Newsroom

WebMCP is a Community Group draft, ACP is in beta with no discovery mechanism, Web Bot Auth has no adopted document. What you can actually implement today.

nosnippet and max-snippet do affect AI Overviews, FAQPage has been deleted from the gallery, and no vendor documents how a passage is chosen. Only what is officially documented.

llms.txt, sitemap, schema and well-known. I went through the official documentation from Google, OpenAI, Anthropic, Perplexity and Microsoft: what is actually read, and what is only promised.

The only public measurement found that major AI crawlers do not execute JavaScript. How to check in two minutes what actually leaves your server, and what to do if the

Training crawlers and answer-time fetchers are two different families with separate user-agents. Who to block, who to leave open, and how to verify in the log that it actually works.

The break-even between self-hosting and an API, why GPU utilisation decides everything, what the calculation leaves out, and why routing by task beats a binary decision.

The approval point is set by reversibility and damage, not by model confidence. Four approval patterns, what belongs on the approval screen, and how to stop it becoming a rubber

Ordinary software testing does not work on AI agents. The full guide to building an eval set, four levels of measurement, regression testing and production monitoring.

Agent boundaries live in the permission model, not the prompt. Meta Rule of Two, a four-layer permission model, prompt injection in practice, and the list of actions an agent never

An AI agent is a loop of planning, action and observation, built from four components: model, tools, memory and control. What actually runs under the hood, what MCP changed, and

Six levers for reducing inference cost, ranked by return on effort: prompt caching, model routing, batch, output shortening, semantic cache and distillation — and the right order to apply them.

95 percent of enterprise GenAI pilots never reach the income statement. Four layers of ROI, a six-week pilot structure, and four signals that justify closing early.

Per seat, per usage or per outcome: how to price a product built on an LLM when gross margin sits at 50-60 percent and 5 percent of users drive 80

Build versus buy is not binary. Five possible routes, the three lenses that decide between them, when fine-tuning genuinely pays, and why an abstraction layer is the one decision to

Token price is one line out of seven. A full cost-of-ownership model for an LLM-backed feature, covering retries, retrieval, evaluation, monitoring and human oversight, plus the formula for cost per

For a feature with an AI component the acceptance criterion becomes an eval. Its four components, how to build a test set without a research team, six sections to add

Public, internal, confidential and restricted: what may be entered into a language model at each tier, anonymisation versus pseudonymisation, the three contract clauses you need, and a four-step implementation plan.

Between 70% and 80% of enterprise RAG projects never reach production, and most failures sit in the retrieval layer rather than the model. The four layers of a knowledge base,

Mention rate, citation rate, share of voice and accuracy – the four metrics replacing position and CTR, and how to feed them from the Search Console Generative AI report, GA4