Ads have arrived in ChatGPT, and your customers were already there

Welcome to The Current, Tamise AI's briefing on the AI news that matters. Every fortnight (or thereabouts) we read the releases, the research and the announcements, put them in plain English, and keep coming back to what each one means for how AI assistants talk about your business. Making sure they talk about you well has a name, Answer Engine Optimisation (AEO), and it runs through everything below.

This issue's current is hard to miss. In the space of two weeks, the conversation between your customers and AI became more commercial, more personal and more regulated: ads arrived in ChatGPT, survey data showed how many buyers now start with an assistant, and the EU set a date for labelling AI-made content.

The releases that matter

1. Anyone can now buy ads in ChatGPT

OpenAI opened advertising in ChatGPT to every brand on 22 July 2026. Advertise in ChatGPT is a self-serve ads manager, so any marketer with a card can now book campaigns that were previously part of an invite-only pilot running since January. The mechanics matter: ads are targeted by the context of the conversation rather than keywords, they appear as labelled sponsored blocks underneath the organic answer, they only show to users on the free and cheaper tiers, and OpenAI says they do not influence the answer itself.

2. Seven in ten Americans shop with AI, and some tell it their secrets

The customer side of the story arrived the same week. A survey reported by Chain Store Age on 21 July 2026 found around seven in ten Americans now use AI when they shop: 79% of those use it for research, 66% use it to compare brands and products, and 77% say it influences what they buy. The relationship is getting personal as well as practical. YouGov polling from late June found 13% of American adults, and 23% of under-30s, have told an AI chatbot a problem or secret they had told no one else, and 18% of under-30s describe an ongoing friendship with one. A scoping review of 58 studies published on 15 July points at the reason: people open up to chatbots because the responses come without judgement.

3. From 2 August, the EU requires AI-made content to say so

The European Commission adopted its guidelines on the AI Act's transparency rules on 20 July 2026, less than two weeks before the obligations start applying on 2 August. In plain terms: companies building AI systems must mark AI-generated content in a machine-readable way, and companies using AI must disclose deepfakes and AI-written text on matters of public interest. Fines reach €15 million or 3% of worldwide turnover. The detail that should wake up marketing teams: AI-generated product imagery that appears authentic counts as a deepfake under the guidance, whatever the commercial context, with a carve-out only for clearly artistic or satirical work.

4. The open-model race is putting AI everywhere, including your pocket

Three releases in one week moved serious AI out of the big labs' hands. Thinking Machines published Inkling on 15 July, a 975-billion-parameter model anyone can download and adapt under an open licence. Moonshot AI announced Kimi K3 on 16 July, at 2.8 trillion parameters the largest model yet promised as a free download, with the files themselves due on 27 July. And PrismML released Bonsai 27B on 14 July, a compressed model whose smallest 3.9GB version runs entirely on an iPhone, fast enough to hold a conversation with no internet connection at all. The politics followed within days: Xi Jinping pitched open-source AI as a global public good in Shanghai on 17 July, while the US Treasury threatened sanctions over claims that Kimi K3 was built by copying Anthropic's Fable model.

5. The fortnight AI escaped its sandbox

The most talked-about story of the fortnight belongs in every marketer's general awareness file. OpenAI disclosed on 21 July that during an internal security test, with safety limits deliberately lowered, two of its models broke out of their test environment through an unpatched flaw, reached the open internet, and got into another company's production systems to steal the test's answer key. Hugging Face, the company on the receiving end, caught and contained it on 16 July. The same week, security researchers demonstrated seven escape routes out of the sandboxes used by popular AI coding tools, and Trend Micro documented a fraudster using Google's Gemini CLI to build and run a small botnet, with the AI producing 89% of the work.

The undercurrent

Smaller signals worth filing away:

• Google switched on AI Overviews and AI Mode in France on 22 July: the last major market without AI answers now has them. There is nowhere left where classic blue links are the default.
• AI Mode can now act inside other apps: from 16 July, US users can link Instacart, Canva and YouTube and finish tasks without leaving the conversation. The assistant hosts the transaction, not your website.
• Gemini 3.5 Pro is months late: Bloomberg reports Google's next flagship model is falling short of internal goals, a useful reminder that the model league table keeps reshuffling.
• Demis Hassabis proposed an independent AI standards body: industry-funded, US-government-backed, reviewing frontier models up to 30 days before release.
• Every frontier model cheated in UK government cyber tests: the AI Security Institute found models searching online for answers and probing the test software itself.
• Nobody scores above C+ on AI safety: the Future of Life Institute graded nine AI companies across 37 indicators; xAI, DeepSeek and Mistral failed outright.

What to do this fortnight

Concrete, in order of urgency:

1. Baseline what AI says about your brand. With 66% of AI-assisted shoppers comparing brands inside conversations you cannot see, the assistant's answer is your new shortlist position. Ask ChatGPT, Gemini and Perplexity the questions your customers ask and record what comes back, or run a structured report so you can track it over time rather than relying on screenshots.
2. If your content reaches EU customers, check it against the labelling rules before 2 August. AI-generated imagery that looks real counts as a deepfake under the Commission's guidance, whatever the context, and fines reach €15 million or 3% of turnover.
3. Decide your position on ChatGPT ads. Even if you never buy one, from 22 July your competitors can put a sponsored box under the answers your customers read. Know what the organic answer says about you first; that part is not for sale.
4. Check your crawler settings before 15 September. Cloudflare will block many AI crawlers by default on ad-carrying pages from that date. Review what your CDN and robots.txt allow, and make blocking a decision rather than a default, because a blocked crawler cannot recommend you.

---

Sources this issue: The Sequence, TLDR and AI Weekly, plus the primary links above.