to import.

Make the ad continue the conversation

On Google, Same-Day Tankless Install. $500 Off. Licensed in Chicago can meet a buyer who has already decided to shop for an installation. In our ChatGPT example, the buyer has not reached that point. They have been asking about a possible hazard, a tenant, and whether replacement is even necessary. A discount-led ad at that moment interrupts rather than helps.

The ad should meet the next decision in the thread. If the conversation has moved past the safety issue and into replacement planning, the angle might address a Chicago rental that needs a water heater before Monday, then explain the installation and permit offer. If the gas concern is still unresolved, safety guidance comes first. Name the problem the buyer is solving before presenting the service.

That is not a call to make an ad pretend it wrote the answer. The ad appears below a separately written response, with a short introduction from the ad system. Its opening needs to make sense after what the buyer has just read. Otherwise, even a relevant service can feel like a random interruption.

The landing page has the same obligation. A generic Water Heater Services page may answer the final cost question but ignore everything that led to it. For the rental example, a focused page would address the safety concern first, then repair versus replace, the tenant’s timeline, and price. If the visitor has to repeat the entire conversation in a form to find out whether you handle their situation, the page has thrown away the context that made the ad relevant.

I would pair one problem description with one ad angle and one page whose headline names that problem. A single page shared across five distinct situations is likely to feel generic to all five. The page does not need to recreate the chat; it needs to show the buyer they landed in the right place.

Measure the research, not only the final click

The matching difference changes how you judge results. Google often catches narrowed transactional searches, while ChatGPT can catch research conversations. Imagine the landlord clicks an ad during a Tuesday conversation, talks through the decision, then returns on Friday through a branded Google search and books the job. Last-click reporting credits Google. Reverse the sequence, and ChatGPT may receive credit for a buyer who was already on the way to purchase.

Neither last click tells you enough about the work each placement did. That does not mean you should award every research click a sale, either. It means you need a way to follow what happens after the click:

  • Tag ChatGPT placements with their own UTMs, separate from organic chat referrals.
  • Connect clicks to CRM stages, not only form fills.
  • Watch assisted pipeline and branded search over a 14- to 30-day window alongside direct results.

Do not grade an early research click solely against the last-click CPA of high-intent Search. You may cut a source that feeds later demand. Equally, do not keep it because a dashboard claims every assisted conversion. Ask whether it adds pipeline or merely collects credit for demand you already owned. This is part of what changed in PPC automation as answer surfaces moved some clicks earlier in the buying process.

Give Google and ChatGPT different jobs

Our landlord can touch both systems. They research the problem in ChatGPT, then search your brand and water heater install on Google when they are ready to act. If you use the same creative and CPA target in both places, you obscure which one helped create demand and which one closed it.

I would separate the work. Google exact match handles the transactional tail. ChatGPT handles the problem conversation. They can both contribute to one job without being judged as if they caught the buyer at the same moment.

There is already a related split inside Google: exact-match keywords are not eligible for embedded AI Overview placements; broad match or keywordless targeting can trigger them. That leaves a classic-result lane and an answer-surface lane. ChatGPT is a third lane, with its own budget and reporting.

ChatGPT budgets operate as seven-day averages, and objectives cannot be changed later. I would start with a budget that can withstand a poor week, choose Clicks with the $3-to-$5 max CPC cap described in that setup guide, and report branded Search separately. The decision rule is plain: keep the lane if it adds assisted pipeline and branded demand over 30 days while holding its own CPC; cut it if it only shifts clicks from demand you already had.

Management starts where match types end

A ChatGPT campaign does not hand you a familiar search-term report full of strings to approve or block. You have conversation categories associated with your hints and delivery data that can reflect both relevance and timing. So the weekly work is different.

Take a broad hint such as water heater replacement. If it draws the wrong conversations, split it into the actual decisions: an emergency rental swap, a repair-versus-replace discussion, and a gas-safety concern. Then check which descriptions pull useful traffic, rewrite the loose ones, and adjust the corresponding ads and pages. I would check that weekly, not wait a month while a $25-a-day test budget drifts toward renters who cannot buy an installation.

That gives you a practical way to question anyone selling ChatGPT ads management. Ask what they change each week. A useful answer covers four things: problem descriptions, ad angles, landing pages, and CRM-linked measurement. A slide deck about conversational futures plus a copy of your Google keyword export does not cover them.

groas runs Search Ads across Google, Bing, and ChatGPT for a flat $1,499 a month, with bids and budgets managed around the clock and a landing page for every ad group. That structure fits the job better than paying for someone to move a keyword list into a system that does not use keywords. Management means refining the match between problem, ad, and page, then checking whether that match produces pipeline.

Test the model on a CRM buyer

Change the business and the reasoning should still work. Say you sell a CRM to a 20-person sales team. Your strongest Google keyword is buy CRM software. That phrase fits a buyer ready to compare vendors. In ChatGPT, the same buyer may describe follow-ups slipping, say HubSpot feels heavy, and worry that the team will not log into a replacement.

Start with what the conversation establishes. The buyer has a follow-up problem and a rollout fear. A description of a 20-person team struggling with spreadsheets and wary of a six-month rollout gives the ad system a situation to match. The ad should address the migration concern, not merely announce CRM software. The landing page should answer the rollout question that appeared earlier in the thread. The company changed; the sequence did not: conversation, problem description, relevant ad, relevant page.

Pull your last 20 closed calls and write down the problem each buyer described before naming your product. That is your starting asset for ChatGPT. Your Google keyword export is not. Porting it unchanged means paying to manage a kind of precision this surface does not offer.