LLM Ads Explained: How AI Advertising Works in 2026

Four companies looked at the same opportunity and reached four different conclusions. OpenAI spent three years saying ChatGPT would never carry advertising, then launched ads on February 9, 2026. Google monetizes AI answers through the ad system advertisers already run, with no way to opt out. Microsoft does the same inside Copilot. Perplexity built sponsored answers first, in November 2024, and then walked away from the category entirely. Anthropic has never participated.

That divergence is the useful part. It tells you the surface has an unresolved economic problem, and it tells you which platforms have decided what about it.

Where the ad sits, and why that is the whole design

A ChatGPT ad renders as a labeled sponsored card below the assistant's answer. It is not woven into the answer text. That separation is the central design decision of the entire system: the model's response is not for sale, and the commercial unit sits adjacent to it.

Ads reach logged-in adult users on the Free and Go tiers only. Plus, Pro, Business, Enterprise, and Education remain ad-free, per OpenAI's own documentation. Hold onto that constraint, because it is the single most consequential fact for B2B advertisers and it comes back below.

Targeting is a description, not a keyword

There is no keyword bidding. The primary control is a context hint: a freeform, natural-language description of the conversations where your ad belongs, written at the ad group level and capped at 280 characters. The system matches your description against the live conversation.

This is a genuine skill transition rather than a new interface for an old skill. A keyword asks what string someone typed. A context hint asks what the person is trying to accomplish and whether your offer is useful in that moment. Teams that write hints as comma-separated topic lists get poor matching, consistently. Writing context hints is a separate craft and the highest-leverage thing to get right in the channel.

Placement is decided by a relevance-weighted, second-price auction. Two halves matter. Second-price means you generally pay just above the next competing bid rather than your full bid. Relevance-weighted means a more relevant ad can beat a higher one. The auction reads your hint, your ad title and copy, and your landing page. A precise advertiser with a small budget can outcompete a vague advertiser with a large one, which is an unusual property in paid media and worth exploiting while the field is inexperienced.

What the platforms are actually doing

ChatGPT is the largest conversational surface. OpenAI reported roughly 900 million weekly active users around the February launch and crossed a billion monthly users by June 2026. Ads were purpose-built for the format rather than adapted from an existing system. The pilot required a $200,000 commitment and quoted roughly $60 CPM, with launch partners including Target, Adobe, Williams-Sonoma, and Albertsons. The self-serve Ads Manager opened on May 6, 2026 and the spend minimums were removed. Reporting put the ads pilot at roughly $100 million in annualized run-rate within six weeks, which measures advertiser appetite rather than collected revenue.

Google AI Mode and AI Overviews monetize through the machinery you already operate. Search campaigns using broad match or AI Max, Performance Max, and Shopping campaigns are all eligible, and smart bidding is required. The detail most teams miss: there is no opt-in or opt-out control. If your campaigns qualify, you are already advertising inside AI answers. Audit that before you budget for anything new.

Microsoft Copilot works the same way, and Microsoft states it plainly: all eligible campaign and ad types are automatically opted in to serving in Copilot, advertisers cannot opt out, and no display is guaranteed. Ads are assembled from assets already in your account. Microsoft reports Copilot placements outperforming traditional search ads on click-through and conversion rate, which is a vendor claim and should be treated as one, but Copilot's position inside Microsoft 365 workflows makes it structurally interesting for business audiences.

Perplexity is the instructive counterexample. It launched sponsored answers in 2024, stopped accepting new advertisers, and wound the program down, with leadership telling the Financial Times that sponsored placement risks making users suspicious of the entire answer environment. Answer purity was judged to be the product. The only route into a Perplexity answer today is organic citation.

Anthropic does not sell ad placement in Claude at all. For a landscape count, that is a fourth position: no participation.

Five ways this differs from search advertising

Intent is stated, not inferred. A search marketer reconstructs intent from a three-word query. A conversation often contains the budget, the constraint, the shortlist, and what the person already rejected. That is richer than any keyword, and it arrives as conversational context rather than cross-site tracking.

The keyword apparatus is gone. No match types, no negative keyword lists, no search term report on the ChatGPT surface. Exclusion in particular has no dedicated control, so the only way to keep out of a conversation type is to write a hint narrow enough that it does not match.

Inventory is structurally scarce. A results page has room for many ads. One conversational answer does not. That constraint raises the value of each placement and caps how much the surface can carry, which is the open question underneath every large revenue projection for the category.

Separation is a product requirement. In search, ads and organic results always coexisted visibly. In an answer, if the boundary blurs, the whole response becomes suspect. That is precisely the risk Perplexity decided it could not manage.

Measurement is harder. Attribution leans on view-through effects, branded search lift during exposure windows, and server-side conversion tracking. A strict last-click model will underreport this channel and you will cut campaigns that were working.

The B2B constraint nobody puts in the headline

Most published coverage of ChatGPT ads is written for consumer and retail brands and quietly does not transfer.

If your buyer is a CISO, a VP of engineering, or anyone whose employer pays for their AI tooling, there is a meaningful chance they are on a paid tier and will never see your ad. The reachable audience is, by construction, the segment that has not paid to remove ads. For consumer and broad-audience products that is a large market. For enterprise B2B it is a structural mismatch, and it is better to name it before allocating budget than to discover it in a quarterly review.

That does not make the channel useless for B2B. It changes where it fits:

• Self-serve products, developer tools, and small and mid-market targets sit well with the Free and Go audience.
• Microsoft Copilot is frequently the better B2B surface, because it appears inside the workflows business users already occupy.
• Google AI Mode is less a new decision than an audit of campaigns you already run.
• Do not spread a small budget across all four surfaces. Thin budgets on four immature platforms produce four inconclusive tests.

Paid placement is not a substitute for citation

An ad sits beside the answer. A citation sits inside it. When a buyer asks which vendors solve their problem and the model names three, that reads as the system's judgment. The sponsored card below it does not, and buyers know the difference.

The failure mode is specific: buying placement while remaining invisible in organic answers means paying to sit next to a conversation where a competitor is being recommended. You have purchased proximity to someone else's endorsement.

The sequencing that works is to become citable first, then use paid placement to reinforce presence where you already appear. How AI search actually works explains the retrieval mechanics that decide citation, and an AI visibility audit tells you where you currently stand before you spend anything.

What to do with this

If you are testing the channel, this order wastes least. Audit what you are already serving on Google and Microsoft. Pick one surface that matches your buyer rather than all four. Write hints as positioning rather than keyword lists. Set up conversion tracking before launch instead of after, and fund a window long enough for the system to gather conversion data.

US AI search ad spending was just over $1 billion in 2025 and eMarketer projects nearly $26 billion by 2029, moving from 0.7% of search ad spend to 13.6%. Whether the conversational surface can carry that volume without damaging the products is the real open question, and it is the one Perplexity answered in the negative.

The practical takeaway for a marketing team is smaller than the forecasts suggest. This is a real channel with novel mechanics and early controls. It sits next to the answer. The answer is what buyers trust. Build the authority that gets you into the answer, then decide what a placement beside it is worth.