---
title: "There Is No Keyword: How ChatGPT Ads Find Their Audience"
description: "ChatGPT ads respond to conversation context, not a typed search query. Here is what replaces keywords, what advertisers can control, and how to track the delayed click."
url: "https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/226cb238-eb02-42f9-90ef-94fd8d5e3cae.png"
published: "2026-11-03T06:40:35.600Z"
modified: "2026-11-03T06:40:35.659Z"
---

November 3, 2026 · 10 min read

# There Is No Keyword: How ChatGPT Ads Find Their Audience

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

![A hand holds a ring of brass keys tagged with search phrases up to a door made of paper speech bubbles that has no keyhole — the keyword has nowhere to go.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/226cb238-eb02-42f9-90ef-94fd8d5e3cae.png)

In this article

1. [First, understand the handle a search advertiser loses](#first-understand-the-handle-a-search-advertiser-loses)
2. [The conversation builds the shortlist; the auction picks the ad](#the-conversation-builds-the-shortlist-the-auction-picks-the-ad)
3. [The answer comes first, and the ad may not appear](#the-answer-comes-first-and-the-ad-may-not-appear)
4. [Replace keyword controls with the levers you actually have](#replace-keyword-controls-with-the-levers-you-actually-have)
5. [Track the click beyond the first visit](#track-the-click-beyond-the-first-visit)
6. [So what does managing ChatGPT ads mean?](#so-what-does-managing-chatgpt-ads-mean)
7. [Test the model against three different buying journeys](#test-the-model-against-three-different-buying-journeys)

Most agencies selling ChatGPT ads right now will show you a keyword list. Ask what those keywords match and the call goes quiet.

I spent nine years building those lists by hand. A query came in, a keyword matched it, a match type decided how close was close enough, and an auction decided who paid what. I could tell you why a search for “emergency plumber Leeds Sunday” triggered an ad at 1am and what I added as a negative the next morning. That chain breaks inside a chatbot.

Take the example I will use throughout: someone tells ChatGPT their gas boiler is 18 years old, their house is a three-bed semi outside Manchester, and they want to know whether a heat pump makes sense before winter. Over six messages, they mention a converted loft and correct their budget. There is no single search query to match. There is a situation taking shape.

**The conversation is the targeting signal, not the last sentence.** The system reads prior turns, constraints and corrections together, decides which ads can be considered, then runs a separate auction for a small Sponsored card under the answer. Keep thinking in keywords and you will buy the wrong controls, read the wrong report, and blame the wrong part of the funnel.

## First, understand the handle a search advertiser loses

In Google Ads, I could draw the basic chain on a whiteboard. Someone typed a short query. Google checked it against keywords in my account. Match type set the tolerance: exact for close variants, phrase for the same idea in a longer string, broad for what Google thought was related. Ad Rank sorted eligible ads by bid and quality. The winner paid the minimum needed to hold position.

That chain gave me two tools I lived inside. The search terms report showed the strings that triggered ads, and negatives let me cut the ones wasting money. Say you spend $20k a month on boiler installs. If “DIY heat pump grant calculator” triggered a broad-match ad 41 times and converted zero times, I could add it as a negative in about 30 seconds. **The keyword was the handle I used to steer the account.**

The Manchester homeowner may never type “heat pump installer Manchester.” They say the boiler is old, mention the semi, explain that the converted loft leaves no room for a hot water cylinder, then correct the budget. Two people can end with the same question and have different ad eligibility because the conversations leading there differ. The next layer is not a new match type. It is a different way of deciding what is relevant.

## The conversation builds the shortlist; the auction picks the ad

[Prior turns, constraints, corrections and budget shape whether an ad is considered](https://kenashe.ai/blog/2026-10-22-chatgpt-ads-start-with-the-conversation-not-the-keyword/). OpenAI uses [a custom-trained LLM orchestrator to turn the multi-turn conversation into semantic retrieval queries, check eligibility and score relevance before sending candidates to a separate second-price auction](https://onemetrik.com/market-insights/chatgpt-ad-stack/). **Orchestration chooses the shortlist. The auction chooses the winner.** They are different jobs.

That is why I would not expect a search terms report or a familiar negative-keyword workflow. [There are no search terms to list and no phrase to block as a negative](https://gptads.chat/posts/chatgpt-ads-negative-keywords/). The match draws on the conversation alongside your landing page, copy and hints; it does not hand you a string to cross out. The closest steering wheel is [plain-language context hints at ad-group level, up to 2,000 per group, which OpenAI says are not exact-match keywords and do not guarantee delivery](https://digitalmatters.me/marketing/chatgpt-ads-audience-targeting/).

For the heat pump installer, a useful hint describes the cases they serve. It cannot force an impression whenever someone mentions a boiler. If leads point to poor fit, I tighten the hint, filter the product feed harder or exclude a location. I do not hunt for a phrase to negate. **Describe the situation you can serve; do not disguise a keyword list as a hint.**

## The answer comes first, and the ad may not appear

The ad does not enter the answer. It appears as [a single labelled Sponsored card below it, visually separate from the answer](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/). Separate systems handle answer generation and advertising, so the card cannot rewrite what ChatGPT just said. Our homeowner reads about cylinder space and winter load first. Then they may see a card for a heat pump survey. The click, if it comes, follows research rather than interrupting it.

Even winning the auction [does not guarantee an impression: final checks can block the ad, and ChatGPT will not hold its answer to wait for one](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/). The eligible audience is narrower too: [ads appear only for logged-in adults on Free and Go plans, not on Plus, Pro, Business, Enterprise or Edu plans, or in under-18 or Temporary Chats](https://www.searchenginejournal.com/openai-details-chatgpt-llm-ad-stack/593857/).

**A relevant conversation is not the same thing as an available impression.** Expect good conversations in which no card can show. That matters before you start blaming the copy for every missing click.

## Replace keyword controls with the levers you actually have

Strip away keywords and the available controls are short enough to list. [Conversation context is always on; advertisers can also use context hints, location, platform and custom audiences. There is no age, gender, interest-category, lookalike or site-retargeting control](https://digitalmatters.me/marketing/chatgpt-ads-audience-targeting/).

- **Context hints** describe the situations in which your ad belongs. They are guidance, not exact-match triggers.
- **Location** can use [up to 2,500 location IDs](https://digitalmatters.me/marketing/chatgpt-ads-audience-targeting/). For our installer, the North West matters more than a speculative homeowner persona.
- **Platform** means iOS app, Android app or web. Split them if click quality gives you a reason.
- **Custom audiences** can help include, adjust for or exclude known people, subject to the list rules.

I can ask to show the Manchester installer’s ad in the North West across all three platforms. I cannot ask to show it to homeowners aged 35 to 54 who like home improvement. **If an agency pitch deck leads with that persona, it was written for a different ad system.**

Custom audiences feel most like old PPC, but the details matter. [A list needs 25,000 matched users for inclusion or bid adjustment, and OpenAI recommends 100,000; a list of any size can be used for exclusion, which wins if a user belongs to both](https://gptads.chat/posts/chatgpt-ads-negative-keywords/). Files accept email, phone, SHA-256 hashed email or phone, and GAID, up to 500MB.

I used to tell clients to build big remarketing pools and let the algorithm find more people like them. I was wrong here. The practical use of a smaller list is defensive: exclude existing customers or the 400 people who already booked a survey. **Hints say where you belong; exclusions say who you should not pay to reach again.**

### Write the card for someone already mid-research

On Google, the click could follow eight words. Here it may follow 400. Our homeowner has read about flow temperatures, cylinder space and why the converted loft complicates a standard setup. “Best Heat Pumps 2026. Get a Quote” sounds like an ad that missed the conversation.

Copy that addresses the constraint has a better reason to earn the click: “No cylinder space after a loft conversion. Survey checks high-temp pump fit for 3-bed semis in Greater Manchester.” The first line recognises the problem; the second gives the reader a reason to continue. **Specific fit beats a generic promise when the reader has already done the homework.**

### Make the landing page feel like the next turn

That click will punish a generic homepage. Start over with stock photos and three vague benefits, and the thread breaks. The page should work like _message seven_: yes, we survey converted-loft semis around Manchester; here is what we check for cylinder alternatives; here is a realistic price band and timeline before winter; here is when a high-temp pump is a bad fit.

The page also helps with matching before anyone clicks. The practical move is to [map the pre-buying dialogue and publish product facts, fit and misfit criteria, and trust signals in crawlable form](https://kenashe.ai/blog/2026-10-22-chatgpt-ads-start-with-the-conversation-not-the-keyword/). **Write for the person checking whether you can handle their awkward case**, not the person hearing about heat pumps for the first time.

## Track the click beyond the first visit

A ChatGPT click can behave like homework, not an impulse buy. Someone reads the answer, clicks the card, returns to the chat, asks two more questions, searches your brand name tomorrow and books next week. One early UK lead-gen test on ChatGPT-only traffic [saw form fills while the platform still showed zero conversions two weeks later, and around 100 reported clicks against 20 in the advertiser’s own tracking](https://aireadycmo.com/p/chatgpt-ads-have-a-measurement-problem). Another operator puts the [conversion delay at 24 to 36 hours, versus a few hours on Google](https://aireadycmo.com/p/chatgpt-ads-have-a-measurement-problem).

Attribution is [click-based, with 7-, 14- and 30-day click windows and 0- or 1-day view windows](https://aireadycmo.com/p/chatgpt-ads-have-a-measurement-problem). A later sale through brand search may appear elsewhere in your reports. I would set up three checks outside the platform from day one:

1. **Log clicks server-side and retain their UTMs** in your own tracking, so a platform-versus-site gap is visible instead of mysterious.
2. **Record identifiable ChatGPT-assisted journeys in the CRM**, including a later brand search when you can connect it to the original visit.
3. **Read assisted conversions weekly**, rather than panicking over daily ROAS while people are still researching.

That does not make every delayed sale yours. It gives you a way to distinguish a slow path from a dead one. **Judge this traffic only on same-session conversion rate and you may kill a useful channel for looking slow.**

## So what does managing ChatGPT ads mean?

The real work is less theatrical than a keyword-build deck, and it happens every week. Tighten context hints when lead quality shows the wrong fit. Keep location IDs clean and exclusion lists current. Split iOS, Android and web when click quality differs. Test copy for a reader who knows the basics. Fix the landing page so it reads as _message seven_, with price bands, fit and misfit criteria, and a next step that does not force the researcher to start over. Then reconcile conversions outside the platform.

The fake work is easier to sell: keyword builds, match type tuning, weekly negative mining and search terms reports with neat pie charts. There is no query string here to mine. When an agency says it can run ChatGPT ads, ask to see its hint structure, exclusion lists, mid-research creative and delayed-click tracking plan. If it hands you keywords instead, it is reselling a Google process.

[groas runs Search Ads across Google, Bing and ChatGPT for one flat $1,499 a month](https://groas.com/pricing), with bidding, targeting and budgets worked around the clock and a landing page for every ad group. That is a better fit for this work than paying someone to perform a weekly keyword ritual the channel does not support. **The job is to manage context, pages and tracking, not pretend the old handle is still there.**

## Test the model against three different buying journeys

The boiler example has a long conversation and a local service at the end. Change the buying journey and the same model should still tell you what to do.

**A 120-day B2B sales cycle makes the first click look worthless if you score it like search.** Take a finance lead evaluating invoice software. Six people ask ChatGPT about approval chains, NetSuite fit and audit logs over three weeks before anyone speaks to sales. Your ad appears during a conversation in week one; the demo comes in week twelve from a direct visit. The relevant context is detailed, but the effect is months away. Judge that first touch on pipeline created using a 90-day lookback in your own records, and it has a chance to earn its keep. Judge it on cost per demo within 14 days, and you may pause the touch that started the deal.

A local emergency job flips the timing. Nobody chats for six turns about a burst pipe at 11pm. The conversation is two lines long, location does most of the work, and the landing page has one job: answer the phone or show the next available slot. **Thin context puts more weight on location and the opening line than on a long hint.** There will also be fewer eligible chats and more week-to-week variance. This will not suit everyone; a low-volume town may not have enough chat volume to learn from.

Now take an ecommerce shopper comparing three air fryers over eight turns. They weigh price, capacity and cleaning, then correct for a small kitchen. There is enough context to identify the constraint before the click. The feed and crawlable product and category pages need to give the orchestrator the facts it can verify: sizes, compatibility, price bands and what does not fit. The hint should read like a shelf label, not a slogan. If the catalogue misstates the fit, the conversation will expose the mismatch faster than a keyword ever did.

That is the test I would put to any operator, including myself: identify the context, show the page that continues it, and explain how a later conversion gets counted. **If the answer begins with a keyword sheet, hire the operator who threw it away first.**

## Frequently Asked Questions

### What replaces keywords in ChatGPT ads?

The multi-turn conversation replaces the keyword. Prior turns, constraints, corrections and budget together decide whether an ad is even considered, then a separate second-price auction picks the winner from the shortlist. There is no single search query to match.

### Can I block specific search terms with negative keywords in ChatGPT ads?

No. There are no search terms to list and no phrase to block as a negative, because matching draws on the conversation alongside your landing page, copy and hints. The closest steering wheel is plain-language context hints at ad-group level, which are guidance rather than exact-match triggers.

### Do context hints in ChatGPT ads work like exact-match keywords?

No. Context hints describe the situations in which your ad belongs, and OpenAI says they are not exact-match keywords and do not guarantee delivery. A hint can guide the system toward the cases you serve, but it cannot force an impression whenever someone mentions a related term.

### Where do ChatGPT ads appear and who can see them?

The ad appears as a single labelled Sponsored card below the answer, visually separate from it, and separate systems handle answers and ads. Ads show only for logged-in adults on Free and Go plans, not on Plus, Pro, Business, Enterprise or Edu plans, and not in under-18 or Temporary Chats.

### What targeting controls do advertisers get in ChatGPT ads?

Advertisers can use context hints, location with up to 2,500 location IDs, platform splits across iOS, Android and web, and custom audiences. Conversation context is always on. There is no age, gender, interest-category, lookalike or site-retargeting control.

### How big does a custom audience list need to be for ChatGPT ads?

A list needs 25,000 matched users for inclusion or bid adjustment, and OpenAI recommends 100,000. A list of any size can be used for exclusion, and exclusion wins if a user belongs to both. Files accept email, phone, SHA-256 hashed email or phone, and GAID, up to 500MB.

### How should I write ad copy for ChatGPT sponsored cards?

Write for someone already mid-research who may have read hundreds of words before the click, so a generic promise sounds like it missed the conversation. Copy that addresses the reader's actual constraint, such as no cylinder space after a loft conversion, gives a specific reason to continue.

### Why do my ChatGPT ad conversions look delayed or missing?

A ChatGPT click often behaves like research: one operator puts the conversion delay at 24 to 36 hours, and an early UK test saw form fills while the platform still showed zero conversions two weeks later, with about 100 reported clicks against 20 in the advertiser's own tracking. Log clicks server-side with their UTMs and record assisted journeys in the CRM to distinguish a slow path from a dead one.

## 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/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#webpage","url":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them","name":"There Is No Keyword: How ChatGPT Ads Find Their Audience","description":"ChatGPT ads respond to conversation context, not a typed search query. Here is what replaces keywords, what advertisers can control, and how to track the delayed click.","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/226cb238-eb02-42f9-90ef-94fd8d5e3cae.png"},"breadcrumb":{"@id":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#breadcrumb"},"dateModified":"2026-11-03T06:40:35.659Z","mainEntity":{"@id":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#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":"There Is No Keyword: How ChatGPT Ads Find Their Audience","item":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#article","url":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them","mainEntityOfPage":{"@id":"https://groas.com/post/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"There Is No Keyword: How ChatGPT Ads Find Their Audience","description":"ChatGPT ads respond to conversation context, not a typed search query. Here is what replaces keywords, what advertisers can control, and how to track the delayed click.","datePublished":"2026-11-03T06:40:35.600Z","dateModified":"2026-11-03T06:40:35.659Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/226cb238-eb02-42f9-90ef-94fd8d5e3cae.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"ads, chatgpt, chatgpt ads, keyword, search, click, conversation, ads find","wordCount":2203,"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/there-is-no-keyword-how-ads-in-llms-decide-who-sees-them#faq","mainEntity":[{"@type":"Question","name":"What replaces keywords in ChatGPT ads?","acceptedAnswer":{"@type":"Answer","text":"The multi-turn conversation replaces the keyword. Prior turns, constraints, corrections and budget together decide whether an ad is even considered, then a separate second-price auction picks the winner from the shortlist. There is no single search query to match."}},{"@type":"Question","name":"Can I block specific search terms with negative keywords in ChatGPT ads?","acceptedAnswer":{"@type":"Answer","text":"No. There are no search terms to list and no phrase to block as a negative, because matching draws on the conversation alongside your landing page, copy and hints. The closest steering wheel is plain-language context hints at ad-group level, which are guidance rather than exact-match triggers."}},{"@type":"Question","name":"Do context hints in ChatGPT ads work like exact-match keywords?","acceptedAnswer":{"@type":"Answer","text":"No. Context hints describe the situations in which your ad belongs, and OpenAI says they are not exact-match keywords and do not guarantee delivery. A hint can guide the system toward the cases you serve, but it cannot force an impression whenever someone mentions a related term."}},{"@type":"Question","name":"Where do ChatGPT ads appear and who can see them?","acceptedAnswer":{"@type":"Answer","text":"The ad appears as a single labelled Sponsored card below the answer, visually separate from it, and separate systems handle answers and ads. Ads show only for logged-in adults on Free and Go plans, not on Plus, Pro, Business, Enterprise or Edu plans, and not in under-18 or Temporary Chats."}},{"@type":"Question","name":"What targeting controls do advertisers get in ChatGPT ads?","acceptedAnswer":{"@type":"Answer","text":"Advertisers can use context hints, location with up to 2,500 location IDs, platform splits across iOS, Android and web, and custom audiences. Conversation context is always on. There is no age, gender, interest-category, lookalike or site-retargeting control."}},{"@type":"Question","name":"How big does a custom audience list need to be for ChatGPT ads?","acceptedAnswer":{"@type":"Answer","text":"A list needs 25,000 matched users for inclusion or bid adjustment, and OpenAI recommends 100,000. A list of any size can be used for exclusion, and exclusion wins if a user belongs to both. Files accept email, phone, SHA-256 hashed email or phone, and GAID, up to 500MB."}},{"@type":"Question","name":"How should I write ad copy for ChatGPT sponsored cards?","acceptedAnswer":{"@type":"Answer","text":"Write for someone already mid-research who may have read hundreds of words before the click, so a generic promise sounds like it missed the conversation. Copy that addresses the reader's actual constraint, such as no cylinder space after a loft conversion, gives a specific reason to continue."}},{"@type":"Question","name":"Why do my ChatGPT ad conversions look delayed or missing?","acceptedAnswer":{"@type":"Answer","text":"A ChatGPT click often behaves like research: one operator puts the conversion delay at 24 to 36 hours, and an early UK test saw form fills while the platform still showed zero conversions two weeks later, with about 100 reported clicks against 20 in the advertiser's own tracking. Log clicks server-side with their UTMs and record assisted journeys in the CRM to distinguish a slow path from a dead one."}}]}]}
```