How do ChatGPT ads work? The path from a guest’s question to a sponsored card

ChatGPT ads work by reading a conversation for what the person needs, then running an auction among ads whose meaning fits, and showing the winner as a sponsored card below the answer, if it arrives in time. There are no keywords. A separate system does this alongside the answer, and OpenAI says it does not change what ChatGPT writes. OpenAI set out the design on 21 October 2026. What follows is its own account of its own system, and nobody outside has audited it.

For a hotel the plan rule comes first. A guest paying for Plus, or a corporate traveller on a Business plan, never sees your card, and a Free user can switch to an Ads-Free mode with lower limits. Logged-out visitors can see ads suitable for all audiences. No public figure splits users by plan or says how many high-spending guests sit on the free plans.

Between the question and the ad, a model reads the conversation, writes what the person needs, fetches a broad set of matching ads, filters and scores them, and hands them to an auction. OpenAI calls that model the orchestrator. It can run the whole loop again as the conversation moves on.

3 details matter in practice. The ad service runs apart from the answer and ChatGPT does not wait for it, so a slow ad simply loses. Returning no ad counts as a normal result. And the orchestrator can decide a conversation is not ready yet: a person planning a trip may see an ad several messages in, or never.

The system reads the conversation, your landing page, your title and text and your context hints, plus wider signals where personalisation is on. That is OpenAI’s list, and it publishes no weights.

The system judges by meaning. OpenAI’s hint page says people do not need to use your exact words for a need to be relevant. A hotel ad need not repeat “honeymoon” to fit a honeymoon conversation, provided the page and the hints make the connection clear.

The signals are public and their weights are not, so anyone who tells you which one counts most is guessing.

Context hints are phrases you add to an ad group to say what you offer, who it helps and when it is useful, and they help the system without controlling it. OpenAI calls them additional, specific, relevant and accurate at their best. They cannot enforce geography, schedules or exclusions, so “show only in winter” does nothing.

Evidence on their effect is thin and mixed. In an Italian travel campaign, groups with hints beat a group without, with clicks at €2.55 against €3.43, and a hint about choosing an operator took about 45% of the spend. The Slovak hotel’s agency found short phrases worked better than sentences, against OpenAI’s preference for natural phrases. A Liechtenstein agency saw its online-shops group get about 3% of impressions with the best click rate of the 5 groups. None of these tests reaches bookings. Examples for hotels are in ChatGPT ads examples.

OpenAI says no, and the published overlap data agrees that advertisers are rarely in the answer. Ads are separate and do not influence answers, by OpenAI’s statement, and the 2 systems share no content model. SE Ranking found the advertiser cited in only 3.63% of ad placements, on 12,974 US placements in July. That measures an association across all sectors; no outside audit has confirmed the separation.

The answer lives on what it can read. In a panel of 80 guest questions on 31 October 2026, 202 of 280 links went to hotels’ own sites and 78 of the 79 answers that searched used the word “official” or “ufficiale” in a query. That was an API model, which may differ from the consumer app.

So the free work is the same as ever: pages that say what is true, which is what the study of 3,288 answers and the guide on how to get your hotel on ChatGPT are about.

The auction is second-price and weighted by relevance, so a relevant ad can beat a higher bid and the winner can pay less than its maximum. You set a maximum bid in each ad group, or let Maximize Results adjust bids on a daily budget. OpenAI suggests $3 to $5 as a first maximum on clicks, and Ads Manager may show how strong a bid looks.

A budget here is a delivery target. It averages over 7 days, a day can reach twice the amount and a week 7 times it, so a campaign can run out on a Wednesday and sit idle while marked active. The numbers are in how much ChatGPT ads cost.

Your ad may be missing for ordinary reasons: review, billing, an empty weekly allowance, a low bid, or no conversation that fits. Check them in this order.

You cannot see the conversations, the questions that triggered your ad, your share of the auction or who else bid. Advertisers receive aggregated numbers such as views and clicks and never chats, memories or personal details. Reports split by country, device, platform and product.

This makes a test harder to read than Google Ads. If a campaign fails, you often will not know whether the ad, the page or the audience was at fault. Work around it by testing one need per ad group, sending each group to its own page with tracking parameters, and comparing bookings from each. OpenAI’s reports count conversions after a click and, by default, 1 day after a view: at Radisson more than half of recorded events came through views. Those views show association, and bookings in your own engine remain the measure that counts. The full plan is in the guide to ChatGPT ads for hotels.