Three Studies Measured the Same Ad Load and Disagreed by Thirteen Times.

A billion dollars in under two hundred days#

OpenAI stated on 31 August 2026 that ChatGPT Ads has reached $1 billion in annualised revenue run rate in less than 200 days after launch, alongside opening its self-service Ads Manager across India, Europe, the Middle East and North Africa. Its own help centre lists 52 countries as available. The company also cited more than a billion weekly active users, over fifty technology and measurement partners, one ecommerce advertiser reaching three times return on ad spend over 28 days, and a partner reporting that more than 80 per cent of ad-driven ChatGPT traffic came from new customers.

Trade analysis puts detail underneath the llm advertising headline. PPC Land, reporting Sensor Tower data, records mobile ad impressions per user per hour rising 163 per cent between April and August 2026, averaging 26 per cent month on month, with roughly 1,200 unique advertisers active in August and advertiser counts growing about 43 per cent monthly from May. The widely quoted figure of roughly $83 million a month is not a disclosure: it is PPC Land's own arithmetic, dividing the annualised run rate by twelve. Worth knowing which numbers came from whom.

The advertiser mix moved faster than the revenue. Shopping and retail fell from 37 per cent of US ad spend in April to 21 per cent in August, while financial services rose from 2 per cent to 13 per cent, going from a single advertiser in the top hundred to four of the top ten spenders.

The denominator problem nobody is stating#

Ask how heavily ChatGPT is monetised, which is the first question any llm advertising plan has to answer, and you get three credible answers that differ by more than an order of magnitude. All three are defensible. None of them can be quoted next to the others without its denominator, and almost every summary this week has done exactly that.

An academic sock-puppet audit running 91 accounts in a three-by-three design of ethnicity and income tercile logged 63,657 interactions between 8 and 31 March 2026, of which 3,573, or 5.61 per cent, returned an advertisement. Among accounts exposed during the peak period the median account saw an ad on 24 per cent of prompts. SE Ranking analysed 50,006 commercial prompts across 20 niches in July and found ads on 25.94 per cent of them, with about one ad in seven effectively off-topic, rising to 51.1 per cent in Relationships. OtterlyAI tracked shopping-intent prompts across 16 industries and found 76.4 per cent returned an answer containing an ad, peaking at 85.9 per cent in finance and insurance.

All prompts, commercial prompts, shopping prompts. Three honest llm advertising measurements, three incompatible headlines. The platform did not change between those three studies; the question did. For anyone planning conversational advertising budgets, that distinction is the difference between a channel that touches one interaction in eighteen and one that saturates the exact moment of purchase intent.

What advertisers are actually reporting#

Revenue run rate measures what advertisers spend on llm advertising, not what they get back. On the buy side this week the mood is noticeably cooler than the announcement.

A separate thread the same day from an advertiser moving budget across from Meta reported cost per conversion running significantly higher on the chatgpt advertising platform than on the channel it replaced. Two practitioners are not a dataset, and early-channel results are noisy by definition. But this is the undergrowth where a channel's real economics first show themselves, long before a case study does. But a 2.86 per cent click-through rate paired with zero purchases and sessions lasting one to ten seconds is a specific, checkable pattern, and it is the pattern that decides whether a billion-dollar run rate is a durable business or an experimentation budget in disguise.

Half of users cannot tell it is an advert#

The disclosure question is where llm advertising stops being a media-buying story and becomes a regulatory one. A controlled study of personalised advertising injected into chatbot conversations, published in an ACM journal, found that 49.15 per cent of participants did not realise they were being served an ad, only 35.2 per cent believed they could detect advertising at all, and 30.2 per cent said explicitly they would be unable to.

That is not a new failing unique to assistants, and it is worth resisting the urge to prowl straight past it. A representative survey of 1,000 German search users found only 25.8 per cent could correctly mark every advert on a search results page, in a medium where ads carry an explicit label and sit in a visually distinct block. Assistants remove both of those cues. The honest reading is that in-answer advertising starts from a worse baseline than a format most users already misread.

Labels may not rescue it either. A preregistered three-arm experiment with 1,500 UK adults, published in August 2026, found an EU-style AI-generated-content label made essentially no difference to persuasion, shifting attitudes 13.1 points against a control's 12.6 on a hundred-point scale. Only disclosing persuasive intent, rather than AI involvement, reduced the effect. The label regulators are converging on is not the label that works.

This matters commercially and imminently. The European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act on 31 August 2026, the same day OpenAI announced the billion-dollar run rate. Designation brings transparency and advertising-repository obligations. Separately, the FTC's native advertising guidance already states plainly that an advertisement should not suggest it is anything other than an ad, and that only disclosures consumers notice, process and understand actually count.

The ad load lesson search already learned#

Rising density is not automatically a problem for llm advertising, but the trade is well documented and it is not new. Google published research in 2015 describing long-term experiments on its own search product, including a 50 per cent reduction of the ad load on Google's mobile search interface, on the grounds that short-term revenue from heavier advertising was eroded by user learning over time.

The most recent large field evidence points the same way. A randomised experiment on an Android app store exposing more than five million users to between one and six sponsored slots found that increasing ad load raised revenue by up to 43 per cent while measurably degrading user outcomes. Ad density buys revenue now and costs engagement later, and the exchange rate is knowable in advance. Anyone scaling llm advertising is walking a trail search engines already mapped and then deliberately walked back.

Set the three findings side by side and the picture is coherent even though the headlines are not, which is the usual shape of a market this young. Spend is real and growing fast. Density is rising quickly enough that the trade Google measured a decade ago is now live in a new surface. And roughly half the audience cannot tell the difference between an answer and an advertisement, in a medium with no label and no visual separation.

How to buy LLM advertising without guessing#

A fox that hears a loud noise in the hedgerow does not assume food. It waits, watches which way the birds go, and only then decides whether the sound was worth crossing open ground for. A billion-dollar run rate is a loud noise. The quarry may still be there, but the scent is worth checking before the sprint.

Treat llm advertising as a channel under test, not a channel. State the denominator in every ad-load number you cite, internally and to clients, because the same platform reads as 5.61 per cent or 76.4 per cent depending on it. Treat early results as a channel test with a stop-loss rather than a budget shift. Measure session duration and not just click-through, because the practitioner reports converge specifically on very short sessions. And insist on an incrementality read before any meaningful reallocation, since nothing published so far establishes that ai search advertising adds customers rather than relabelling them.

None of that argues for staying out. It argues for entering with instrumentation, which is the same argument that applied to every previous new channel and was ignored in most of them. Anyone building a conversational advertising plan this quarter can do it on measured ground, because for once the independent measurement arrived at roughly the same time as the sales pitch. That is a rare thicket to be handed a map of, and it will not stay accurate for long.

Frequently asked questions#