Cloudflare charges crawlers, Google widens AI: July 2 brief

Cloudflare pushed the biggest shift today: AI traffic is moving from unpaid scraping toward priced access. That hits publishers, niche site owners, SaaS teams with data, and any operator counting on search traffic. Google also widened its AI product surface, while coding agents kept turning into a budget line. Fast week.

Cloudflare turns AI access into a billable request

Cloudflare announced its Monetization Gateway on July 1, opening a waitlist for customers that want to charge for web pages, datasets, APIs, or MCP tool calls behind Cloudflare. The payments settle through x402, an open HTTP payment protocol. Cloudflare says the gateway will verify payment at the edge before traffic reaches the origin.

This matters because the old content bargain is cracking. Search crawlers used to send traffic back. AI agents can read once, summarize, and move on. Cloudflare says AI crawlers already request content anywhere from hundreds to tens of thousands of times for every visitor they send back. That is a brutal exchange for anyone funding content with ads, leads, or subscriptions.

The operator move is not to block every bot on sight. Too blunt. First separate three buckets: pages that need discovery, assets worth protecting, and endpoints that agents might pay to call. A buying guide may need to stay crawlable. A proprietary dataset, calculator, or premium API probably does not.

This is the same pressure behind our AI Overviews briefing and the AI recommendations piece. Discovery is getting narrower. Ownership of the underlying data is getting more valuable.

Google packs more work into Gemini and NotebookLM

Google used its June AI roundup, published July 1, to group a long list of product updates: Gemini 3.5 Live Translate, Gemini 3.5 Flash with computer-use abilities, Gemma 4 12B for local agents on 16GB laptops, NotebookLM upgrades, Google Finance AI features, and Gemini Spark updates.

The most operator-relevant piece is NotebookLM. Google says it added advanced reasoning, a secure cloud computer for code, and the ability to generate charts, spreadsheets, and slide decks. Access is global for Google AI Ultra subscribers and specific Workspace accounts. That makes NotebookLM less like a note tool and more like a research workbench for messy inputs.

The move is to test it on internal research before customer-facing output. Feed it sales calls, support themes, survey notes, and competitor pages. Ask for the spreadsheet and chart first, then use a human to decide what is true and useful. Do not push generated slides straight to customers. Rough edges still show.

The broader read: Google is making AI a default layer across Search, Workspace, Finance, Android, and home devices. If your visibility plan still assumes a user clicks ten links and compares manually, it is behind. Our AI search topic page tracks the same shift from search result to answered task.

GitHub adds a lower-cost open model to Copilot

GitHub made Kimi K2.7 Code generally available in Copilot on July 1. It is the first open-weight model offered as a selectable option in the Copilot model picker, hosted by GitHub on Microsoft Azure and billed at provider list pricing under usage-based billing.

Rollout starts with Copilot Pro, Pro+, and Max. GitHub says Business and Enterprise administrators must enable the Kimi K2.7 Code policy before teams can select it. That admin toggle matters. The cheap model still has to clear your security, compliance, and data rules.

The operator read is simple: coding AI is no longer one model choice. It is a routing problem. Cheap models handle small edits and boilerplate. Expensive models get architecture, review, and high-risk changes. Even non-technical teams should care because the same pattern is coming to marketing ops: one assistant for drafts, another for analysis, another for automations.

Cursor’s CursorBench 3.1 page shows why price now belongs in every AI-agent test. It compares scores against average cost per task and includes models such as GPT-5.5, Sonnet 5, Gemini 3.5 Flash, Kimi K2.7 Code, and GLM 5.2. The standout detail is not one winner. It is the spread: higher scores can cost many times more per task.

For small teams, the move is to stop buying the most expensive model by habit. Make a two-column test: quality good enough, cost per completed task. That discipline applies just as much to AI writing tools and content workflows as it does to code.

Worth watching

Meta is reportedly exploring a cloud infrastructure business to sell AI compute and model access. If that happens, model access gets more fragmented, but compute-heavy tools may get cheaper.

Venice AI raised a $65 million Series A at a $1 billion valuation, according to TechCrunch. Privacy-first AI keeps pulling demand, but operators should treat uncensored tools carefully around brand and customer data.

HubSpot’s reported Warmly deal is worth tracking, but we could not verify the primary source cleanly today. If confirmed, it points to CRM moving closer to visitor identification and AI-driven sales timing.