AI Projects - November 6, 2026

Week of: November 6, 2026

Overview

This week's most concrete applied-AI work came from Zapier, which published a production-vs-preview testing method for how AI agents read its own web pages, and from HubSpot, which pushed further into answer engine optimization (AEO) measurement for marketers. Alongside those case studies, new tooling landed for document-to-JSON extraction and for running LLM prompts inside Datasette, while Latent Space used a quiet news day to publish behind-the-scenes business commentary.

Stories

1. Zapier tested how Claude reads its pages by running the same task against production and preview

Source: Zapier Link: https://zapier.com/blog/what-we-learned-watching-claude-read-zapier

Zapier's team writes that SEO feedback loops take weeks, so they wanted something faster for agents and built a tool that runs the same Claude Code task against a production page and a preview variant of that page, with the URL held constant. The first real test surfaced three things the team says it did not expect: their Markdown mirrors are already in the agent's request path, WebFetch hands the page to a separate model before the calling model ever sees it, and that intermediate model has its own instructions shaping what survives.

For teams building agent-facing content or automated site testing, this is a useful reminder that the "reader" of a page is often not one model but a chain, with an extra summarization/retrieval step in the middle that can rewrite or drop content. Holding the URL constant while swapping the served variant is a practical experimental design others can copy when debugging why agents cite or ignore a page.

Impact Analysis: If your product or content is consumed by agents, test it the way agents actually fetch it — the intermediate model in the pipeline, not just your own page, determines what survives.

2. Latent Space publishes behind-the-scenes edition covering "Kill My SaaS," Jev, and AI in 2026

Source: Latent Space Link: https://latent.space/p/ainews-behind-latent-space-kill-my-saas-jev-ai-in-2026-aiei-shanghai

The November 3, 2026 AINews edition is framed by the publication as a quiet-news-day roundup that shares "some behind the scenes," with the title pointing to segments on "Kill My SaaS," Jev, and AI in 2026, tied to AI Engineer Shanghai. Specific claims and details are not included in the available snippet — see source.

For builders, the "Kill My SaaS" framing is the notable hook: it signals continuing discussion of how AI-native tooling displaces or reshapes existing software businesses. Treat the framing as a pointer to the source rather than a conclusion.

Impact Analysis: Worth reading for the business-of-AI framing around SaaS disruption, but verify the specifics in the source before citing any claims.

3. Dharma-AI publishes a field extractor that turns any document into structured JSON

Source: Hugging Face Link: https://huggingface.co/blog/Dharma-AI/dharma-field-extractor

Dharma-AI posted a Hugging Face blog describing a "Dharma Field Extractor" that takes any document as input and returns structured fields as JSON output. The post's snippet contains no further implementation detail — see source for the pipeline, models, and usage.

Document-to-structured-data is one of the most commonly deployed automation patterns in applied AI, since it sits directly in front of accounting, intake, and CRM workflows. A published, general-purpose extractor is worth evaluating against your own schema and failure modes rather than assuming out-of-the-box accuracy.

Impact Analysis: A candidate component for document intake pipelines; benchmark it on your own documents before wiring it into production.

4. HubSpot outlines an "AEO keywords gap" method for measuring brand citations in AI answers

Source: HubSpot Link: https://blog.hubspot.com/marketing/aeo-keyword-gap-analysis

HubSpot's marketing blog argues that content and SEO marketers should add AEO keyword gap analysis to their strategies to improve visibility in AI answer engines. The post draws a direct analogy to traditional SEO: where a classic keyword gap shows terms a company isn't ranking for versus competitors, an AEO gap identifies the prompts and questions where a brand is cited in AI search but a competitor isn't.

This is a concrete, implementable workflow rather than a general plea — it turns "are we visible in AI answers?" into a comparison set that can be tracked over time. For anyone building internal or client-facing AEO tooling, the prompt-and-citation unit of analysis is the useful design pattern.

Impact Analysis: Treat AI answer citations as a trackable gap analysis against competitors, not a one-off audit.

5. Zapier rounds up beginner AI courses, citing AI skill gaps as an adoption barrier

Source: Zapier Link: https://zapier.com/blog/best-ai-courses

Zapier published a guide to the best AI courses for beginners in 2026, arguing that upskilling in AI is among the most productive uses of free time given rising demand for AI-related skills. The piece cites a Zapier survey finding that 35% of enterprise leaders say AI skill gaps among employees are a barrier to AI adoption at their company, and notes that big tech companies have been publishing high-quality AI learning content to help close the gap.

The survey figure is the more actionable data point: it frames internal enablement, not model access, as the constraint on adoption for a meaningful share of enterprises. That is a useful argument for teams pitching internal AI training or automation onboarding.

Impact Analysis: Skill gaps, not tooling, are the stated blocker for roughly a third of enterprise leaders surveyed — invest accordingly.

6. Phonon launches an on-device TTS model it says tops the benchmarks

Source: Hugging Face Link: https://huggingface.co/blog/phonon-tts

Phonon published a Hugging Face blog post introducing Phonon TTS, described in the title as high-quality on-device speech that tops the benchmarks. The post's snippet contains no further detail on model size, licensing, or evaluation methodology — see source.

On-device speech synthesis matters for applied projects where latency, cost per call, or data residency rule out a hosted API — voice agents, offline apps, and embedded devices. Benchmark claims should be checked against your own audio and target hardware before adoption.

Impact Analysis: A new on-device TTS option for latency- or privacy-constrained voice projects; validate the benchmark claims on your own hardware.

7. datasette-llm 0.1a10 released

Source: Simon Willison Link: https://simonwillison.net/2026/Nov/4/datasette-llm/

Simon Willison published a release note for version 0.1a10 of datasette-llm, the plugin that brings LLM prompting into the Datasette data tooling. The post is a version announcement; the provided snippet does not list the changes in this alpha — see source for the changelog.

Alpha-numbered releases of a plugin in this niche are mostly relevant to people already using Datasette to explore data and wanting model calls close to the database layer, which can shortcut small analysis and annotation workflows. Check the changelog before upgrading anything in a pipeline.

Impact Analysis: Incremental tooling for LLM-in-the-database workflows; read the changelog before upgrading.

Source Links

• Zapier - Zapier tested how Claude reads its pages by running the same task against production and preview
• Latent Space - Latent Space publishes behind-the-scenes edition covering "Kill My SaaS," Jev, and AI in 2026
• Hugging Face - Dharma-AI publishes a field extractor that turns any document into structured JSON
• HubSpot - HubSpot outlines an "AEO keywords gap" method for measuring brand citations in AI answers
• Zapier - Zapier rounds up beginner AI courses, citing AI skill gaps as an adoption barrier
• Hugging Face - Phonon launches an on-device TTS model it says tops the benchmarks
• Simon Willison - datasette-llm 0.1a10 released