Act II begins: wiring AI into your stack. This episode moves you from re-pasting your brand voice into every chat to building one saved, grounded assistant your whole team can use.
In the news
- HubSpot's July product wave (July 6, 2026): Breeze AI reaches five new surfaces (summarized global search, inline email help, assignment-notification CTAs, mobile meeting briefings, and a HubSpot Agent for Microsoft 365 Copilot). The Prospecting Agent is now on every paid seat with a new Seamless integration for net-new contacts. Breeze Projects gets file uploads, CRM @mentions, team permissions, and templates. Breeze Copilot is renamed Breeze Assistant; Commerce Hub becomes Revenue Hub (locked quotes, auto sales-tax, Klarna/Affirm BNPL, HubSpot Capital in US/UK). Trailing context: outcome-based pricing since April (~$0.50/resolved conversation, $1/lead); HubSpot reportedly claims 65% resolution and 39% faster handling across ~8,000 customers.
- Google Ads (tool live July 6, change effective Aug 17): Budget-limited target-based campaigns will hold targets instead of beating them, across Search tROAS/tCPA, Shopping, Performance Max, Demand Gen, and Travel. Use the new Bid Target Adjustment Tool to reset targets before Aug 17.
- Models: OpenAI reportedly delaying its next public launch after a US government early-access request (July 3). Digiday (July 1) reports OpenAI hiring an ad-formats engineer for ChatGPT ads. Anthropic shipped Claude Sonnet 5 (June 30). Trailing US assistant share: ChatGPT ~58%, Gemini ~19%, Claude ~13%.
- AI-search check: Google's "Further Exploration" links and ask-a-follow-up into AI Mode are rolling wider (~July 6). Trailing data: Ahrefs ties AI Overviews to a ~58% CTR drop for top pages; Chartbeat sees ~33% fewer Google referrals; AI search ~22% of queries. Trackers worth a look: Profound, Topify, LLMsRefs, Cloro, and Semrush's AI Visibility Toolkit.
The tutorial
What grounding is, the three rungs (custom instructions, uploaded files, true retrieval/RAG), how Custom GPTs, Claude Projects, Gemini Gems, and NotebookLM each handle it, the eight docs to feed your assistant, how to structure and clean them, a sample instruction block, a test battery to confirm grounding took, version control against stale docs, team governance, privacy rules, and the pitfalls, starting with the big one: grounding reduces hallucination, it doesn't end it.