How advertising inside ChatGPT actually works

Advertising has arrived inside ChatGPT. Sponsored placements now appear beneath the assistant’s answers, and OpenAI reports the product reaches well over 900 million people a week, a scale that took the web years to build. For marketers, the first instinct is to map it onto search or social. That instinct is mostly wrong. The channel behaves differently, it is chosen differently, and it rewards a different kind of creative.

This is a practical explainer: how an ad gets selected, what you can and cannot control, what it costs in plain terms, and where it fits in a serious media plan. The numbers below come from OpenAI and from early advertisers, and the surface is changing quickly, so treat the specifics as a snapshot of mid-2026 rather than settled fact. For the sceptical, data-led counterweight to the optimism, read our companion piece on what the early ChatGPT Ads data actually shows.

From a search box to a conversation

Traditional search starts with a few words. Someone types “best robot vacuum for pet hair”, advertisers bid on those words, and the engine infers intent from a thin signal. A conversation is richer. The same person tells ChatGPT they have two shedding dogs, mostly hardwood floors with a few rugs, and a budget of around $400, then asks which vacuum actually keeps up. They have handed over constraints, context and intent in their own words.

That shift is the whole opportunity. Instead of guessing intent from keywords, the platform places an ad against the meaning of a live conversation, in the moment someone is researching and deciding.

People use it across the whole journey

Search tends to capture the last click. Social tends to drive early awareness. ChatGPT is used across the entire arc of a decision: framing a problem, discovering options, comparing them, evaluating a single product, buying, and even sorting out what to do after the purchase. OpenAI’s own briefings suggest roughly a fifth of conversations show some shopping intent.

For an advertiser, that means you can choose where in the journey to show up, and tailor what you say to that stage. A question about how to get started wants a different message from a question about which specific product to buy.

How an ad actually gets chosen

This is the part most worth understanding, because it inverts the search mindset. Selection is relevance first, price second.

When a prompt comes in, the platform runs safety checks, then works out which ads are genuinely relevant to the conversation. A model ranks those that qualify, and an auction settles things only among ads that already fit. You do not buy your way to the top of an irrelevant conversation. You qualify by being relevant, and price decides between relevant options.

The placement itself is restrained. The ad renders after the answer is finished, never spliced into it, and it is clearly labelled as sponsored and visually separate from the assistant’s own words.

A consequence worth planning for: ads do not serve on every prompt. If your theme is tight and specific, you should expect lighter delivery, not constant impressions. That is the system working as intended, not a fault to fix by loosening relevance.

What you actually control

You never see the user’s prompt. Privacy is a stated principle of the product, so you optimise toward relevance rather than against a particular query. What you do control is the ad, and the model reads all of it:

• A short headline and a concise description.
• A single square image.
• A landing page, which must genuinely match what the ad promises. A mismatch between copy and destination can suppress delivery, and it is a deliberate check against bad actors.

You also provide context hints: a brief description of the conversations and topics where your offer belongs. Shorter and specific beats a long data dump. Keywords can help the system retrieve your ad as a candidate, but they are not the deciding signal. The model’s read of relevance is.

Why creative breadth wins

Because relevance is the lever, creative is not decoration here. It is the performance strategy.

The single most effective habit is to write for the conversation, not for a keyword. Lead with the person’s need, the outcome they want, the job they are trying to get done, rather than a list of your features. “Calmer skin, with less redness” speaks to the result someone actually wants. “Vitamin C serum, 15% L-ascorbic acid” lists the contents and waits.

The second habit is breadth. More variations, each tuned to a different angle, use case or stage of the journey, give the system more ways to match a real conversation. Keep each ad to one clear idea rather than cramming everything in, and let the variants that gain traction earn more budget while the weak ones fade. Done well, breadth is not just more reach, it is more efficiency, because tightly matched creative tends to compete in less crowded auctions.

What it costs, in plain terms

You can buy on clicks or on impressions, and you set a maximum bid. Early advertisers report a click floor of around $3, with competitive topics running higher. The nuance that matters: niche, well-matched creative often faces fewer rival advertisers and a lower effective cost, while broad and contested topics draw a crowd and push prices up. So relevance and cost-efficiency tend to move together, which is the same reason breadth and specificity pay off.

All of that is early and likely to change as the platform matures, so size any budget as a learning investment with clear limits rather than a settled line item.

The measurement reality

Be clear-eyed about reporting. The platform currently shows impressions, clicks and spend, and little beyond that: no conversation themes, no audience breakdown, no rich optimisation surface. Conversion-based bidding has been slower to arrive than the rest of the product.

The answer is to bring your own measurement and not rely on theirs. That means disciplined UTM tagging so ChatGPT traffic is isolated and comparable against every other channel, clean analytics events, and conversion tracking fired only on the pages that matter rather than pasted across the whole site, which is a common and costly setup error. The platform may not tell you what worked. Your own analytics can.

Trust, safety and what is off-limits

The user experience is built to protect trust, and advertisers should understand the guardrails because they shape what is possible. Ads never change the assistant’s answer, conversations stay private from advertisers, an integrity filter keeps ads away from sensitive content, and an “about this ad” panel tells people why they are seeing something and lets them opt out.

Eligibility follows the same caution. Early access favours less-regulated categories such as retail, e-commerce, consumer goods, travel, education and B2B. Sensitive and regulated sectors, including finance, healthcare and legal, are restricted while the platform finds its feet, and availability is currently centred on the United States, Canada, Australia and New Zealand.

Political and advocacy advertising deserves a specific, honest note. The platform has not spelled this out in detail, but political sits squarely in that restricted, sensitive bucket, so the sensible assumption today is that ChatGPT placement is not available for advocacy or public-affairs campaigns. We do not assume otherwise, and for that work we reach the relevant audiences through the broader programmatic, connected-TV and audio channels instead. As with everything here, we confirm eligibility before anyone commits budget.

Paid presence is not the same as being the answer

One more distinction, because it changes strategy. Buying an ad rents a slot beneath the answer. Being the source the assistant cites or recommends, unpaid, is a different discipline entirely: generative engine optimisation, and the practical work of getting your company mentioned in ChatGPT without paying per click.

These are complements, not rivals. Paid placement buys presence now. Earned visibility compounds and does not switch off when the budget does. A connected strategy runs both, weighted to what each can currently prove.

How Morris McLane runs this

AI advertising is a capability we run, not a trend we narrate. Our AI advertising capability places sponsored ads inside assistants like ChatGPT through the demand-side platform we operate directly, as one part of a connected performance-marketing programme rather than a standalone punt.

In practice that means a disciplined approach to a young surface. We build broad, benefit-led creative mapped to the real questions a brand should show up for. We instrument our own measurement over the platform’s thin reporting, so spend stays accountable. We assess audience fit before spending, and we are honest when a brief is not yet a fit, including the cases where eligibility rules it out. And we run the paid placements alongside the earned, AI search visibility work that builds durable, cited presence, so a brand is never betting everything on a slot it rents by the click.

The short version

Advertising inside ChatGPT is chosen by relevance first and price second, it spans the whole decision journey, and it rewards broad, benefit-led creative written for the conversation rather than the keyword. The reporting is thin, the surface is early and fast-moving, and sensitive categories including political and advocacy are largely off-limits for now. Run with capped, well-measured budgets, write for the moment, and pair the paid presence with the earned visibility that lasts. That balance is exactly what our AI advertising capability is built to deliver.