Research guide · AI Ads
AI Ads Conversational AdsBelow an AI answer,
will your ad appear?
AI ads (AI Advertising) are paid content that appears in an AI assistant conversation, matched to the context of that conversation. In ChatGPT, an ad appears below the answer, is labeled Sponsored, and is selected by a separate ad system based on the conversation context, the ad content and the bid.
This page covers how ChatGPT ads are matched, how familiar ad terms read differently for AI ads, how to measure results, where ads are available, and LS-Labs’ ad campaign research.
01
AI ads reading guide
From the Taiwan launch and ad formats, to campaign figures, to tracking the traffic ChatGPT brings.
- News
ChatGPT Ads in Taiwan: the September 2026 launch and its implications
After Taiwan opens ChatGPT ads, how delivery, early results and budgets may shift.
- News
ChatGPT Sponsored Agents: how branded ad conversations work
From Sponsored Agents, see how conversational ads work and who sees the ads.
- Growth
ChatGPT Ads Cost and Performance: 9 Public Case Reports
A one-by-one check of CTR, CPC and ROAS across 9 named campaigns.
- Growth
What is UTM, and why do ChatGPT links add utm_source=chatgpt.com?
Use UTM and GA4 to tell apart the traffic ChatGPT brings.
Related terms: Attribution · Incremental conversions · GA4 · Channel groups.
Further reading: Could Google become the next Yahoo? (Traditional Chinese) · How information agents change advertising value (Traditional Chinese).
02
Four research areas for AI ads
LS-Labs splits AI ads research into four areas: how ads match a conversation, what context signals advertisers can provide, how to measure results after the click, and how far the market is open. The first two look at how the system selects ads; the last two look at campaign results and scale.
- 01
Ad matching Matching
Why did this ad appear in this conversation?
OpenAI explains that the ad system selects ads on expected relevance and performance, drawing on the context and intent of the current conversation, the ad’s landing page, headline and copy, and the Context Hints the advertiser provides. When a user has turned on personalized ads, broader ChatGPT usage signals are also considered.
- 02
Context signals Context Hints
How do advertisers tell the system what situations a product suits?
Context Hints are set at the ad group level, using one natural sentence to say what the product is, who it suits and in what situation it helps. They help the system judge relevance, but they are not exact-match keywords and do not guarantee an ad appears in a specific conversation.
- 03
Results measurement Measurement
After the click, did it turn into business?
Ads Manager reports provide impressions, clicks, spend, CTR, average CPC, average CPM and conversions. Advertisers can also add UTM parameters to the landing page URL and track the traffic ChatGPT ads bring with existing tools such as GA4.
- 04
Market watch Market
Who is advertising, and in which countries?
On October 21, 2026, OpenAI announced that ChatGPT Ads is available in 89 countries, with more than 100,000 advertisers having run ads. Ads appear only on Free and Go plans, and businesses in Taiwan can run ads themselves in Ads Manager.
1Find a restaurant for two tonight at 19:00 and book a table.
Consider Restaurant A and Restaurant B; both offer online reservations.
2Match Context Hints, copy and landing page3Rank by relevance and bid
4Relevant to your reservation request
Seats for two still open at 19:00 tonight
Chef’s anniversary menu; book online and skip the wait.
Book a table- 1
Understand the request
Read the intent from this turn: tonight, two people, a reservation. If the request is still broad or the topic is sensitive, no ad is shown this turn.
- 2
Find candidates
Compare each ad’s Context Hints, headline, copy and landing page, and retrieve a batch of relevant restaurant ads first.
- 3
Rank by relevance and bid
Candidate ads are ranked by relevance and bid, and the one that best fits this conversation is selected.
- 4
Place below the answer
The ad is shown separately from the answer, labeled Sponsored, with a one-line intro explaining the connection; Restaurants A and B are listed in the answer as usual.
The full six-stage flow is in “How ads are selected”.
03
AI ads terms explained
CPC, CTR and ROAS keep the same formulas below an AI answer, but they are read differently. The terms below are grouped into four sets by campaign stage. Each term first gives its original meaning, then what changes for AI ads, and how to reassess it.
How to readOriginal meaning → For AI ads → Reassess
Performance metrics, from impressions to ROAS
Impressions
- Original
- An ad counts as one impression each time it is seen. Ratios such as CTR and CPM are calculated from this.
- For AI ads
- Each answer carries at most one ad unit, and an ad is shown only if it arrives before the answer finishes writing. The system also deduplicates at the advertiser and campaign level, so the same brand does not flood a single answer.
- Reassess
- Read impressions together with the conversation that triggered them. A thousand impressions in comparison and shortlisting conversations are closer to a sale than the same number in casual chat.
CTRClick-through rate
- Original
- Clicks ÷ impressions; shows whether an ad attracted the people who saw it.
- For AI ads
- The ad appears after a complete answer, so users have already read comparisons and recommendations; those who click are people who want to go deeper after reading. In the 9 overseas cases LS-Labs compiled, CTR was about 0.35%–1.3%.
- Reassess
- Do not compare CTR directly with search ads. Look at the gap after rewriting Context Hints within the same ad group, then at conversions after the click.
CVRConversion rate
- Original
- Conversions ÷ clicks; shows how many site visitors complete a purchase, booking or form.
- For AI ads
- The conversation has already narrowed the options for the user, so visitors arrive closer to a decision and conversion rate becomes the core signal of whether the match is accurate. Ads Manager supports conversion measurement and can optimize delivery toward standard conversion events.
- Reassess
- Set up conversion events before launch, and break reports down by ad group and landing page to find which context description converts best.
ROASReturn on ad spend
- Original
- Revenue from ads ÷ ad spend; shows how much revenue each dollar spent returns.
- For AI ads
- In the overseas cases LS-Labs compiled, e-commerce ROAS ranges from 1.49x to 3.3–6.8x. Someone who finished comparing in the chat may search the brand name another day or go straight to the official site to order, so counting only revenue at the moment of the click understates the ad’s contribution.
- Reassess
- Extend the attribution window and include revenue from brand search and direct visits during the campaign; only then does ROAS show the full picture.
Pricing, bids and auctions
CPCCost per click
- Original
- Spend ÷ clicks, and the most common pricing model in digital advertising.
- For AI ads
- ChatGPT ads are billed on valid clicks. OpenAI recommends a starting maximum bid of US$3–5 per click for CPC campaigns; across the 9 overseas cases, CPC ranges from US$1.72 to US$22.89.
- Reassess
- Work backward from the post-click conversion value to set the bid. The bid-strength indicator in Ads Manager shows whether your bid is limiting delivery.
CPMCost per mille
- Original
- Spend per thousand impressions, mostly used for campaigns aimed at brand exposure.
- For AI ads
- Choosing the Views objective bills by impressions, and you set a maximum CPM bid. Ads appear only below answers in relevant conversations, so CPM buys people who are discussing this kind of need.
- Reassess
- Read CPM campaigns alongside brand search volume, direct traffic and new-customer share to gauge the value of being seen in conversations.
CPACost per acquisition
- Original
- Spend ÷ conversions; shows what it costs to get one purchase, booking or lead.
- For AI ads
- The Conversions objective (oCPC) optimizes toward conversions and still bills on valid clicks; when you set it through the Ads API, the maximum bid you enter is the CPA bid.
- Reassess
- Build up conversion data with click-based campaigns first, then switch to conversion optimization, capping the CPA target at what your margin can bear.
AuctionAuction and quality score
- Original
- Search ads are ranked by bid and quality score; quality score looks at expected click-through rate, ad relevance and landing page experience.
- For AI ads
- ChatGPT ads use a relevance-weighted second-price auction. The system first evaluates semantic relevance, predicted views, clicks and conversions, and past performance, then decides ranking and price.
- Reassess
- Relevance directly affects ranking and the clearing price. Writing precise Context Hints and matching copy and landing page to the conversation’s need is as much a lever as raising the bid.
Keywords, audiences, creative and brand safety
Keyword→ Context Hints
- Original
- Advertisers buy the terms users search for, and the system matches on those terms to decide whether to show an ad.
- For AI ads
- ChatGPT ads use Context Hints instead: at the ad group level, one natural sentence describes the product, who it suits and in what situation it helps, and the system judges relevance from that rather than by exact match. OpenAI’s example rewrites “running shoes” as “everyday cushioned running shoes for beginners training for a first 5K.”
- Reassess
- Rewrite keyword lists as need scenarios, write one need per ad group, and test different phrasings in separate groups.
AudienceTarget audience and remarketing
- Original
- Select audiences by demographics, interests and behavior, then run remarketing ads to people who have already seen the brand.
- For AI ads
- The current conversation is the most immediate intent signal; when a user turns on personalized ads, the system also draws on past chats, memory and ad interactions. Advertisers do not receive conversations, memory or personal data, only aggregated reports.
- Reassess
- Shift audience strategy from selecting demographics to describing needs and situations; handle remarketing through your own members, lists and site traffic.
CTACreative and call to action
- Original
- Headline, copy, image and call to action decide whether a user clicks after seeing the ad.
- For AI ads
- A ChatGPT ad unit includes the advertiser name, icon, headline, copy, landing page and image, and the system adds a neutral one-line intro written by a model explaining how the ad relates to the conversation. The creative follows a complete answer, so it should add information the answer did not give.
- Reassess
- Write copy as specific information that responds directly to the need, such as price, time slots and eligibility, and rotate several creatives within the same ad group to test.
Brand safetyNegative Phrases
- Original
- Avoid content environments you do not want to appear in, and use negative keywords to exclude irrelevant searches.
- For AI ads
- OpenAI rules that ads do not appear next to sensitive topics such as personal health, mental health and politics, and does not accept political ads; eligible advertisers can also use Negative Phrases to exclude placements that conflict with brand policy.
- Reassess
- Brand safety lists shift from blocking sites to listing the conversation situations you do not want to appear in; advertisers who can use Negative Phrases should write them as short phrases.
Landing pages, tracking and attribution
Landing page
- Original
- The first page seen after clicking an ad; it decides whether a visitor stays and converts.
- For AI ads
- The landing page is itself a matching signal: the ad system consults it to judge relevance, and the ad review crawler also visits submitted landing pages.
- Reassess
- Keep the landing page consistent with the situation the ad group describes, and write key information as text that both the system and readers can read directly.
UTMTracking parametersUrchin Tracking Module
- Original
- Parameters added to a URL to tell apart a visit’s source, medium and campaign.
- For AI ads
- Once static UTM parameters are added to the landing page URL, they travel with the ad click to your site; links inside ChatGPT’s organic answers carry utm_source=chatgpt.com on their own.
- Reassess
- Define your own utm_source and utm_medium for ad traffic, separate from the chatgpt.com traffic that organic answers bring, so the paid and organic contributions can be told apart.
Paid traffic
- Original
- Traffic arriving through paid ads, counted separately from organic traffic.
- For AI ads
- The ad review crawler may visit the landing page with tracking parameters, and analytics tools count it as paid traffic, but these visits are not billed. Impressions, clicks and CTR in Ads Manager update about every 15 minutes, and spend can lag by 7–8 hours.
- Reassess
- Reconcile against valid clicks in Ads Manager, and use paid traffic in GA4 to see post-click browsing and conversion behavior.
Attribution
- Original
- Credit for a conversion is assigned to the channels the user touched; the most common approach is last click.
- For AI ads
- A user can finish comparing inside the chat, then search the brand name or go directly to the official site to order later, so last click credits search and direct traffic.
- Reassess
- Track brand search volume, direct traffic and new-customer share during the campaign, and use geo-split or time-split incrementality tests to estimate ChatGPT ads’ actual contribution.
04
How ads are selected
On October 21, 2026, OpenAI published “How we built ChatGPT’s novel ad stack for user trust and relevance”. Ad selection runs in a separate service, at the same time as answer generation, with a post-trained LLM orchestrator calling tools in sequence; returning no ad is also a normal outcome.
- 01
Understand the conversation
The orchestrator condenses the multi-turn conversation into a summary and rewrites it as one or more queries that express commercial intent.
- 02
Retrieve candidates
Queries are converted to semantic vectors and searched with approximate nearest neighbor (ANN) search, deliberately retrieving a large batch of candidate ads first.
- 03
Filter out ineligible ads
Exclude ads with exhausted budgets, region mismatches, policy violations, or that do not suit this kind of conversation.
- 04
Score
Evaluate semantic relevance, predicted views, clicks and conversions, and the ad’s past performance.
- 05
Auction
A relevance-weighted second-price auction decides ranking and price; the orchestrator only nominates candidates and does not decide delivery directly.
- 06
Final check and display
If the ad still seems out of place in the conversation, or did not arrive before the answer finished, no ad is shown this turn. When shown, it carries a one-line neutral intro written by a model and is labeled Sponsored.
The conversation runs: “I’m spending two days in Yosemite. Which sights should I plan?” “Will it be cold next week?” “What new product lines does [a brand] have?” The user never linked the brand to the trip, yet the ad system may still distill the conversation into two queries: shoes appropriate for hiking trip and new [brand] product lines good for travel.
05
Ad performance calculator
Adjust budget, bid and two rates to estimate the impressions, clicks and conversions this budget buys. The reference ranges for CTR and CPC come from the 9 overseas campaigns compiled in “ChatGPT Ads Cost and Performance: 9 Public Case Reports”; OpenAI recommends a starting maximum bid of US$3–5 per click for CPC campaigns.
Adjust the numbers
The pink band is the 9 cases’ US$1.72–22.89; OpenAI recommends a starting bid of US$3–5
The pink band is the 9 cases’ 0.35%–1.3%
Share of clicks that complete a purchase, booking or form; enter your own site’s number
Results
Cost per conversion (CPA)US$200.00
- Impressions
- 31,250
- Clicks
- 250
- Conversions
- 5
- Cost per thousand impressions (CPM)
- US$32.00
Clicks = budget ÷ CPC, impressions = clicks ÷ CTR, conversions = clicks × conversion rate, CPA = budget ÷ conversions
US$1,000 buys about 250 clicks and 5 conversions, about US$200.00 per conversion.
Report refresh times
Impressions, clicks and CTR in Ads Manager update about every 15 minutes. Spend can lag by 7–8 hours, and spend showing zero does not mean nothing is being billed.
Track traffic with UTM
Add UTM parameters to the landing page URL. They travel with the ad click to your site, so existing tools such as GA4 can separate ChatGPT ad traffic.
Visits from the review crawler
The ad review crawler may visit the landing page with tracking parameters, and analytics tools count it as paid traffic, but these visits are not billed clicks. Billing follows the valid clicks in Ads Manager.
06
Conversation context and ad performance study
LS-Labs ChatGPT Ads Pilot Study 01 asks one question: for the same set of ads, if the Context Hints are written differently, do the conversations the ads appear in and their performance change?
- What is compared
- Two sets of Context Hints: one describes product features, the other describes usage situations.
- What is held constant
- The same ad headline, copy, image, landing page and bid.
- Which numbers
- CTR, CPC, conversion rate and CPA, with impressions for the two sets recorded separately.
- How to read it
- Compare the two sets’ impressions and traffic sources first, then look at performance differences, to avoid treating traffic differences as a matching effect.
- 01
Preliminary research
Map the product’s usage situations and competitor positioning, and write the two sets of Context Hints.
- 02
Campaign test
The partner provides the Ads Manager account and budget; LS-Labs takes part in setup and monitoring, for two weeks.
- 03
Performance analysis
Compare CTR, CPC, conversion rate and CPA for the two sets, checked against GA4 visit and conversion records.
- 04
Publication
With both parties’ consent, publish the campaign settings, test limitations and shareable data anonymously.
07
Common questions
When did ChatGPT ads start?
OpenAI began testing ChatGPT ads in the US on February 9, 2026, and expanded to Southeast Asia and Taiwan in September. On October 21, OpenAI announced that ChatGPT Ads is available in 89 countries, with more than 100,000 advertisers having run ads.
Do ads affect ChatGPT’s answers?
OpenAI says they do not. Ads are selected by a separate system that runs apart from answer generation, and advertisers cannot influence, rank or modify ChatGPT’s answers. Ads are clearly labeled, and seeing an ad does not mean OpenAI endorses that advertiser.
Who sees ChatGPT ads?
Users on the Free and Go plans may see ads; Plus, Pro, Business, Enterprise and Edu plans have no ads. OpenAI does not show ads to accounts it determines to be under 18, and temporary chats do not show ads either.
Which conversations do not show ads?
OpenAI rules that ads do not appear next to sensitive topics such as personal health, mental health or politics, and political ads are not currently accepted. Eligible advertisers can also use Negative Phrases to exclude placements that conflict with their own brand policy.
Are Context Hints keywords?
No. According to OpenAI’s “Writing context hints for ChatGPT Ads”, Context Hints supplement the product, its intended audience and use situations to help the system judge relevance; they are not exact-match keywords or delivery rules, and they do not guarantee an ad appears in a specific conversation.
Can businesses in Taiwan run ads themselves?
Yes. Taiwan is among the Ads Manager self-serve countries, provided the legal entity that runs and pays for the campaigns is based in an available country. As of November 6, 2026, the list shows 89 countries and regions.
What CPC bid should I set?
OpenAI recommends a starting maximum bid of US$3–5 per click for CPC campaigns, and Ads Manager also indicates whether a bid is competitive. In the 9 public cases LS-Labs compiled, CPC ranged from about US$1.72 to US$22.89.
How do ChatGPT ads differ from GEO?
GEO deals with brand mentions and source citations in organic answers, while ads are a paid placement below the answer, labeled Sponsored. OpenAI says ads do not affect answers, so running ads does not bring a brand into organic answers; manage and measure the two separately.
08
How the four research topics connect
GEO examines mentions and citations in organic answers, TSO examines which tool an agent selects, AX examines whether a task gets completed, and AI ads examine how the paid placement below an answer is matched and converts. The four address adjacent questions and can be read across by content.
