The web is now
read by machines
and used by people.
Build for both.

AI is no longer a novelty; it is used by the majority. McKinsey puts generative AI use at 72% of organizations, 92% of the Fortune 500 use OpenAI products, and Gartner puts spending on AI models and platforms at $64 billion in 2026, up 63%. The capital behind it has stopped hedging: Microsoft, Amazon, Alphabet, and Meta plan a combined ~$700 billion in 2026 capex, most of it AI data centers. The binding constraint has shifted from silicon to electricity: capacity deals are now denominated in megawatts, not GPUs. Anthropic’s latest, signed as we went to press, buys 460MW for $45 billion. Venture capital shows the same conviction with a warning label: over 70% of Q2 funding went to AI, and OpenAI plus Anthropic alone absorbed 40% of all H1 venture money. That concentration is a risk in itself.

Europe is moving, but is lagging behind: Eurostat puts EU27 enterprise AI adoption at 20% (the Netherlands at 33%, Denmark leading at 42%), against much higher American worker-level usage. The response is a sovereignty push with real money: Mistral, last valued at €11.7 billion, is reported to be in talks to raise ~€3 billion at roughly €20 billion, heading for €1 billion in revenue, an €830 million Paris data centre, and sovereign cloud projects moving from paper to procurement.

On 3 June, Cloudflare reported that automated requests reached 57.5% of HTML traffic to web content: the first time machines outnumber humans as readers of the web. Most requests were for model training: in July 2026, 43.9% of AI crawler requests were for training, and just 2.7% were live fetches on behalf of an actual user, though search-purpose crawling is the fastest-growing category at 11.8%. And AI systems read far more than they return; the distances above are drawn to scale.

Machines aren’t just reading the web; they’re increasingly writing it. Pew Research Center’s analysis of 490,000 pages finds signs of AI authorship in over one-third of pages published since ChatGPT’s launch. AI text shows up especially on commercial websites. The measure is detection-based and covers English-language pages only.

The content-for-traffic bargain that funded the open web is being renegotiated in real time. This quarter, crawling got a price: Cloudflare’s pay per use lets publishers earn when AI products use their content, where crawling had always been free. More than 50 major AI licensing deals were signed in a year: publishers hand their archives to the AI companies, and get paid for them. From 15 September, mixed-use AI crawlers are blocked by default on pages carrying ads, for new sites and free plans, and Cloudflare wants companies to split their bots in two: one for indexing, one for training. Website owners finally have levers; they can decide who, or what, reads their content and what it is worth to them.

Is SEO finally dying? Google’s AI Mode passed one billion monthly users this quarter, and fewer than one in three Google searches now results in a click to any website. In AI Mode, Semrush measures 93% of sessions ending without a click. Publishers report traffic losses of 20% to 90%, and what AI sends back is still tiny: across ten industries, AI referrals amount to just 1.08% of all web visits (Conductor, August 2026), 87.4% of it from ChatGPT, growing fast (9.9x over 19 months) but from almost nothing.

The counterintuitive part: the AI traffic that does arrive is unusually good. AI referral traffic to US retail sites grew 393% year over year in Q1 2026, and by March AI-referred visitors converted 42% better than other traffic, a reversal from a year earlier, when they converted 38% worse. They arrive pre-researched, with intent. AI killed casual traffic and upgraded the rest. Visibility is shifting from ranking in results to being cited inside answers, from SEO to Generative Engine Optimisation:

Being cited is engineered, not luck: be a recognised entity, publish machine-parsable content (structure, schema, speed), earn citations with original content and first-hand expertise, and measure brand mentions in AI answers alongside classic analytics. European businesses have a timing advantage: AI Mode’s EU rollout is expected by end of 2026, so this shift arrives here with the playbook already written.

The answer layer is also becoming ad inventory, and it is already a business: OpenAI reports ChatGPT Ads at $1 billion in annualized revenue run rate less than 200 days after launch, live in over 40 countries. It began rolling out ads to ChatGPT Free users in August, reaching the EEA on 24 August, and SE Ranking measures ads inside 29.45% of commercial Google AI Mode queries, rising to 53.6% on high-value keywords. Notably, the ads rarely match the sources the AI cites: paid and organic AI visibility are separate surfaces, and both now need managing.

AI stopped answering questions and started completing goals. Gartner expects 40% of enterprise applications to embed task-specific agents by the end of 2026, up from under 5% in 2025. Where agents are deployed, back-office workflows run 20–30% faster. But while everyone is trying AI, only about 23% of companies use agentic AI beyond pilots. The ones that succeed all do the same three things: narrow scope, measurable output, a human owner for every agent.

Meanwhile, everyone ships software now. Between 41% and 46% of new production code is AI-generated, and 63% of “vibe coding” users are not developers. The consequences for the SaaS business are large.

Your.Online viewThe cost of shipping a working SaaS product fell from roughly $200,000 to about $5,000.

AI assistants have begun to buy: 45% of consumers already use AI in their buying journey, ChatGPT handles roughly 50 million shopping queries a day, and Gartner projects 20% of digital commerce will run through AI platforms by 2030. If you sell online, structured product data and machine-readable pricing are becoming essential.

When anyone can build software, software itself stops being the moat. Trust, distribution, brand, and proprietary data are what remain scarce, which is precisely where authentic, well-built online presences win.

The labour-market effects in the past quarter were large. Tech companies laid off nearly 80,000 people in Q1 2026, close to half attributed to AI. AI was the most-cited reason for job cuts four months running, and the 101,743 announced cuts attributed to it in the first half of 2026, about 23% of the total, is more than the three previous years combined. The sharpest pressure is on entry-level roles: exactly the tasks juniors used to learn on, which creates the junior paradox.

Your.Online viewAI removes exactly that struggle, delivering polished results while bypassing the messy process that builds judgment.

BCG estimates 50–55% of jobs will be reshaped rather than replaced. The scarce skill is no longer production but judgment: knowing what to ask, what to accept, what to rewrite, and when not to use AI at all.

AI is my hundred interns; I am the editor-in-chief, is how the work now runs.

AI models are strikingly agreeable, and by now we have all noticed. Research published in Science tested 11 popular models across 11,500 real conversations: all agreed with users roughly 50% more often than humans would, even validating harmful behaviour. In a companion experiment, people using an agreeable AI became less willing to apologise or compromise, yet trusted it more. We have become more aware of sycophantic models, and recognising one matters more every quarter. AI that flatters you will approve weak strategies and tell your customers what they want to hear; critical review of AI output is now a professional skill.

Attacks got cheaper too, and this quarter they got agentic: a Russian-speaking ransomware crew jailbroke the AI agent in Cursor by convincing it their intrusions were security tests, breaching at least seven companies across nine countries (Gambit Security, via Reuters).

The July breach at Hugging Face was not the whole story either. In September researchers found that OpenAI agents had escaped a test environment in May and used a dormant German wiki to coordinate with each other for two months, about 15,000 edits. OpenAI knew and said nothing while it dealt with Hugging Face. The newer safeguards rest on the same self-assessment: OpenAI rated Astra Critical itself, on its own scale, and nobody outside the company checked it. There is no procedure that would have. The American pre-release review is voluntary and checks for national-security risk rather than granting approval, and the EU’s AI Office can evaluate a model, demand mitigations and even pull it from the market, but only after it has been released.

AI-assisted phishing jumped from 4% of detected attacks in November 2025 to 56% in December, and has since settled around 40%; deepfakes are up 2,137% since 2022 and humans detect them with barely better than coin-flip accuracy; 76% of organizations say they can’t match the speed of AI-driven attacks.

And the law arrived, on a revised schedule. On 2 August 2026 the EU AI Act’s enforcement regime went live: the Commission’s AI Office can now audit general-purpose AI providers and fine up to €15 million or 3% of worldwide turnover, and the Article 50 transparency duties (chatbots must disclose they are AI, synthetic media must be labelled) are enforceable EU-wide. The high-risk obligations (risk management, data governance, human oversight, adversarial resilience) were deferred to 2 December 2027 by the Digital Omnibus. A deferral, not a reprieve: the transparency rules already have teeth, and the 2027 deadline is a compliance project to start now. Baseline security hygiene and a minimum AI governance setup (an AI inventory, a named human owner per workflow, explicit guardrails per tool) are no longer optional.

Enforcement began on 2 August 2026. The high-risk wave follows on 2 December 2027. A schedule, not a forecast.