Review: deck-wide AI edits, a company context playbook, and the truth about shopping agents

This week's useful launches and guides have a common thread: AI works better when it has the right context, and the hype works better when nobody checks the evidence. As usual, these are first assessments from what has been published, not test results.

The reviews

Gemini in Google Slides. Verdict: try it on your next deck refresh. Gemini can now edit multiple slides or a whole presentation from one prompt. It can update content throughout a deck, refresh visual styles in bulk, or match the style of another presentation, and it shows a summary of what it changed when it finishes. That last part matters. Every marketing team has a sales deck with five slightly different versions of the same message. A deck-wide edit with a change summary is a fast way to bring them back into line, as long as someone reads the summary before the deck goes out.

Ramp's guide to giving AI company context. Verdict: use it this week. The guide argues that AI gives confident but wrong answers because it cannot guess your internal definitions, policies and workflows. Its fix is practical: a one-to-two-page context document, live connections to the tools where work happens, permissions set first, a named owner, and a test of 10 to 20 questions you already know the answers to. For marketing, the obvious version is a brand context page: positioning, audience, words you use and words you avoid. A related finding supports keeping it short. Claude Code removed more than 80% of its system prompt for newer models without measurable loss on its coding evaluations. Clear and brief beats long and exhaustive.

Substack's recommendation loop. Verdict: worth testing if real people on your team have something to say. One piece claims Substack can send a newsletter roughly 50 times more subscribers than Google, because every signup shows a few newsletters the writer recommends. Companies are using that with behind-the-scenes posts from the people who build the product, and by swapping recommendations with newsletters that share their readers. It works only if the writing is genuine. A rebranded press release will not earn recommendations.

Algos Wiki. Verdict: a handy reference, but check the source. It is an unofficial, AI-powered encyclopedia of Instagram's algorithm, organized from platform announcements and leadership interviews, with links to the supporting sources. Useful for settling internal debates about posting times and shares. Because it is unofficial and AI-assembled, click through to the original statement before you change your strategy on it.

The finding that matters more

The best piece this week was a reality check on AI shopping agents. Its argument: the agents with the strongest evidence of sales impact are the ones run by retailers, not the general chatbots. Amazon reports nearly $12 billion in incremental annualized sales from its assistant, and Walmart says users of its Sparky assistant spend 40% more per order. ChatGPT and Google have enormous reach, but no published purchase results.

DoorDash made the same point from the other side. It said it has received only "several hundred" orders from AI agents, while processing 3 million deliveries a day. Most orders still come straight from its own app.

My read: AI-driven discovery is real, but in most categories nobody has yet shown it moves sales, outside the retailers' own stores. Reach is not proof.

What I would do

• Write a one-page context document for your brand and test it with ten questions you know the answers to.
• Make your product data complete and accessible, because that is what any agent, retailer-owned or not, reads first.
• Ask for purchase results, not reach, before you fund an agent channel. If a platform cannot show sales, treat it as a test.

The teams that win with AI this year will be the ones that feed it the right context and keep asking for proof.