Prompt set funnel from prompts tested to target brand appearances

# Prompt Mapping: The New Keyword Research for AI Advertising # Prompt Mapping: The New Keyword Research for AI Advertising OpenAI began testing ads in ChatGPT for US logged-in adults on the Free and Go tiers on February 9, 2026, while Pro, Business, and Enterprise stay ad-free ([OpenAI, "Testing ads in ChatGPT"](https://openai.com/index/testing-ads-in-chatgpt/); [TechCrunch, "ChatGPT rolls out ads," 2026](https://techcrunch.com/2026/02/09/chatgpt-rolls-out-ads/)). That change moves the unit of ad targeting from the keyword to the prompt. Prompt mapping ChatGPT ads is the new keyword research: you systematically discover which conversational prompts surface sponsored placements, then organize them by buyer-journey stage. > **Key Takeaways** > - ChatGPT ad targeting is contextual, matched on conversation topic, chat history, and prior ad interactions, not exact-match keywords ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). > - There is no public ad directory, so the only way to see triggering prompts is to run many varied prompts in eligible sessions and log results ([Search Engine Journal, "How To See If Competitors Are Placing Ads In ChatGPT Answers," 2026](https://www.searchenginejournal.com/see-competitor-ads-chatgpt-trendos-spa/575883/)). > - Per prompt, capture ad title, ad description, final URL, and impression share ([Search Engine Land, "What ChatGPT Ads data reveals about your competitors," 2026](https://searchengineland.com/what-chatgpt-ads-data-reveals-about-your-competitors-479301)). > - Reported ChatGPT CPCs run roughly $2.50 to $8.00, above Google Search at about $1 to $3 ([Maciej Turek, "ChatGPT Ads 2026"](https://maciejturek.com/resources/chatgpt-ads-2026.html)). ## What is prompt mapping? Prompt mapping is the practice of cataloging which conversational prompts trigger sponsored placements inside ChatGPT, then grouping those prompts by buyer-journey stage. It mirrors keyword research, but the unit changed. OpenAI matches ads on conversation topic, chat history, and prior ad interactions, so the prompt and its surrounding context are what you target ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). Think of a prompt map as a spreadsheet where each row is a real buyer question. You record the prompt text, the journey stage it represents, whether it triggered any ad, which advertisers showed, and how often. Over enough runs, patterns emerge. Some prompts reliably pull sponsored answers. Others never do. Why does this matter now? Because the old playbook assumed a keyword bid in an auction you could inspect. Conversational ads hide that auction inside private threads. To plan spend or read a competitor's footprint, you first need to know which prompts even produce ads. That discovery step is prompt mapping. For the broader monitoring program this feeds, see [monitor ChatGPT ads at scale](https://www.joinmassive.com/blog/how-to-monitor-chatgpt-ads). > **Citation capsule:** As of early 2026, ChatGPT ad targeting is contextual rather than keyword-based, matched on conversation topic, chat history, and prior ad interactions ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). Prompt mapping catalogs which prompts trigger placements and sorts them by buyer stage. ## Why don't keywords work for ChatGPT ads? Keywords do not work because advertisers cannot buy exact-match terms in ChatGPT. They give ad groups "context hints," meaning topics and conversations, and those hints do not guarantee placement. OpenAI decides delivery based on relevance ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). The lever is context, not a bid on a string. That difference reshapes research. A keyword like "best CRM" maps to one search query. In a chat, the same intent spreads across dozens of phrasings: "which CRM should a 10-person agency use," "is HubSpot worth it for a small team," "cheapest CRM with email automation." Each is a distinct prompt with its own context, and each may or may not surface an ad. [IMAGE: Side-by-side comparison of a single search keyword versus a branching tree of conversational prompts - search terms "conversation tree diagram", "branching dialogue flowchart"] There is a second wrinkle. Targeting also factors chat history and prior ad interactions, so the same prompt can behave differently across sessions and users ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). You are not measuring a fixed auction. You are sampling a probabilistic system, which is why volume and repetition matter so much. To see how this contrast plays out in the answer body itself, compare [organic versus paid placements](https://www.joinmassive.com/blog/organic-vs-paid-chatgpt). ## How do you build a prompt map? Building a prompt map starts with real buyer questions, run at volume in eligible sessions, with results logged. Because there is no public ad directory and ads match per private thread, running many varied prompts and capturing what appears is the only practical way to see triggers ([Search Engine Journal, "How To See If Competitors Are Placing Ads In ChatGPT Answers," 2026](https://www.searchenginejournal.com/see-competitor-ads-chatgpt-trendos-spa/575883/); [cloro.dev, "How to Monitor ChatGPT Ads," 2026](https://cloro.dev/blog/monitor-chatgpt-ads/)). Here is a method that holds up in practice. ### Step 1: Build prompt sets from real buyer questions Start with how people actually talk to an assistant. Mine sales call notes, support tickets, and search queries, then rewrite each as a conversational prompt. Group them by journey stage: awareness ("what is X"), consideration ("X versus Y"), and decision ("is X worth the price"). Aim for 15 to 30 phrasings per intent so you cover natural variation. ### Step 2: Run the prompts at volume in eligible sessions Ads only test on Free and Go tiers for US logged-in adults, so your sessions must qualify ([OpenAI, "Testing ads in ChatGPT"](https://openai.com/index/testing-ads-in-chatgpt/)). Run each prompt many times, because placement is probabilistic and history-dependent. A single run tells you little. Dozens per prompt start to reveal a stable signal. Running prompt sets at this scale is its own engineering problem, covered in [running prompt sets at scale](https://www.joinmassive.com/blog/how-to-scrape-chatgpt-ads). ### Step 3: Log what each prompt returns For every prompt, capture the ad title, ad description, final URL, and an impression share, calculated as appearances divided by total runs ([Search Engine Land, "What ChatGPT Ads data reveals about your competitors," 2026](https://searchengineland.com/what-chatgpt-ads-data-reveals-about-your-competitors-479301)). Impression share is the column that turns anecdotes into a map. A prompt that shows a competitor 8 times in 10 runs is a different signal than one that shows them once.

Illustrative prompt-set funnel. Capture method per Search Engine Land, "What ChatGPT Ads data reveals about your competitors," 2026.

## How do you read a finished prompt map? A finished prompt map tells you three things: which buyer questions trigger ads, which advertisers compete for them, and how dominant each advertiser is. Reading it well means treating impression share as the spine of analysis, since per-prompt capture of title, description, final URL, and share is what the data supports ([Search Engine Land, "What ChatGPT Ads data reveals about your competitors," 2026](https://searchengineland.com/what-chatgpt-ads-data-reveals-about-your-competitors-479301)). Sort prompts by stage, then by impression share. High-share decision-stage prompts are where money concentrates. Those tend to be the most expensive to compete in, which the pricing data supports: reported ChatGPT CPCs sit near $2.50 to $8.00, above Google Search at roughly $1 to $3, reflecting high-intent research-mode users ([Maciej Turek, "ChatGPT Ads 2026"](https://maciejturek.com/resources/chatgpt-ads-2026.html)). Here is the part most teams miss. In keyword search, share of voice is a slice of a known pie, because the auction and its terms are visible. In ChatGPT, the pie itself is hidden, so a prompt map is the only instrument that estimates the pie's shape. That inverts the workflow: you build the measurement surface before you can measure. Treat your prompt map as infrastructure, not a one-off report, and re-run it on a schedule, because targeting that factors chat history will drift as the system learns. For the metric this feeds, see [measuring share of voice](https://www.joinmassive.com/blog/chatgpt-ads-share-of-voice). ## Where does the heavy lifting happen? The hard part is volume and geography: you need many prompt runs from real-looking sessions across regions, because placement is probabilistic and partly history-dependent ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). One laptop running prompts by hand cannot produce a stable impression-share signal. This is an infrastructure problem before it is an analysis problem. Massive is a device-access network plus a rendering stack built for exactly this layer. Its Web Render API `/ai` endpoint returns ChatGPT, Gemini, Perplexity, and Copilot completions through real-user-device origins in any geo, returned as a completion plus sources HTML and a subqueries array, synchronously or asynchronously. That is the layer a team runs large, geo-varied prompt sets through. The network spans 1M+ verified residential devices across 195+ countries, ethically sourced via an opt-in SDK, with SOC 2, GDPR, and AppEsteem compliance. [IMAGE: A world map dotted with device nodes representing geo-distributed residential origins - search terms "global network map nodes", "distributed devices world map"] ## Frequently Asked Questions ### What is prompt mapping for ChatGPT ads? Prompt mapping is the practice of cataloging which conversational prompts trigger sponsored placements in ChatGPT, then grouping them by buyer-journey stage. It replaces keyword research because targeting is contextual, matched on conversation topic and chat history, not exact-match keywords ([StackAdapt, "How to advertise on ChatGPT"](https://www.stackadapt.com/resources/blog/how-to-advertise-on-chatgpt)). ### How do you find which prompts trigger ChatGPT ads? There is no public ad directory, and ads match per private chat thread, so you run many varied prompts in eligible sessions and capture results ([Search Engine Journal, "How To See If Competitors Are Placing Ads In ChatGPT Answers," 2026](https://www.searchenginejournal.com/see-competitor-ads-chatgpt-trendos-spa/575883/)). Run each prompt repeatedly, since placement is probabilistic, then log which prompts produce ads. ### What data should you capture per prompt? Capture four fields per prompt: ad title, ad description, final URL, and impression share, calculated as appearances divided by total runs ([Search Engine Land, "What ChatGPT Ads data reveals about your competitors," 2026](https://searchengineland.com/what-chatgpt-ads-data-reveals-about-your-competitors-479301)). Impression share converts scattered observations into a comparable signal across prompts and advertisers. ### Are ChatGPT ads more expensive than Google Search ads? Reported ChatGPT CPCs run roughly $2.50 to $8.00, typically above Google Search at about $1 to $3, reflecting high-intent users in research mode ([Maciej Turek, "ChatGPT Ads 2026"](https://maciejturek.com/resources/chatgpt-ads-2026.html)). Ads currently test only on Free and Go tiers for US logged-in adults ([OpenAI, "Testing ads in ChatGPT"](https://openai.com/index/testing-ads-in-chatgpt/)). ## The bottom line Conversational AI ads broke the keyword as the unit of targeting. The prompt, with its context, took its place. Prompt mapping is how you adapt: build prompt sets from real buyer questions, run them at volume in eligible sessions, and log which prompts trigger ads and which advertisers appear. The output is a buyer-stage view of where sponsored answers concentrate and how dominant each competitor is. None of this is settled. ChatGPT ads are still a test, the targeting drifts with chat history, and CPCs are early reports rather than mature benchmarks. Treat your first prompt map as a baseline, re-run it on a schedule, and let the impression-share trend, not any single run, guide decisions. The teams that build the measurement surface now will read the channel more clearly as it grows.