---
title: "ChatGPT Ads Cannot Change the Answer. Here Is What They Can Do."
description: "ChatGPT’s answer and its sponsored ad come from separate systems. Follow a water-heater example to see how matching works and why copy, landing pages and conversion tracking matter more than prompt strategy."
url: "https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ecb92422-d600-484a-b48b-fc34d7739822.png"
published: "2026-11-02T06:58:38.888Z"
modified: "2026-11-02T06:58:38.948Z"
---

November 2, 2026 · 11 min read

# ChatGPT Ads Cannot Change the Answer. Here Is What They Can Do.

[Alexander PerelmanHead Of Product @ groas](https://groas.com/author/alexander-perelman)[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

![A handwritten water-heater sizing sheet sealed under a glass bell jar, with a business card and coins pressed outside it, unable to get into the answer.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ecb92422-d600-484a-b48b-fc34d7739822.png)

In this article

1. [First principle: an ad does not buy a sentence in the answer](#first-principle-an-ad-does-not-buy-a-sentence-in-the-answer)
2. [Machine one writes the water-heater answer](#machine-one-writes-the-water-heater-answer)
3. [Machine two matches an ad to the conversation](#machine-two-matches-an-ad-to-the-conversation)
4. [Why that boundary matters to the person reading](#why-that-boundary-matters-to-the-person-reading)
5. [The answer does some work before the ad appears](#the-answer-does-some-work-before-the-ad-appears)
6. [A landing page must qualify the click you cannot inspect](#a-landing-page-must-qualify-the-click-you-cannot-inspect)
7. [Four things a manager can work on](#four-things-a-manager-can-work-on)
8. [Test the model on an invoicing-software comparison](#test-the-model-on-an-invoicing-software-comparison)
9. [Ask which machine the sales pitch claims to move](#ask-which-machine-the-sales-pitch-claims-to-move)

A homeowner spends twenty minutes asking ChatGPT to size a tankless water heater for a three-bath house. An installer wants to reach that homeowner. Paying for an ad will not put the installer’s name in the sizing answer: one system writes the answer, and a separate system considers an ad for the conversation after that answer exists.

**The answer and the ad are built by two different machines.** I used to tell clients to think of this as paid search with prompts instead of keywords. I was wrong. The split changes what an advertiser can buy. You cannot pay to shape the model’s explanation; you can compete for the next step once the reader has it. That is the job groas is built to do.

## First principle: an ad does not buy a sentence in the answer

In paid search, money buys a chance at position. Bid on _tankless water heater installation_, write a tighter ad, and you can appear above organic results. It is easy to carry that model into ChatGPT: pay OpenAI, get your brand into the answer when someone asks which heater to buy. Agency talk of _prompt strategy_ and _visibility engineering_ can make that sound like a service someone has figured out how to sell.

OpenAI’s [testing page says ads do not influence answers](http://openai.com/index/testing-ads-in-chatgpt/). It also says they are labeled sponsored and visually separated. In our homeowner’s conversation, the model explains BTUs, gas-line size and why a 199,000 BTU unit might fit three baths. An advertiser did not buy a word of that explanation. Only after the answer exists does the ad system consider whether an installer belongs below it.

That gives us the rule the rest of this article builds on: **a paid placement is not a paid recommendation.** If an agency promises to get your business into the answer by buying an ad, it is selling control OpenAI says the ad product does not provide.

## Machine one writes the water-heater answer

Stay with the homeowner. They paste room counts, winter gas bills and a photo of the old closet into the chat. The model uses the conversation and the background and product information available to it to write a sizing answer. The [ad system runs separately, on its own models and code](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/); ad selection does not feed back into the response.

The order matters more than the machinery’s names. First, the model gives the homeowner an answer about flow rate and a possible gas-line upgrade. Then an ad might appear underneath. If the ad arrives too late, the answer still appears and the ad is dropped. The homeowner does not have to wait for a sponsor before seeing the sizing explanation.

This is not a claim that every answer will be right. It is a claim about what the ad purchase does not do. A good installer may have a better view of the job once they inspect it, but an ad cannot edit sentence three to make the model recommend that installer. **You cannot bid your way into the sizing math.**

## Machine two matches an ad to the conversation

Once the answer exists, a second system considers whether an ad fits. OpenAI calls its coordinating component the _orchestrator_. It turns relevant parts of the conversation into embeddings, finds candidate ads without requiring an exact keyword match, checks eligibility for location, policy, budget and pacing, and scores fit alongside predicted views, clicks and conversions. A separate auction tool then sets ranking and price for the ads that survive. OpenAI does not disclose the signal weights, so an outsider cannot tune this as if every dial were visible.

### It reads context, not a purchased keyword

Our homeowner may never type “tankless water heater installer.” They spend twenty minutes working through sizing instead. The ad system can treat that progression as a sign that a local installer is relevant. [Advertisers describe that relevance in natural-language context hints rather than buying keywords or audiences](https://www.tink.ca/en/insights/advertising-on-chatgpt).

The advertiser can set a context description, country geography, budget and bid type. They do not get to read the homeowner’s conversation or history. That distinction matters when someone pitches a clever list of prompts to target. The advertiser is describing a situation in which an offer fits, not taking ownership of a phrase or peering into someone’s chat. **Write for the situation the buyer is in, not a keyword you hope they typed.**

### It puts the ad below the answer, if it puts one there at all

In the current rollout, the sponsored placement sits underneath the answer, visibly separate from it. In the [UK rollout, ads appear for Free and Go users, not Plus, Pro or Enterprise users](https://hallam.agency/blog/oh-no-another-ad-platform-chatgpt-ads-launch-in-the-uk/). Even winning an auction does not guarantee an impression: a [post-auction fit check can suppress an ad, and a late ad can miss the delivery window](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/).

Picture the homeowner’s screen. The sizing explanation stands on its own. An installer card may sit below it with a sponsor label. If the match is weak or slow, the homeowner just gets the answer. **A plausible match is not a guaranteed placement**, much less a guaranteed mention in the response.

## Why that boundary matters to the person reading

The homeowner may spend $3,000 on a heater plus installation after acting on the sizing advice. If readers think that advice bends toward whoever paid most, they have a reason to doubt the advice itself. Keeping the ad team out of the answer and the sponsored card outside the text protects a boundary the reader can see.

That concern is not just theoretical. In an [Ipsos Consumer Tracker of 1,085 US adults in late January 2026](https://www.ipsos.com/en-us/ai-skepticism-still-high-and-ads-could-hurt-trust-even-more), 63% said ads in AI search would make them trust results less; 24% disagreed, and 36% thought ads would simplify shopping. Perplexity tried sponsored follow-up questions beginning in November 2024, then [phased ads out in late 2025](https://www.brafton.com/blog/paid-search-blog/perplexity-ads/). In February 2026 it said it would rely on subscriptions and enterprise instead, after even labeled ads raised doubts about whether money shaped a synthesized answer.

OpenAI’s two-system design is its way of drawing that line. For an advertiser, the consequence is practical: **the stricter the wall around the answer, the more your spend has to accomplish after the answer.**

## The answer does some work before the ad appears

By the time our homeowner sees an installer card, they may already understand BTUs, flow rate and the gas-line problem. An ad does not need to start with “What is a tankless water heater?” It needs to offer a useful next step: a licensed install quote, a same-week slot or a page that confirms what the installer will check and names an installation price range.

OpenAI says matching can use conversation topic, past chats and past ad interactions; its [own example pairs recipe research with meal-kit or grocery-delivery ads](http://openai.com/index/testing-ads-in-chatgpt/). The mechanism is the same for the heater. A conversation establishes a need. The ad has a chance to meet the next part of it. **The copy should pick up where the answer leaves off**, not repeat what the reader just learned.

That does not make every click a booked job. The homeowner might still be comparing options, or the ad might offer installation when they only want to understand their existing system. Offer fit decides whether the handoff works. A sponsor label can get an ad noticed; it cannot make a mismatched offer useful.

## A landing page must qualify the click you cannot inspect

In search, a query often tells you what the buyer wants. In ChatGPT, the advertiser does not see the prompt. Internal logs track no-candidate searches and eligibility failures, but [advertisers cannot see which queries triggered ads as they can in Search](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/). UK reporting so far has [limited targeting and very limited reporting, with reach and click optimization only](https://hallam.agency/blog/oh-no-another-ad-platform-chatgpt-ads-launch-in-the-uk/).

That missing query shifts work to the page. For our installer, a generic “home comfort solutions” page throws away the context the conversation built. A page about three-bath sizing, whether a gas-line upgrade is included, the installation price range, the license number and a booking calendar gives the homeowner a way to check fit and act. It also gives the advertiser a better chance to distinguish a serious inquiry from a curious click.

Measurement has to follow the same logic. I would not spend a dollar here without tying the click to booked jobs through offline conversion tracking. Otherwise, a report full of clicks can look healthy while the calendar stays empty. The point is not to admire a new source of traffic. It is to find out whether that traffic became work worth paying for.

Price makes that discipline more important. Launch pricing ran [around $60 CPM with $200k-plus minimums through partners; observed CPMs later fell to $25–$45, with CPC mode recommended at $3–$5 and competitive SaaS and finance at $8–$18](https://www.tink.ca/en/insights/advertising-on-chatgpt). The draft comparison is not cheap display reach: Google Display sits roughly at $2–$5 CPM and Meta at $6–$10, while ChatGPT CPCs land near Google Search. Say you are planning around $20k a month of spend at a $35 CPM. The page and the measurement must justify buying that attention. A trick with prompts will not rescue a generic page. **If the page cannot turn context into a qualified action, the math has nowhere to work.**

## Four things a manager can work on

The controllable work is a shorter list than most sales decks suggest. It is also enough to keep a serious operator busy:

1. **Context description and copy.** State, in plain language, when the offer fits. For our installer, that is the homeowner moving from three-bath sizing to a licensed installation quote, not a pile of heater-related terms.
2. **Offer and landing page.** Make the next step clear. Show the quote path, installation price range and booking option so an educated reader can decide whether to proceed.
3. **Budget, geography and bids.** Set the country, daily caps and payment approach, then watch whether spend is reaching useful opportunities rather than merely buying activity.
4. **Conversion measurement.** Connect the click to booked jobs and revenue instead of stopping the story at a tap on the ad.

That is [what there is to manage in a ChatGPT ads account](https://groas.com/post/what-is-there-to-manage-in-a-chatgpt-ads-account-the-levers). Most of the useful work happens on the page and in the tracking, not in a dramatic tour of the dashboard. Context match can drift as conversations shift, so copy and budgets need attention while those changes still matter. A monthly review is a slow way to catch a problem that is already costing money.

The other side of the list is just as important. No bid, budget or partner badge edits the water-heater explanation, inserts your brand into the answer or guarantees that an ad appears under the right thread. You cannot see the prompt, buy the keyword or force the post-auction fit check to pass. I have heard owners ask for negative-prompt lists and mention monitoring as though either steers ad placement. Neither does.

So _which agencies can actually manage ads inside ChatGPT right now?_ Anyone with access can work those four levers. Nobody can manage the answer through an ad account. The firms worth hiring say that plainly, then do the less glamorous work on pages and pipeline. groas runs that work as [Search Ads across Google, Bing and ChatGPT for one flat fee](https://groas.com/pricing), with the page and tracking included rather than invoiced separately. **If a pitch spends ten minutes on prompts and one minute on your landing page, walk away.**

## Test the model on an invoicing-software comparison

Now take a case with no water heater in sight. A finance lead asks ChatGPT to compare two invoicing tools for a 40-person agency. The answer lists tradeoffs from reviews and documentation. A third vendor wants consideration.

Work through it in order. The answer system writes the comparison; the third vendor cannot buy a row in its table or rewrite a tradeoff. After that answer exists, the ad system may consider the conversation relevant to an offer. The vendor can describe the context it serves: a team outgrowing per-seat billing and looking for flat pricing with migration help. If its ad appears, the click needs to lead to a page that makes the migration case quickly, with tracking that follows a trial through to a paid customer.

A weak page breaks that chain. A prompt trick cannot repair it because the prompt is not the advertiser’s lever. **The machine that writes decides what answer the reader gets. The machine that matches decides whether an advertiser gets a chance at the next click.**

## Ask which machine the sales pitch claims to move

On a sales call, ask one question: _Which machine are you claiming to move?_ A promise to buy mentions, sentiment or a place in the answer claims access to the machine OpenAI keeps separate from ads. A plan for context descriptions, offer fit, situation-specific pages, paced budgets and booked revenue fed back as conversions describes the work an advertiser can actually do.

That second list is unglamorous. I spent years doing versions of it by hand at 1am. Now groas runs it continuously, with page and tracking work included. Once the answer is already written, winning is not about finding a clever way into it. It is about being useful when the reader is ready for the next step.

## Frequently Asked Questions

### Does purchasing an ad placement allow control over the content of the AI response?

No, an ad placement does not provide control over the response content or guarantee a specific phrase within the answer. The AI uses separate systems to generate the text and then, after the answer is created, a different system considers placing an ad underneath it.

### How does the ad system determine relevance to a conversation if keywords were not typed?

The ad system uses embeddings to turn relevant parts of the conversation into contextual data. It finds candidate ads by analyzing the natural language context and situation, rather than relying on an exact match to a purchased keyword.

### If an advertiser wins an ad auction, is there a guarantee that an ad will appear on the page?

A winning auction does not guarantee an impression; a post-auction fit check can suppress an ad. Furthermore, if an ad arrives too late in the conversation, it may miss the delivery window completely, meaning the user only sees the response.

### Since the AI already provides educational information, what should the advertiser's ad offer?

The ad needs to offer a useful next step that picks up where the AI's explanation finishes. This means providing a licensed installation quote, a same-week slot, or a page detailing the installation price range and qualifications.

### What is the impact of not being able to see the user's specific prompts on advertising strategy?

Since advertisers cannot see which specific queries triggered an ad, the focus must shift to the landing page. The page must be designed to qualify the click by addressing the context the AI conversation built for the user.

### What are the four main areas an advertiser should manage in a ChatGPT ads account?

Advertisers should focus on context descriptions and ad copy, the offer and landing page, budget, geography, and conversion measurement. Useful work primarily happens on the external page and in the conversion tracking setup.

## Paid Ads & SEO, Run For You

Businesses pay one flat monthly fee per service, agencies resell it under their own brand, and groas does the work either way.

[See Pricing](https://groas.com/pricing)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is a done-for-you marketing service, not software: elite people working with proprietary technology systems run search ads, Meta ads, SEO and AI search for businesses, and white label for agencies, with a dedicated senior strategist owning every account.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#webpage","url":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two","name":"ChatGPT Ads Cannot Change the Answer. Here Is What They Can Do.","description":"ChatGPT’s answer and its sponsored ad come from separate systems. Follow a water-heater example to see how matching works and why copy, landing pages and conversion tracking matter more than prompt strategy.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ecb92422-d600-484a-b48b-fc34d7739822.png"},"breadcrumb":{"@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#breadcrumb"},"dateModified":"2026-11-02T06:58:38.948Z","mainEntity":{"@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"ChatGPT Ads Cannot Change the Answer. Here Is What They Can Do.","item":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#article","url":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two","mainEntityOfPage":{"@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"ChatGPT Ads Cannot Change the Answer. Here Is What They Can Do.","description":"ChatGPT’s answer and its sponsored ad come from separate systems. Follow a water-heater example to see how matching works and why copy, landing pages and conversion tracking matter more than prompt strategy.","datePublished":"2026-11-02T06:58:38.888Z","dateModified":"2026-11-02T06:58:38.948Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ecb92422-d600-484a-b48b-fc34d7739822.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"answer, ads, cannot, chatgpt, homeowner, installer, page, chatgpt ads","wordCount":2231,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/how-ads-in-llms-actually-work-the-answer-and-the-ad-are-two#faq","mainEntity":[{"@type":"Question","name":"Does purchasing an ad placement allow control over the content of the AI response?","acceptedAnswer":{"@type":"Answer","text":"No, an ad placement does not provide control over the response content or guarantee a specific phrase within the answer. The AI uses separate systems to generate the text and then, after the answer is created, a different system considers placing an ad underneath it."}},{"@type":"Question","name":"How does the ad system determine relevance to a conversation if keywords were not typed?","acceptedAnswer":{"@type":"Answer","text":"The ad system uses embeddings to turn relevant parts of the conversation into contextual data. It finds candidate ads by analyzing the natural language context and situation, rather than relying on an exact match to a purchased keyword."}},{"@type":"Question","name":"If an advertiser wins an ad auction, is there a guarantee that an ad will appear on the page?","acceptedAnswer":{"@type":"Answer","text":"A winning auction does not guarantee an impression; a post-auction fit check can suppress an ad. Furthermore, if an ad arrives too late in the conversation, it may miss the delivery window completely, meaning the user only sees the response."}},{"@type":"Question","name":"Since the AI already provides educational information, what should the advertiser's ad offer?","acceptedAnswer":{"@type":"Answer","text":"The ad needs to offer a useful next step that picks up where the AI's explanation finishes. This means providing a licensed installation quote, a same-week slot, or a page detailing the installation price range and qualifications."}},{"@type":"Question","name":"What is the impact of not being able to see the user's specific prompts on advertising strategy?","acceptedAnswer":{"@type":"Answer","text":"Since advertisers cannot see which specific queries triggered an ad, the focus must shift to the landing page. The page must be designed to qualify the click by addressing the context the AI conversation built for the user."}},{"@type":"Question","name":"What are the four main areas an advertiser should manage in a ChatGPT ads account?","acceptedAnswer":{"@type":"Answer","text":"Advertisers should focus on context descriptions and ad copy, the offer and landing page, budget, geography, and conversion measurement. Useful work primarily happens on the external page and in the conversion tracking setup."}}]}]}
```