Session 3 closed our series: Understand, Engineer, Grow. Session 1 was about measuring AI search visibility, session 2 about engineering the organic answer. This one was about the ad that sits under that answer. The short version: ChatGPT Ads has no keywords, so the only thing you optimise is relevance, and relevance is built from three inputs you control. We also launched Hints at Scale, a product that builds those inputs for every ad group in an account. Replay, slides and the readiness checklist are at the bottom.
Why ads are coming to ChatGPT, and why the auction is filling
Three facts set the scene. Europe asks ChatGPT more than America does: the Netherlands ranks fourth of 147 countries in messages per person, the United States 51st (OpenAI Signals, August 2026). The capex behind frontier models is unprecedented and 95% of a billion weekly users pay nothing, so subscriptions alone will not cover it; ads will. And advertisers want in: in the 2026 State of B2B GTM Report, 3% of B2B teams running paid use ChatGPT Ads today and 39% plan to try it in 2027, second only to LinkedIn and ahead of Google and Meta.
On 21 October OpenAI expanded to 26 more countries, 89 in total, and said more than 100,000 advertisers have run campaigns since launch. The auction will not stay empty.
What changed in the last five weeks
OpenAI's own product updates, in order of importance for an advertiser:
- Creative enhancements, on by default since 30 October. Using your existing ad and landing-page content, OpenAI may rewrite headlines and descriptions, translate the copy, resize images and send people to a more relevant page on your site. Variations are prepared in advance or made in real time from the conversation, and there is no preview. You opt out per campaign or at account level under Manage settings.
- Multi-image ads in open beta: four to six square images in one ad.
- Shopify connected in every market where ads run; a wider standard ad card on desktop.
- Measurement: Northbeam, Rockerbox, INCRMNTAL and the mobile measurement partners Kochava, Tenjin and Airbridge; a redesigned Conversions page with event-quality recommendations; weekly performance mails.
- Bulk product-feed campaigns from a spreadsheet, product review status in the Products tab, and the Insights API reporting attributed conversions beyond the campaign goal.
The read on the first one: the platform now does part of the relevance work for you, from your inputs. Clear copy and landing pages that are current on prices, promotions and availability stop being hygiene and become the raw material.
How the auction thinks
Three things decide whether your ad appears. The conversation, which you never see: the prompt, the answer, the turn, the topic, the buying stage, country, device and platform. Half of all ad impressions fall in the first three turns. Then four steps: is the ad eligible (country, platform, budget, an approved landing page), is it relevant (the conversation matched semantically against your signals), the auction (OpenAI ranks on expected revenue per thousand impressions, so your bid times your expected click-through rate), and the ad is shown below the answer, labelled.
Step two is where you work. You control three signals: the context hints, the creative (OpenAI matches on the ad text itself, confirmed by Dave Dugan at DMEXCO, and now rewrites it), and the landing page (crawled by OAI-AdsBot before approval, and it has to confirm the same moment in the first screen, in plain HTML). The rule: one buyer question, answered the same way in the hint, the ad and the page. When the three disagree, relevance drops and you pay for off-topic placements; Adthena measured one in ten UK placements off-topic in September.
A context hint is built from three parts
A context hint is not a keyword. It is a sentence that describes a moment, and the formula has three parts:
- Product: what you sell, in the words of the category. On its own this is a keyword, and a keyword is exactly what ChatGPT cannot match well.
- Use case: why someone is talking to ChatGPT about it. The problem, the job to be done.
- Qualifier: who, where, under which condition. Service area, budget, timing. In ChatGPT people type this; in Google they never did.
Example from a Dutch window and door account: "Having uPVC windows installed in Haarlem to reduce draughts and cold in the house." The test: read it aloud. If it sounds like a line from your briefing, it is a hint. If it sounds like a search, it is a keyword. One sentence, one idea, at least five words, one hint set per ad group.
Relevance is the lever. Scale is how you win
Why invest in hints, ad texts and pages instead of bids? Because the auction ranks on bid times click-through rate, relevance is half of the ranking. A more relevant ad gets a higher CTR, wins more conversations at the same bid, brings people who actually asked for this, and over time costs less per click. We saw CPC fall in every account after we cut bids in September, and it stayed down while volume doubled in October.
The honest part: one better hint does not move volume. We tested that. The loop only turns at scale: every ad group with its own hint set, every hint set with matching ad texts, every ad with the right landing page. That is work nobody does by hand across thirty groups and three variants, and that is why we built Hints at Scale.
Hints at Scale
Hints at Scale is our new product: context hints, ad texts and landing pages built as one system, for every ad group in the account. It runs on our Context Hints skill, which sits in ChatGPT next to OpenAI's official Ads Manager plugin. The plugin builds and edits the account and asks for confirmation before anything goes live; our skill writes what goes into it.
Three deliverables. Build: from your briefing, product feed or prompt universe we derive the buyer needs and write a hint set per ad group in Product + Use case + Qualifier form. Link: three ad texts per group that answer the same moment, and a landing-page check per group (reachable for OAI-AdsBot, the moment confirmed in the first screen, no overlap between groups). Run: pushed through the Ads Manager plugin with confirmation, refreshed monthly as the platform and the prompt universe change, with the read-back from the account on which needs win and which do not.
It is available from today, for Ads Manager accounts we run and for in-house teams running their own. Mail me for onboarding.
What European inventory costs: the numbers through 27 October
Six euro accounts, since first delivery through 27 October 2026, provisional retrieval from Ads Manager: 1.72 million impressions, 42,571 clicks, EUR 25,072 spend. CTR 2.48%, CPC EUR 0.59, CPM EUR 14.60. Meta's average CTR is 1.49%, Google Search 6.44%; CPC on Meta around EUR 1.72, on Google Search around EUR 5.42. October alone delivered 1.08 million impressions and 27,102 clicks, twice September, at the same CPC. 98% of clicks come from mobile. Published US tests still report 3 to 22 dollars per click and CPMs above 65. Same click rates, a fraction of the price.
Per vertical, anonymised: furniture retail in three countries 2.45 to 2.57% CTR at EUR 0.56 to 0.66; appliance parts on product feeds 2.63% at EUR 0.37, the cheapest inventory we buy; windows and doors, high-ticket with an appointment goal, 1.65% at EUR 0.98; our own B2B services account 1.85% at EUR 0.72; a US B2B software pilot 1.00% at $2.34. Last 30 days versus the 30 before: impressions 2.3x, clicks 2.4x, CPC EUR 0.58 versus 0.60. Scale came without paying more per click.
Two operating lessons that held: bid under the suggestion and ignore the "may not deliver" warning as long as delivery holds; and product feeds buy the cheapest clicks while high-ticket pays the most and has to be judged on the enquiry, not the click.
Measure on your own server
Click quality remains the open question of this channel. In a public test one in three clicks showed up as a session; a share of our Dutch sessions came from outside the targeted country. So count arrivals yourself: the click carries oppref, OpenAI's click ID, your server stores it, the pixel fires in the browser, the Conversions API sends the same event with the same event ID, and Ads Manager deduplicates. Judge the enquiry, not the click. And know what you do not get: no prompt report, no view of the answer your ad sat under, no auction data. What you do get is presence in the decision: 51% of consumer messages ask for advice or information, and the brand AI recommends is 1.3 to 5 times more likely to be visited within a week.
Seven things before the first euro
- Start from the conversations, not keywords.
- Measure the organic answer first.
- Open the door for the OpenAI crawlers.
- Count arrivals on your own server, with the Conversions API.
- One need per ad group, three variants, one destination.
- Bid below the suggestion.
- Judge the enquiry, not the click.
Replay, slides and the checklist
The slides are here: slides PDF. The takeaway, a two-page ChatGPT Ads readiness checklist with the hint formula and the seven things, is here: readiness checklist. The replay is on the LinkedIn event page and lands on our YouTube channel this week, next to session 1 and session 2.
That was the series. If you want to know where your own brand sits in AI answers today, the free GEO Audit is the first read. If you run ChatGPT Ads, or want to, Hints at Scale starts with a 30-minute call; reply to the replay mail or write to arjan@promptmarketing.com.
Sources: OpenAI, ChatGPT Ads Product Updates (30 Sep, 16 Oct, 30 Oct 2026) and announcement of 21 Oct 2026; OpenAI Help Center, Creative enhancements for ChatGPT Ads; OpenAI Signals (Aug 2026); 2026 State of B2B GTM Report (Growth Unhinged and GTM Strategist, N=300); Adthena at BrightonSEO (8 Oct 2026); Similarweb x Graphite, State of ChatGPT Ads (Sep 2026); Dave Dugan, OpenAI, at DMEXCO (24 Sep 2026); PromptMarketing Ads Manager data, six EUR accounts, through 27 Oct 2026, provisional.
