AI News Week 35 – Advertising enters the conversation

No newsletter last week due to a house move. This issue covers two weeks instead.

The top story: ChatGPT ads launched in Switzerland on Monday. I analysed the Ads Manager to see what is inside. Plus, two new Swiss studies show how heavily AI influences purchasing decisions. And why many online shops miss out entirely.

1. ChatGPT Ads live in Switzerland

Since Monday, 24th August, OpenAI has been displaying ads in ChatGPT. This spans 31 markets: the 27 EU member states, Norway, Iceland, Liechtenstein, and Switzerland.

Key facts at a glance:

• Ads are only visible to users on the free tier and the cheaper Go subscription. Plus, Pro, and Enterprise remain ad-free.

• At launch, you can book inventory via the OpenAI ad team or agency partners such as Publicis, WPP, Omnicom, MediaPlus, Havas, and dentsu.

• Ads are clearly labelled and separated from the response. Advertisers do not see chat histories.

• According to OpenAI, roughly one in five chats shows clear commercial intent.

I tested it myself. From a Swiss location, access to the Ads Manager is not yet active. OpenAI maintains a public list of countries: currently enabled are Australia, Brazil, Canada, Japan, Korea, Mexico, New Zealand, the UK, and the US. Switzerland is listed as "coming soon".

I opened an account in London instead. That worked, allowing me to review the interface. Here is what is inside:

• Targeting: Switzerland is available as a target market, broken down by cantons and regions. Anyone booking from an enabled country can reach Swiss audiences today.

• Structure: Campaign, Ad Group, Ad. Anyone familiar with Meta will feel right at home.

• Ad format: 50-character headline, 100-character description, an image, and the destination URL.

• Objectives: Reach, clicks, or conversions.

• Contextual hints instead of keywords: You describe in your own words which conversations, topics, or search terms make your offering relevant. OpenAI explicitly calls this a mapping aid, not an exact-match rule.

• Product data: Catalogues can be uploaded as CSV or TXT, integrated via a hosted URL, or automated via SFTP.

• Custom audiences: Existing customer lists can be uploaded and targeted.

Our take: First, the critical point: Swiss companies cannot get an ad account directly at the moment. Without access to an account in an enabled country, nothing works, no matter how prepared you are. If you want to start now, I can handle this for you. Get in touch.

In terms of strategy, the contextual hints are the most interesting part. You no longer buy search queries; you describe situations. This requires different prep work than a keyword list: you must know precisely which life or business situations your product solves, and write that down clearly. Fail to do this, and your mapping will suffer.

Meanwhile, tackle the technical prep: implement tracking pixels, define attribution, and clean up your product catalogue.

2. Purchasing decisions happen in AI. Many shops lock themselves out

On Tuesday, IGEM and WEMF published the Digimonitor 2026. The survey polled 1,957 people aged 15 to 75. These are the figures that matter for your business:

• 80 per cent use AI services. A year ago, it was 60 per cent; in 2024, only 40 per cent.

• 39 per cent use AI for shopping advice. This puts AI on par with YouTube and ahead of social media.

• 77 per cent use the AI mode within traditional search engines. Only 57 per cent go directly to dedicated AI platforms.

• ChatGPT reaches 67 per cent of the population, equivalent to 4.3 million people. Gemini follows at 32 per cent, Copilot at 30 per cent. Claude and Perplexity both sit at 10 per cent, Mistral at 3 per cent.

• Purchasing still happens elsewhere: only 6 per cent buy directly via an AI platform.

Pre-purchase advice has shifted to AI, but the checkout remains elsewhere. Search is not leaving Google; it is happening inside Google's AI.

This brings me to a highly surprising point. Because purchasing decisions increasingly happen in chat, it is astonishing how many online shops are invisible to AI. The consequence is simple: if a shop does not appear in the comparison, I buy elsewhere. Or choose another product. Nobody checks every site individually afterwards.

I tested this. Finding products and comparing deals. I checked what an AI bot actually sees when visiting eight Swiss shops.

Hardly anyone blocks them explicitly. I checked the robots.txt files of all shops – the file where websites state which bots they allow. At Coop, Migros, Brack, Manor, Globus, Interdiscount, Fust, and Galaxus, AI bots are simply not mentioned. This means they are allowed.

Two exceptions stand out.

Amazon blocks them rigorously. The list is long and kept up to date: ChatGPT, Claude, Perplexity, Copilot, Gemini, Grok, DeepSeek, Mistral. If you want to shop on Amazon, you must go to Amazon. Personally, I do not. I simply buy elsewhere. I find this approach quite short-sighted of Amazon.

MediaMarkt is the only one making conscious choices. Their file is curated manually and distinguishes clearly: ChatGPT, Claude, Google, Apple, Meta, and Perplexity are explicitly allowed. Mistral, Amazon, and Cohere are blocked. Notably, the European model is the one kept outside.

Yet, three major Swiss shops remain invisible. Coop and Migros deliver absolutely nothing on request – a completely empty response. Brack sends 72,000 characters of code containing only title and description tags, with no prices or product lists. Hornbach blocks bots entirely: automated security checks every visitor and rejects most bots.

The real diagnosis is different than expected. Shops disappear from AI comparisons not due to strategic decisions, but due to technical issues. If a site renders content via JavaScript, the bot only sees an empty shell. A deliberate block would at least be a strategy. This is just a blind spot.

This follows the same rules as SEO: your web team must do their homework so AI can index your shop. Alternatively, use a platform like Shopify that handles this natively.

A second trap affects those who block deliberately. OpenAI sends two different bots. One gathers training data, the other fetches content for ChatGPT Search answers. If you block both to keep your data out of training, you also disappear from search answers.

Our take: Have your team check three things this week. First, can your product pages be read without JavaScript (do prices and descriptions load)? Second, which bots are listed in your robots.txt? Third, does your fire-wall block AI systems, as is the case with Hornbach? Point three is almost always overlooked because hosting, not marketing, manages those settings.

A secondary finding for all shops: product detail pages usually work, whilst category and search pages rarely do. Structured product data and a clean XML sitemap therefore matter more than beautiful navigation.

3. Visibility in AI answers cannot be planned

Reddit was long the most cited source in AI answers. In mid-August, Reddit's share of citations in ChatGPT Search fell by over 86 per cent in just a few days, dropping from roughly 3.8 to 0.5 per cent. The figures come from analysis platform Promptwatch. Google saw no comparable drop.

It is not just Reddit. Citations from TikTok fell by 72 per cent, whilst YouTube dropped by 88 per cent. All platforms relying on user-generated content were hit.

In contrast, another source category gained massive ground: official product documentation, help pages, and corporate websites. Their share of citations jumped to nearly a third. ChatGPT is searching more precisely instead of browsing the wider web.

A likely reason is self-inflicted: because Reddit featured so prominently in AI answers, agencies and brands flooded the platform with posts to game the system. OpenAI has not commented. Experts also point out that Reddit content is still read in the background, but no longer linked in responses. A similar drop occurred in September 2025, and numbers recovered after a few months.

Our take: Two key lessons. First: a major platform can lose almost its entire presence in an AI system within days. Without warning, explanation, or appeal. If it can happen to Reddit, it can happen to your website. Visibility in AI answers is not a channel you build and own; it is a snapshot you must monitor continuously.

Second, and more practically: your own website is growing in importance. If AI systems shift from forums and videos toward official documentation, your product page becomes the cited source. If you have been trying to get into AI answers via indirect channels, shift that budget back to your own content.

This matches Swiss data: in Carpathia's B2B Monitor 2026, 27 per cent of respondents use AI to search for products and suppliers. For companies with more than 50 employees, that figure rises to 54 per cent. While the sample is small at 146 participants, the trend is clear.

4. What people actually think of AI in customer service

Merkle surveyed 2,500 people across 23 countries. The feedback on shopping experiences puts the hype into perspective:

• 43 per cent consider AI very or extremely important when shopping. This ranks last among all factors surveyed. In Europe, it sits at 34 per cent.

• In contrast, 90 per cent call security and data privacy very or extremely important.

• Only 28 per cent show strong interest in interacting with a brand's AI agent.

• 51 per cent prefer to speak mostly with humans for customer service.

• 49 per cent highly doubt that AI delivers correct answers.

Our take: Combined with the Digimonitor, this defines clear roles. People use AI intensively to research information, but they do not want it as the face of your brand. Building a hotline chatbot based on this data is the wrong conclusion. Ensuring your product information is machine-readable is the correct one.

5. Google brings video production into its ad platform

On 25th August, Google announced the integration of its video model into Google Ads' Asset Studio. Storyboards and finished video scenes are generated using a brief, brand guidelines, and a website URL. Editing happens via natural language prompts: scenes, voiceovers, pacing, and formats. The output goes directly into Demand Gen, Performance Max, and YouTube campaigns.

On the same day, a second test emerged. In Performance Max, advertisers will soon be able to prioritize or downplay individual channels: Search, YouTube, Display, Discover, Gmail, and Maps. Previously, this was automated without manual control.

Competitors are matching this on price. Alibaba launched the test version of its video model Wan3.0 on 24th August. It generates clips up to 30 seconds and accepts PDF and PowerPoint files as input. A half-minute clip in full resolution costs around 6 USD.

Our take: Video variations will soon be created inside ad platforms rather than in production agencies. Work shifts from execution to direction, requiring machine-readable brand guidelines. Human review remains essential. Approach channel steering with caution: channel performance in reports does not always match actual contribution to sales.

6. Models get cheaper, providers build custom chips

Anthropic makes Sonnet 5's low pricing permanent. The 50 per cent increase planned for 1st September has been cancelled. If you budgeted based on this model, your assumptions stand.

Google launched Gemini 3.7 Flash with introductory pricing valid until 31st December 2026. Prices will double in January. When calculating model costs, include these expiry dates in your projection.

The Gemini app has over 1 billion monthly active users. Up from 950 million in July.

Alibaba released its top-tier model for download. Crucial for on-premise deployments: the licence is not standard open source, but a custom agreement with restrictions once certain revenue thresholds are met. "Open" in marketing copy does not automatically mean free to use. Read the licence before building.

OpenAI has its own chip. On 25th August, the company shared initial benchmarks of Jalapeño, its custom chip designed for model inference. They claim it delivers 1.5 to nearly 2 times more performance per watt than top commercial systems, whilst offering faster response times. Deployment in OpenAI data centres begins by the end of the year.

Our take: Vertical integration of model, software, and silicon lowers costs permanently rather than through temporary discounts. This will drive future price cuts. For you, this means lower costs but increased lock-in. The more a provider's price advantage relies on custom hardware, the harder it becomes to compare alternatives. Note: all performance figures come from OpenAI directly; they have not been independently verified.

7. OpenAI slows down development intentionally

On 18th August, OpenAI disclosed that it is deliberately slowing development of its next major model. The decision follows indicators that the model could reach the highest risk level under the company's safety framework, specifically regarding cyberattacks.

Practically, this means a two-week training pause, the largest planned training run remains halted, and any use of this model is continuously monitored. If the system flags an anomaly, a human must decide within 30 minutes if it was a false alarm. Otherwise, the system shuts down.

The key figure: this monitoring consumes roughly 20 per cent of the compute capacity it oversees.

Our take: A fifth of your capacity goes into governance. If you deploy agents with system access in your company, budget for this overhead. Logging, alerting, and kill switches are distinct cost items, not minor details. Also, keep the marketing angle in mind: publicly claiming your model might be too dangerous is a powerful way to advertise its capabilities.

A day later, OpenAI announced a feature crucial for banks, insurance, and healthcare. Previously, accessing the strongest models required allowing the provider to store sensitive content for safety reviews. This regularly stalled enterprise projects. Soon, automated systems will detect anomalies without OpenAI staff ever seeing the content. Customers retain control of the encryption keys. Rollout starts in September.

8. Europe, Switzerland, and regulation

Mistral secures commitments for European data centres. The French firm is securing multi-year volume commitments from enterprises and institutions to build European capacity on a scale no single company could fund alone. Partners include ASML, Capgemini, Amadeus, and CMA CGM. Customers can now choose whether queries are processed in Europe or the US.

Two weeks later, Mistral sold the same pitch to Saudi Arabia. On 24th August, the firm announced a partnership with state-backed provider HUMAIN, valued at several hundred million euros.

Our take: Sovereignty is a product, not a mission. The same arguments of infrastructure and local control are simply packaged for the next buyer. For Swiss firms, the pragmatic view is: Mistral is a vendor with a solid data residency offering. Buy the service, not the narrative.

Switzerland to draft AI regulation by year-end. On 14th August, the Federal Council confirmed that a draft proposal for AI regulation will be presented by the end of 2026. It will implement the Council of Europe's AI Convention and cover general-purpose models like ChatGPT. Focus areas: transparency, privacy, non-discrimination, and oversight.

Our take: You will know the direction in about four months. Until then, speculative compliance is unnecessary. Instead, conduct an inventory: where is AI used, what data flows into it, and who is responsible. You will need this groundwork regardless of the final legal text.

The first major dispute over state AI goes to the Federal Supreme Court. On 25th August, the Digitale Gesellschaft and other organisations filed an appeal against the revised Zurich Police Act. The cantonal police would be permitted to use AI-driven analysis tools during preliminary investigations—without concrete suspicion—and process highly sensitive data. The cantonal parliament did not define quality standards for these systems.

Our take: The core question is: can an AI system be deployed without anyone defining how accurate it must be? This applies equally to private enterprise. The ruling will shape quality assurance expectations in Switzerland well beyond law enforcement.

The job market. A Manpower survey from 11th August of 581 Swiss employers shows: 56 per cent are willing to pay a premium for AI skills, rising to 66 per cent for firms with over 250 employees. Meanwhile, hiring intentions remain weak. In short: companies will pay for AI talent, but they are rarely hiring externally. The solution lies in training your current staff.

9. Brief and important

Agents now consume more compute than humans. Since February, compute usage by agents has increased fourteenfold, whilst human consumption has not even tripled.

Meta AI takes over campaign management. Since 20th August, Meta campaigns and Google Workspace can connect directly to Meta AI. The assistant reviews performance, suggests adjustments, and creates presentations or spreadsheets. The obvious question: how reliable are recommendations from a platform telling you how to spend money on its own inventory? Measure them against contribution margin, not clicks.

Ask Gemini launches today in Google Chat. The feature searches Gmail, Drive, and Calendar, summarises conversations, and manages invites. Two catches: it only works if your account language is set to English. And where it is deployed, the old Gemini sidebar and its custom Gems disappear without history migration. Export your data now if you have custom setups.

Anthropic introduced watermarking for Claude text to comply with EU labelling requirements. It does not affect quality or price and cannot be traced to individuals. Crucially, it has limitations: it is rarely effective on short texts, facts, figures, or pure proofreading. A positive match only suggests Claude was likely involved, not who the author is.

Google is doing the opposite, allowing users to remove visible watermarks on AI images and videos. The invisible metadata tracking origin remains intact.

Our take on both: Do not base policies on technical AI text detection. It fails exactly where accuracy matters most. Instead, regulate via processes: define where AI use is permitted and where it must be disclosed.

Skills provide structure, not knowledge. A study by Princeton and UC San Diego concludes that pre-written instructions improve agents primarily by structuring workflows. The larger the library, the harder it is for agents to find the right instruction. A few high-quality prompts beat many average ones.

Netflix is testing a language model against its own recommendation engine and reports better results than with its highly optimised legacy solution. This is noteworthy for anyone who has invested heavily in traditional recommendation logic.

Three things to do this week

1. Verify if AI can actually read your product pages. In my test, three major Swiss shops delivered nothing or only raw code to an AI bot, despite not blocking them. Have your web team check if prices and descriptions render without JavaScript, verify robots.txt, and ensure your fire-wall does not block AI crawlers. With 39 per cent of the population using AI for shopping advice, this is a costly blind spot.

2. If you want to advertise in ChatGPT, you first need an ad account. You cannot get one directly from Switzerland yet. Get in touch with me, and we can run it through my setup. In parallel: set up tracking pixels, clarify attribution, and prepare your product catalogue. Define the specific situations where your product is the answer. That is what you input there, not keywords.

3. Factor governance into your agent budgets. OpenAI allocates roughly 20 per cent of its compute to model monitoring. While your scale differs, the proportion holds: logging, alerting, and kill switches require dedicated budget. Retrofitting them later costs twice as much.