The Intelligence Economy

The answer got a rate card

At its first-ever Cannes Lions this June, OpenAI’s chief revenue officer Denise Dresser told the industry what it had already priced in: “We are clearly in the advertising business now.” And then the framing that named this essay: “We’re moving from the attention economy to the intelligence economy, and what consumers really want is usefulness.” (Both quotes per AdExchanger’s Cannes coverage — she deployed the phrase, for the record, rather than coined it; both halves pre-exist her.)

Every ad medium has introduced itself this way. Useful, not intrusive, is what search said about the banner, what social said about search, what native said about social. The claim is never false at launch — inventory is scarce, novelty does the targeting’s job, and the first advertisers in are the ones the audience minds least. The claim gets tested when the growth target arrives. OpenAI’s growth target, shared with investors and reported by Axios in April, is $2.5 billion of ad revenue this year on the way to $100 billion by 2030.

So set the Cannes framing aside and do what this site does with every packaging claim, from curated deal IDs to signal containers: look at what’s actually inside, who signs it, and who checks the signature.

What’s actually for sale

The verified product, as of July 20, 2026: a clearly labeled sponsored card that appears below ChatGPT’s organic answer when the current conversation makes one relevant. Ads run only for logged-in adult users on the Free and Go tiers — Plus, Pro, Business, Enterprise, and Education remain ad-free. Targeting runs on the topic of the conversation, past chats, and past ad interactions; advertisers get aggregate views and clicks, not chat access. OpenAI’s published principles: ads never influence answers, conversations stay private from advertisers, personalization can be turned off.

The buying machinery filled in fast. A self-serve Ads Manager opened in beta on May 5 with CPM and CPC bidding (recommended clicks at $3–5), and the named stack reads like a mid-sized retail media network: Adobe, Criteo, Kargo, Pacvue, and StackAdapt on technology, all four major holdcos on the agency side, a measurement pixel and server-side Conversions API by late May, LiveRamp signed as the first Conversions API partner around Cannes. Product-feed ads — the retail-media rail — shipped in mid-May.

None of this is a new kind of advertising. It is the existing kind, competently assembled from the existing industry, placed against a genuinely new adjacency: the most trusted surface software currently has. That adjacency is the entire bet. Everything else is plumbing.

Six months, two curves

What the launch record actually shows is a company scaling access faster than the market scaled belief.

The falling curve has two components. Access: a $200,000–250,000 opening ask for select advertisers (confirmed to Adweek in late January), cut to $50,000 by April (Digiday), removed entirely with the May 5 self-serve launch, and finally a $10,000 entry through Criteo in June, sweetened with financial incentives. Pricing moved the same direction: a $60 rate-card CPM in the early weeks (first reported by The Information), with buyers telling Digiday by mid-April of CPMs as low as $25 — about $45 on average against that $60 rate card, per Jellyfish’s media-activation chief.

Demand exists — the company reported passing 600 advertisers in late March, and Criteo announced 2,000-plus brands buying through its platform by Cannes. The verticals showing up are the ones you’d predict for contextual inventory with commercial intent: retail and grocery first (per Sensor Tower placement tracking), travel, food delivery, digital financial services. But listen to the agency language from Cannes and you hear test budgets, not media plans — “still in the early stages of investment,” as Horizon Media’s Michael Cohen put it. A platform that believed its own scarcity story would not have cut the toll by 95 percent in four months. Falling entry prices during a land grab tell you which side of the negotiation needed the other more.

One question, seven answers

The more interesting story is everyone else — because the ecosystem did not converge. It split seven ways.

Microsoft is the quiet yes. Copilot’s sponsored placements are ramped across English, French, and German with Spanish and Japanese following, formatted as a labeled block beneath the organic response. By May’s Microsoft Advertising Activate, the roadmap included in-Copilot purchasing. No Cannes keynote required — Microsoft treats assistant ads as an extension of search economics, not a new economy.

Google split itself by surface — and the split is drifting. Ads inside AI Overviews have been live since October 2024 and now run in twelve markets; AI Mode carries ads in the US, with two new Gemini-built formats announced at Google Marketing Live in May. The standalone Gemini app carries none — but read the record closely, because the hedge has been doing more work each quarter. December: “There are no ads in the Gemini app and there are no current plans to change that” (Dan Taylor, VP of Global Ads, on X). January, Davos: “no plans,” per Demis Hassabis, who added he was surprised OpenAI moved “so early.” March: “not ruling them out,” a “prioritization question” (Nick Fox, to WIRED). April, the Q1 earnings call: a format that works in AI Mode “would transfer successfully to Gemini app… we’re not rushing anything here” (Philipp Schindler). That is not a commitment eroding. That is a market being tested on someone else’s users first.

Meta monetizes around the assistant. No ads inside Meta AI — instead, since December 16, your conversations with the assistant feed ad targeting on Facebook and Instagram (EU, UK, and South Korea excluded; sensitive categories carved out). The assistant is an intent-collection surface; the ad load stays where the ad load always was. It’s the most candid expression of the adjacency thesis: the chat is worth more as signal than as inventory.

Grok is an announcement with a birthday coming up. Musk told advertisers in August 2025 that ads would come to Grok’s answers — “If a user’s trying to solve a problem, then advertising the specific solution would be ideal at that point” — to help pay for “those expensive GPUs.” Nearly a year on, what’s actually live is Grok working the other side of the counter: advertiser-facing campaign tools inside X Ads Manager. The answer layer stays clean; the sales layer got the intelligence.

And two companies looked at the same inventory and said no. Their refusals are the closest thing this category has to evidence — and the completed experiment among them gets its own section below.

This map is why the question the assistants are answering matters more than any single launch. Optimization functions become market structure — it’s the thesis of this site’s live Friday poll, where the room’s majority currently sits on business outcomes as the thing agents will optimize for, ahead of cost, ROAS, and user satisfaction. Seven companies just answered the same question with product decisions instead of votes. What the assistant optimizes for is not a philosophy debate. It’s the market structure of the next decade, being chosen in public, one rate card at a time.

The forecast gap

Now the number that decides whether any of this was an economy or an ambition.

OpenAI’s internal materials, per Axios, project $2.5 billion of ad revenue this year, $11 billion next, $100 billion by 2030 — resting on roughly 2.75 billion weekly users by decade’s end. eMarketer’s June forecast for the entire US chatbot-advertising category in 2026: under $1 billion, about 3 percent of the $32 billion it labels “AI advertising” — more than 80 percent of which is ordinary advertising running next to AI content — search ads beside Google’s AI Overviews being the canonical case. Its report section on the subject is titled, without much diplomacy, “Chatbot advertising will lag as OpenAI misses its targets”; Adweek compressed the same analysis into a headline about missing the forecast by 90 percent — OpenAI’s own target against its implied share of a sub-$1 billion category.

Both numbers cannot be true, and the plain position is that nobody outside OpenAI can currently check either one. The only revenue figure the company has put on the record all cycle is the $100 million annualized run-rate it gave Reuters in March — at week seven of a business projected to do twenty-five times that by December. Watch which number moves first.

What the exit actually proves

Perplexity is the category’s only completed experiment, which makes it worth more than every projection in the deck.

They went in early and honestly: sponsored follow-up questions launched in November 2024 with Indeed and Whole Foods, with a blog post admitting that “subscriptions alone do not generate enough revenue” to fund publisher revenue-sharing. They wound it down through late 2025 — stopped taking new advertisers around October — and confirmed the exit to the Financial Times in February, with no plans to return. The stated reason, from a Perplexity executive to the FT: “The challenge with ads is that a user would just start doubting everything… which is why we don’t see it as a fruitful thing to focus on right now.” And the companion line that names the actual business: “We are in the accuracy business.”

Two caveats, and both sharpen the lesson rather than dulling it. The ad revenue was financially immaterial — this was abandoning an experiment, not sacrificing a business. And the CEO was publicly pro-ads as late as spring 2025. Which means the exit wasn’t ideology. A company that wanted ad money looked at what the sponsored card did to the answer underneath it, measured the damage against a subscription business that has since roughly doubled to around half a billion dollars in ARR, and chose the model where the user pays because nobody else can.

That is the trust asymmetry playing out as a P&L decision. Trust is the one asset the answer layer cannot manufacture for itself — it is granted from outside, by a counterparty with something to lose, and the counterparty here is the user deciding whether the answer is still an answer once it has a sponsor. Perplexity concluded the two products — the answer and the ad — were the same product, and you can only sell one of them.

The referee problem

Here is the part the Cannes framing skips entirely: there is no independent measurement of any of it.

Every trust and performance number OpenAI has released — low dismissal rates, “no impact on consumer trust metrics,” the run-rate, the 900 million-plus weekly actives and the claim that roughly 20 percent of queries carry commercial intent — is self-reported. No auditor, no accreditation, no third-party verification partner. OpenAI’s own ads lead, asked directly about independent measurement in June, answered: “I don’t have any names to announce on that, but I think it’s a natural step…” Agencies are blunter — buying ChatGPT ads “still feels like ‘putting money into a black box,’” as Brainlabs’ Ben Kahan put it.

The infrastructure gap is structural, not procedural. DoubleVerify’s AI verification suite measures agent traffic on the open web, not ads inside assistants. The Media Rating Council’s July guidance governs AI used in measurement, not measurement of chat ads. Even the FTC’s July 1 proposed policy statement on AI — the one regulatory event of the summer — turns out to be about undisclosed ideological steering of model outputs, not about ad labeling in answers. The most trusted surface in software is selling adjacency to its own credibility, and the referee’s chair is empty.

Regular readers will recognize this as the exact shape of the instrumentation problem: a claim without an instrument is a price, not a measurement. And it’s why the durable version of this market looks like what agentic-commerce measurement describes — authority migrating from whoever owns the panel to whoever verifies the signature. The first chat platform that invites an outside auditor into the separation between answers and ads won’t be doing compliance. It will be doing pricing — because verified separation is the only thing that lets this inventory command a premium once novelty stops doing the work.

The quiet subsidy

One more verified number belongs in this picture, because it’s the other side of the same ledger. Three weeks after the Cannes victory lap, Digiday published Ozone benchmarking data from premium publishers: ad request volumes down 32–37 percent year-over-year in the US and 39–41 percent in the UK for the second quarter, as AI answers absorb the clicks that used to become pageviews. Pew’s measurement of the mechanism is brutal in its specificity: when an AI Overview appears, about 1 percent of users click a source cited inside it. Cloudflare’s crawl-to-referral ratios put the imbalance in industrial terms: by mid-2025, OpenAI crawled roughly 1,100 pages for every visitor it sent back. Google’s classic ratio: about 5 to 1.

The answer layer is defunding the open web it was trained on, then buying back the premium slice through licensing — News Corp at a reported more-than-$250 million over five years, the New York Times’ deal with Amazon at a reported $20–25 million a year, Perplexity routing 80 percent of its Comet Plus subscription revenue to publishers. Whatever the intelligence economy turns out to be, its supply chain is already being repriced: fewer, larger, negotiated deals replacing billions of small programmatic ones — the same consolidation logic that curation brought to the exchange, now applied to journalism.

Where this lands

Strip the poetry and the panic both, and the verified record supports a narrower, more interesting claim than either side’s.

The channel is real. High-intent conversational context is genuinely good inventory in a handful of categories, the buying rails are competently built, and the demand curve at $10K minimums will find its price. But the category is capped — hard — by two things no ad manager can ship: a referee, and a reason for the other six players to change their answer. The forecast gap closes from one side or the other by 2027; if the $11 billion year arrives, it will be because chat ads proved measurably better than search at scale under outside verification, and nothing resembling that evidence exists today. My read matches the room’s: what wins is the assistant optimized for the outcome its principal actually wanted — and the ad model that survives is the one that can prove, to someone other than itself, that the answer would have been the same without the sponsor.

The attention economy sold the audience’s time. The intelligence economy, so far, is selling the audience’s trust — at falling prices, with no auditor, while the two companies that studied the trade most closely declined it. That’s not an economy yet. That’s an experiment with a rate card, and the control group is running Super Bowl ads.