Sam Pajot-Phipps

Building | Enabling entrepreneurs to open for business inside AI assistants

[Sam Pajot-Phipps @sampajotphipps]

Today @courtyardai is officially live! 🎉

After spending the last two months working closely with a select group of amazing small businesses, we’re ready to open the doors!

The more AI assistants like ChatGPT or Gemini know your small business, the more it can help your potential customers.

If you’re running a small business, Courtyard makes it easy to tell AI everything about your business.

Within minutes of signing up, Courtyard:

- Identifies and automatically pulls key knowledge out of 40+ software tools (ex: scheduling, availability) your human visitors can see on your website but AI assistants can’t

- Finds any relevant PDFs (ex: menus, price lists) on your website then converts and optimizes them into knowledge

- Conducts deep research across the web on relevant questions about your business and drafts it into your knowledge base

- Crawls your site for all relevant images, then organizes and optimizes them into a media library for AI retrieval into relevant responses (coming very soon!)

All of this gets you started with a comprehensive, living knowledge base of your business. Made available to AI assistants like ChatGPT, Claude and Gemini in the variety of formats, standards and locations of how they retrieve knowledge.

You can then easily add knowledge by dropping in a doc, image, text or voice. Launching email forwarding soon 👀

We’re also continuously rolling in new opportunities and emerging standards:

- Every knowledge base is also an MCP server, paired with a discoverable agent skill and server card

- Auto-translations so your knowledge is optimized for potential customers prompting in their preferred language

- Supporting new AI search crawlers: @ExaAILabs @p0 @firecrawl ..etc

Just getting started. https://t.co/U9yK6T7j8C

[Sam Pajot-Phipps @sampajotphipps]

🚢 New updates to analytics in @courtyardai

- Easily see what AI assistants are visiting your knowledge base to use in training and live responses to your potential customers

- Explore how AI assistants understand your business: we automatically monitor relevant prompts, and see where they are getting their answers from, including your knowledge base

Much more to come on how AI understands your small business, and quick actions to directly address gaps that can cost you customers.

[Sam Pajot-Phipps @sampajotphipps]

ChatGPT visited your website 10 times last week. So what? Why?

This is a topic we're constantly exploring at @courtyardai

How are AI assistants using Courtyard knowledge bases?

There are many ways we look at it, but let's start with AI visits. Each knowledge base exists at its own subdomain, hosted by Courtyard so we can track all server traffic.

We break down AI traffic by their known bot identifiers, across three activity categories: live retrieval, indexing and training.

All three work together to paint a picture of AI assistants finding, understanding and then using a knowledge base.

Live retrieval (e.g. ChatGPT-User) is the highest signal that tells us a relevant conversation was happening with an AI assistant that resulted in a live request to the knowledge base to help form an answer. A close AI equivalent of calling the front desk.

However, not all usage of the knowledge base requires a live retrieval. When it's been indexed and cached by AI assistants, answers can use the cached version in which case there is no signal of that interaction, which is a limitation we need to work around with other signals.

Tracking indexing traffic helps us understand what categories in the knowledge base have been discovered and ingested by AI, and looking at frequency tells us how often they come back to see what's new.

Tracking training has fewer short-term benefits, but it provides a signal that this knowledge becomes eligible to be included in future model training. This is important as some answers come from a combination of training data and search. Courtyard knowledge bases often contain new knowledge that was previously inaccessible to AI assistants like availabilities, schedules, and pricing as they were locked behind human web experiences (i.e JS, widgets) or in a desktop folder or email thread. Building awareness that this knowledge is now available for a business into training data could result in better responses in future models.

Later, I'll share how we monitor prompts across models to see how a knowledge base is shaping responses.

[Sam Pajot-Phipps @sampajotphipps]

Little update on this experiment, in two weeks:

- Visited 985 times to answer live user questions by ChatGPT (ChatGPT-User traffic)

- Per showtime seat map pages have the most AI assistant visits (177 distinct seat pages visited)

- Ran a sample of 25 relevant prompts across 4 models to see how the new knowledge is being used in responses: ChatGPT citing it in 28% of prompts, Perplexity is 95%, Gemini started at 30% but has dropped off.

- Seeing major AI crawlers return almost every day

As showtimes near, data is refreshed every 15mins and updates pushed for indexing.

[Sam Pajot-Phipps @sampajotphipps]

Recently, one of us at @courtyardai was using ChatGPT to research movie showtimes and when asking what seats were available, noticed the similar response pattern of when AI hits a wall “I can’t access any live seat availability information.”

So we decided to fix it. Using our internal Courtyard 'ivy' agent, which closes the gap between the human interactive web and the agent-accessible web. Ivy takes the live business data locked inside human web experiences: what's available, what it costs, when can you go, and turns it into structured, fresh data for AI. So we pointed it at this theatre chain.

Ivy:
- Created a feed that now serves structured and detailed data like showtime granular seat availability, per theatre accessibility, seat types and translations.

- Automatically publishes this data on opentheatreseats(dot)com how AI wants it: simple html, markdown twins, llms, llms-full, json-ld, schema markup and pushes it across indexing engines.

Within 8 days:
- Went from zero to fully ingested by AI training crawlers, steadily indexed by Google + OpenAI (over 4K daily AI crawlers)

- Being pulled in ~100×/day to answer live seat-availability questions in ChatGPT and Claude via ChatGPT-User and Claude-User bot traffic

- Being cited in 14/15 prompts in ChatGPT, and 5/10 in Gemini based on the specific data unlocked

More experiments to come!

[Sam Pajot-Phipps @sampajotphipps]

Since OpenAI added image search results into the Responses API a few weeks ago, anyone else seeing ChatGPT increasingly bring and suggest images into conversations?

https://t.co/fGscMd9cBf

[Sam Pajot-Phipps @sampajotphipps]

The more knowledge AI has about your small business, the more it can bring you business.

@courtyardai we're always pushing to make this easier. Now when you create a new category, you can hit ‘draft it’.

We’ll run an extensive deep research across the web (i.e: your site, 3rd party aggregators, reviews, blogs..etc) to pull together this knowledge about your business. Then just review, edit and approve. Now this knowledge about your business is further in your control, increasingly making your business the source AI goes to.

[Sam Pajot-Phipps @sampajotphipps]

🚀 As of today, every @courtyardai knowledge base now automatically has an MCP!

All knowledge inside your knowledge base, whether that’s schedule availability, past work, treatments or media is now also accessible via your own MCP. As agents increasingly dynamically discover MCP servers when relevant to their users requests, the key knowledge about your business is readily accessible, up-to-date and structured the way AI wants.

👉 Running a small business? Sign-up today, free during our beta.

[Sam Pajot-Phipps @sampajotphipps]

Recently, one of us at @courtyardai was using ChatGPT to research movie showtimes and when asking what seats were available, noticed the similar response pattern of when AI hits a wall “I can’t access any live seat availability information.”

So we decided to fix it. Using our internal Courtyard 'ivy' agent, which closes the gap between the human interactive web and the agent-accessible web. Ivy takes the live business data locked inside human web experiences: what's available, what it costs, when can you go, and turns it into structured, fresh data for AI. So we pointed it at this theatre chain.

Ivy:
- Created a feed that now serves structured and detailed data like showtime granular seat availability, per theatre accessibility, seat types and translations.

- Automatically publishes this data on opentheatreseats(dot)com how AI wants it: simple html, markdown twins, llms, llms-full, json-ld, schema markup and pushes it across indexing engines.

Within 8 days:
- Went from zero to fully ingested by AI training crawlers, steadily indexed by Google + OpenAI (over 4K daily AI crawlers)

- Being pulled in ~100×/day to answer live seat-availability questions in ChatGPT and Claude via ChatGPT-User and Claude-User bot traffic

- Being cited in 14/15 prompts in ChatGPT, and 5/10 in Gemini based on the specific data unlocked

More experiments to come!

[Sam Pajot-Phipps @sampajotphipps]

@paw_lean @vercel 👋 excited for it!

[Sam Pajot-Phipps @sampajotphipps]

Codex + seo-geo skill + webflow mcp = 🤯

[Sam Pajot-Phipps @sampajotphipps]

@illscience @figma 💯 really helpful to visually explain a plan or a new feature.

[Sam Pajot-Phipps @sampajotphipps]

NotebookLM for repos? 👀

Excited to try it. Soon we'll be listening to generated podcasts explaining PRs.

https://t.co/Q5eP40f0Zf

[Sam Pajot-Phipps @sampajotphipps]

"We already see some of the most popular coding agents today – like Claude Code and OpenCode – send these accept headers with their requests for content"

Such a turn key way to better serve agentic traffic 👏

[Cloudflare @Cloudflare]

Time to consider not just human visitors, but to treat agents as first-class citizens. Cloudflare’s network now supports real-time content conversion to Markdown at the source using content negotiation headers.
https://t.co/B7wYH4PtA8

[Sam Pajot-Phipps @sampajotphipps]

@OpenAIDevs This is gold 🙏 super helpful

[Sam Pajot-Phipps @sampajotphipps]

@benln Got a link for Calgary? 🤠

[Sam Pajot-Phipps @sampajotphipps]

@OpenAI Are all connected apps able to be used in deep research tasks? Noticing the apps I can toggle are only a few from my connected.

[Sam Pajot-Phipps @sampajotphipps]

Tip: using the diagram tool for @figma MCP with Codex has been super useful to visually learn and summarize a feature or a decision the model has made.