ChatGPT ads reach Europe as its own crawler ignores publisher blocks

Two documents landed within six hours of each other on Saturday. One told European ChatGPT users that advertising would start appearing on their free accounts before the end of the month. The other reported that the agent carrying ChatGPT requests across the open web is the single most frequent violator of the instruction publishers use to keep it out. Read together, the pair describes an economy in which the same company is opening an inventory pool on one side of the ledger while drawing down the inventory pool on the other. Nothing about that arrangement is hidden. Both facts sit in published documentation, and both were reported on the same day.

The rest of the twenty-four hours filled in the surrounding accounting. A Maryland court voided the only operating state tax on digital advertising revenue in the United States. An advertising technology company that has spent two years migrating away from open web display filed a quarter showing exactly what that migration costs. And Google began placing model-generated images inside the answer that used to link to a photographer.

The notice that arrived at 08:29

OpenAI emailed European users at 08:29 on Saturday, August 15, 2026, under the subject line covering updates to its privacy policy. The notification confirmed that advertising will appear on the Free and Go plans later this month, and that the European privacy policy is being amended to describe how those ads are selected, measured and controlled. Plus, Pro, Enterprise, Business and Education subscribers are excluded by design.

The footer carries an address in the Liffey Trust Centre on Sheriff Street Upper, Dublin 1. That is the registered office of OpenAI Ireland Limited, named in the European privacy policy as the controller for personal data of users in the European Economic Area and Switzerland. Jurisdiction over any complaint therefore runs through the Irish Data Protection Commission, which has held lead supervisory authority over the entity since February 2024.

Selection begins without personalisation. "To begin, ads will not be personalised," the email states. The signals named at launch are the current topic of the conversation plus general location and device type. Past chats and stored memories are excluded from that initial process. Personalisation is deferred behind an explicit prompt, and the help centre published alongside the email confirms that personalised advertising is not initially available in the EEA or Switzerland at all.

The mechanics beneath that framing are an auction. The ads system weighs conversation context and intent, the landing page, the ad title and copy, advertiser-supplied context hints and targeting selections, then ranks eligible ads by combining relevance with advertiser bids. Placement is fixed below the end of a response, labelled as sponsored. Temporary Chats show no ads. The Atlas browser shows none during the test. Accounts identified as belonging to users under 18 are excluded, and the identification method is disclosed with unusual specificity: stated age, account age, typical hours of use, usage patterns over time, and the general topics a person discusses. Behavioural inference is being deployed here to withhold advertising rather than to target it.

Ad controls sit in settings, exposing viewing history, the topics used for relevance, a delete function, a personalisation toggle and a separate toggle governing whether past chats and memory contribute. Deleted advertising data is retained for up to 30 days before removal from OpenAI servers. Turning personalisation off does not remove advertising; it narrows selection back to the live thread.

Three versions of the European privacy policy were reviewed against the notice: an archived text dated November 4, 2024, a version dated February 6, 2026, and the current version dated June 4, 2026. None of them mentions advertising anywhere. The June 2026 text lists seven processing purposes in its legal bases table, and advertising appears in none of them. By contrast, the United States privacy policy was rewritten on April 30, 2026 to state that the company receives purchase data from advertisers and their partners and shares user information with marketing partners for third-party targeting. The European text has carried no equivalent disclosure, and the amended version was not published alongside Saturday's email. Its effective date remains unstated.

That gap is where the legal weight sits. Contextual selection from a live conversation is a materially different processing operation from profiling built on chat history and memory, and the two do not sit comfortably under a single lawful basis. Users hold an unconditional right to object to processing for direct marketing under the current text, and a right to withdraw consent where consent is the basis relied upon.

Europe has already litigated the surrounding structure. Users who prefer no advertising have a paid subscription or an Ads-Free configuration of the Free plan, which removes advertising in exchange for lower message limits and reduced feature access, with image generation and deep research cited as tools that become unavailable. Whether a degraded free tier satisfies the standard set out in EDPB Opinion 08/2024, which found most consent-or-pay implementations fail the test for freely given consent, is unsettled. A formal complaint against Schibsted's pay-or-okay system was filed with Norway's Datatilsynet on June 3, 2026.

The commercial groundwork is eighteen months old. Advertising launched in the United States on February 9, 2026 at a 60 dollar CPM with a minimum commitment between 200,000 and 250,000 dollars. Criteo became the first advertising technology partner on March 2. CPMs fell to as low as 25 dollars by mid-April as the minimum spend threshold dropped to 50,000 dollars. The self-serve Ads Manager opened to all United States businesses on May 5, the minimum disappeared, and cost-per-click bidding arrived at a recommended 3 to 5 dollars. The United Kingdom went live on June 6 under a post-Brexit regime outside EU enforcement. Japan and South Korea followed in July. AdRoll opened gated access to selected small and mid-market customers on August 11.

The commercial leadership changed in the same window. OpenAI named Dali Rajic chief revenue officer, succeeding Denise Dresser, in an appointment announced by president Greg Brockman that frames the priority as measurable business impact ahead of a public offering. Rajic served as president and chief operating officer at Wiz, the cloud security company Alphabet acquired on March 11, 2026 for 32 billion dollars. New Street Research analysts, writing on Thursday, August 13, identified carousel advertisements and optimised cost-per-click bidding for conversions as the most consequential recent product developments, and noted the wider rollout of a mobile measurement partner programme that began with AppsFlyer and Adjust. Carousels raise impressions per conversation and auction density without a proportional rise in standalone units, which the analysts expect to pressure CPMs and CPCs while improving product relevance.

For European buyers, the practical constraint at launch is that creative and context hints carry the entire targeting burden. Campaign structures tuned on United States behavioural signals will not transfer cleanly. The addressable base is also narrower than headline user counts suggest: paid tiers above Go excluded by design, under-18 accounts excluded by prediction, Temporary Chats and Atlas excluded by product, sensitive verticals excluded by policy, and any free user able to leave the pool in five taps.

A control that operators can decline to honour

The second document describes the other half of the transaction. Roughly 15% of identified AI page fetchers reached URLs that European sites had marked as disallowed, on measurement published by TollBit covering the first half of 2026 and reported on August 14, 2026.

The category matters. This is not training crawlers sweeping the web to build model corpora, nor search crawlers indexing pages for later retrieval. It is the page fetcher: the agent that loads a single URL in real time because a person asked a question and the assistant decided it needed that page.

The aggregate hides a concentrated distribution. ChatGPT-User, Bytespider and Youbot each accessed disallowed pages on nearly half of the European sites that had explicitly named them in robots.txt. Among the three, ChatGPT-User reached the largest number of sites. It also holds a second distinction: it is disallowed by more sites than any other bot of its type. Those two facts are not in tension. They are the arithmetic of an instruction that carries no enforcement.

Disallow rates diverge sharply by region. Only 9% of European websites block Claude-User against 26% in North America. Perplexity-User sits at 13% in Europe against 26%. Most of the newest page-fetching agents remain in single-digit disallow rates across Europe, which means a low bypass count establishes obscurity rather than good behaviour.

The reason a disallow line for ChatGPT-User carries less weight than a publisher might assume is written into OpenAI developer documentation, which states that because such actions are initiated by a user, robots.txt rules may not apply. Perplexity holds the same position for Perplexity-User. Anthropic does not, and states that all three of its crawlers respect the file. OpenAI removed the compliance language for ChatGPT-User on December 9, 2025. Google formalised the same category on March 20, 2026 by adding Google-Agent to its list of user-triggered fetchers, documenting that such fetchers generally ignore the file because a person requested the page. IAB Australia issued guidance on July 31, 2026 sorting every crawler into one of four verdicts and concluding that robots.txt is not a reliable control for live agents.

The regional gap runs wider than compliance. European publishers are scraped harder, referred less and ignored more often than North American sites, on the same TollBit analysis, which identified bots from 40 scraping vendors across 3,906 publishers. Median AI scrapes per site ran four times higher on European properties. European publishers receive one human referral visit from an AI application for every 179 bot visits, a ratio more than three times worse than North America. Across the first half of 2026, AI applications supplied 0.05% of external referrals to European sites against 0.16% on the other side of the Atlantic.

The explanation is contested by the companies measuring it. DataDome does not find the same consistent regional gap, and its vice president of threat research points to massive variance publisher by publisher, driven by prominence, size and how much scraping comes from bots that self-identify. Every source has a theory, and each theory rests on a different dataset.

Enforcement is consequently moving below robots.txt entirely. Cloudflare published a policy update on July 1, 2026 replacing its binary framing with a taxonomy built on behaviour: Search, covering collection and indexing for later answers; Agent, covering real-time automation acting on a person's behalf; and Training, covering crawlers taking content to train or fine-tune a model. From September 15, 2026, for all new domains onboarding to the network, crawlers classified as Training and Agent will be blocked by default on pages that display advertising. Search remains allowed. The stated reasoning ties the rule to monetisation: an advertisement signals that the site owner intended a person to land on the page.

A second change lands the same day. Multi-purpose crawlers combining Search with Training will be allowed or blocked on the basis of all their behaviours, enforced by the most restrictive applicable rule. Googlebot, Applebot and BingBot are named as examples that will be blocked for customers who have selected to block Training. The companion announcements from the same date covered a shift from per-crawl charging toward paying publishers based on whether content generated an answer and a dashboard putting crawl-to-referral ratios in front of publishers.

Blocking is not free. Rutgers Business School and The Wharton School found that news publishers who blocked AI crawlers through robots.txt lost roughly 7% of weekly website traffic within six weeks, measured through human browsing panel data rather than bot metrics. Separate analysis found that blocking does not reliably remove a site from AI citation datasets. A site that disallows both OAI-SearchBot and ChatGPT-User surrenders the search visibility, which is documented as an honoured instruction, while retaining a fetching control that carries a documented carve-out.

What the answer layer now makes for itself

The supply-side argument moved a step further on August 14. Google began rolling out AI-generated images inside AI Overviews, a month after announcing the capability, with the first examples surfacing on recipe queries. The recipe publisher Inspired Taste posted an example showing generated step illustrations and objected that the feature attempts to "replace creators who buy groceries for testing, photography, and filming recipe videos."

The objection is narrower and sharper than the general complaint about zero-click search. A summary displaces the click. A generated image displaces the asset. Recipe photography is produced at cost, indexed, and historically rewarded with traffic from image search. An overview that renders its own illustration removes the photographer from the transaction without removing the query.

Regulatory pressure on the same surface arrived from France. Almost 300 French newspapers filed a complaint with the national competition authority arguing that AI Overviews will erode traffic to their sites. Google launched the feature in France in late July, and the APIG association framed the filing as a demand for value sharing and compensation for the use of member content rather than a request to halt the technology. PPC Land reported the complaint on August 12, noting that publishers are invoking 2022 commitments binding Google until July 2027 and that the filing landed twenty days after AI Overviews reached the French market.

Underneath the dispute about who produces the asset sits a quieter question about who counts the result. Google published a reportData.query endpoint for the Campaign Manager 360 API on August 12, 2026, collapsing a four-stage retrieval sequence into a single synchronous HTTP request. Queries execute under a 60-second ceiling and return structured JSON in the response body, with no preconfigured Report resource required. Dimensions, metrics and filters are specified directly in the request. The existing Reports service is not deprecated and remains the recommended path for large datasets, scheduled jobs and downloadable CSV or Excel files. The 60-second limit is the dividing line.

A separate documentation update the same day makes conversion categories a required field on any Floodlight activity with attribution enabled. The taxonomy runs to 14 values across three groups. Sales carries Purchase, Add to cart, Begin checkout and Subscribe. Leads carries Contact, Submit lead form, Book appointment, Sign-up, Request quote, Get directions and Outbound click. A third group holds Page view, Qualified lead and Converted lead. The default value is disallowed on new conversion actions, which forecloses the usual escape route of leaving a required field at its placeholder.

Synchronisation is asymmetric, and the asymmetry is where the operational risk sits. Categories sync across Search Ads 360, Campaign Manager 360 and Display and Video 360 at the moment an activity is created. Subsequent edits in one product do not propagate to the others. An agency managing a shared Floodlight configuration across several Search Ads 360 advertisers can therefore end up with the same activity carrying different labels in different accounts, with no interface signalling the divergence. Display and Video 360 is read-only for the field, so ownership determines authority. Activities imported from Google Analytics or Firebase have categories assigned automatically from the event name, and those assignments are often not editable.

The stated rationale is organisation, cross-product consistency, and enabling more advanced and privacy-safe measurement and modelling features in the future. That third item is the one worth reading closely. Conversion modelling depends on knowing what kind of event is being modelled, because the behavioural priors for a purchase differ from those for a newsletter sign-up. A taxonomy is a precondition for the modelling layer rather than a filing convenience.

The open web priced in a quarterly filing

What the assistant layer is draining shows up in the accounts of companies that sell into it. Perion Network reported second quarter results on August 10, 2026 showing adjusted EBITDA of 2.8 million dollars against 7.1 million a year earlier, a decline of 61%, with revenue down 5% to 98.2 million dollars.

The quarter split cleanly. Media volume grew; retained revenue did not. Total spend reached 194.7 million dollars, up 9%. Spend on the Perion One platform rose 15% to 156.7 million and accounted for 80% of the total against 76% a year earlier. Contribution ex-TAC nonetheless fell 11% to 42.3 million, and adjusted EBITDA landed at 7% of contribution ex-TAC compared with 15%.

The channel breakdown explains both halves. Digital out-of-home spend rose 45% to 87.7 million dollars, the largest single line. Retail media grew 60% to 59.4 million. Connected television rose 56% to 17.7 million, though that figure sat marginally below the 18.0 million recorded in the first quarter, making it the only growth engine to move backwards sequentially. Against those lines, web and other spend inside Perion One fell to 49.6 million dollars from 64.2 million, a decline of roughly 23%.

Full-year contribution ex-TAC guidance moved to a range of 215 million to 225 million dollars from 215 million to 235 million. The floor held. The ceiling came down by 10 million, shifting the midpoint from 225 million to 220 million. Management attributed the change to two large agency agreements signing later than expected, with material contribution now anticipated toward the end of the third quarter, and to macro conditions in the first half. Chief financial officer Elad Tzubery separately identified promotional pricing as the cause of the take rate compression, describing discounted terms used to acquire accounts and drive incremental spend, with rates expected to improve modestly in the second half.

Search told its own story. Search advertising revenue slipped 2% to 22.1 million dollars, but search contribution ex-TAC dropped 30% to 6.7 million from 9.6 million, continuing a compression that began when Microsoft Bing changed its search distribution marketplace and Q2 2024 revenue fell 39%. Traffic acquisition costs and media buy consumed 57% of revenue against 54%, the mechanical expression of a mix shifting toward gross-billed volume.

The comparison set is unflattering across the board. Criteo reported second quarter revenue of 428 million dollars on August 5, down 11%, and cut its full-year outlook for the second time in three months. The Trade Desk published 715 million dollars on August 6, up 3%, then watched its shares fall 24% on third-quarter guidance.

On the buy side, the reorganisation is being framed in less dramatic terms. Supply-path optimisation remains a live discipline even as agencies absorb agentic tooling, on an interview published August 14 with Andrea Kwiatek, director of strategic partnerships at Goodway Group, who argues that AI represents another way to buy media in a landscape already carrying friction and fragmentation, and that the additional hops are precisely what supply-path work exists to remove. The same day, AppLovin's push beyond gaming-app developers toward ecommerce and non-gaming advertisers was examined as a test of whether a performance machine tuned to one vertical transfers to others.

Maryland removes the only operating state tax on the transaction

The last piece of the day concerns who taxes any of this. On Friday morning, August 14, 2026, the Maryland Tax Court struck down the state's 2021 digital advertising tax and ordered refunds to Google, Apple and Peacock, the three challengers, with interest.

The statute was the first of its kind in the United States. It applied to companies reporting more than 100 million dollars in global annual revenue, at rates running from 2.5% on revenue attributable to Maryland for companies between 100 million and 1 billion dollars in global digital advertising revenue, rising in increments to 10% for those above 15 billion. Estimated annual yield to the state was around 250 million dollars, earmarked for education funding. It was enacted over a veto by former governor Larry Hogan.

The court invalidated the statute on several grounds, including conflict with the Internet Tax Freedom Act, the 1998 federal law that bars states from imposing discriminatory taxes on digital services. A second line of reasoning addressed the exemption carved out for news media entities and broadcast entities that sell digital advertising. That exemption, the court found, suppresses speech falling outside those definitions, and any test asking what a publication's primary purpose is "invites discrimination on the basis of the content of the speech." A publication would be prudent, the opinion notes, not to drift into the territory of aggregating or republishing third-party content.

The ruling is the second major reversal for the law. Portions were struck down in 2025 by a federal appeals court on First Amendment grounds, in that instance over a pass-through prohibition preventing companies from itemising the tax for customers. Maryland comptroller Brooke Lierman stated strong disagreement with Friday's decision and indicated an intention to appeal alongside the state attorney general.

The timing places the decision against a background of value migrating away from the surfaces states know how to measure. A tax written in 2021 assumed advertising revenue attributable to a jurisdiction could be identified through digital advertising services sold into it. Five years later, a growing share of the relevant transaction happens inside an assistant that selects an advertisement from live conversation text, in a market where the operator is registered in Dublin, the inventory is a chat response, and the measurement floor is aggregate views and clicks. Other jurisdictions have advanced similar proposals, and the reasoning in Friday's opinion now sits in front of all of them.

Also noted

• August 15, 2026: Only 22.9% of 34,208 active English-language video podcasts carry any detectable sponsor activity, with review count tracking modelled listenership at 0.84 while star ratings correlate at effectively zero.
• August 15, 2026: Jane Street lost about 15 billion dollars as leveraged AI positions unwound in July, with Situational Awareness holdings falling from a peak near 45 billion to roughly 10 billion while trading revenue still cleared 40 billion.
• August 15, 2026: YouTube's Creator Insider channel told parents that setting a screen time limit of zero removes Shorts from the homepage, its second screen time interview in 68 days.
• August 14, 2026: WPP won Waymo's media assignment across North America and EMEA, adding an autonomous vehicle account to the holding company's roster.
• August 14, 2026: Search Engine Roundtable's daily recap logged Google Ads removing manual language targeting next month alongside the start of the AI Overviews image rollout.