Agent Experiences

Flowers by Text: The Small Business an Agent Could Order From

By Alan WeibelPublished 8 min read

This month a personal agent (an AI assistant that carries out tasks for you, such as booking or buying) ordered birthday flowers for my wife. I never opened the florist's website, and I never typed a card number. Apart from one page of options, the whole order happened in a text thread.

The florist is a small independent shop, not a national chain. It could only be on the list because the agent could read its site, meaning it could open the pages and see what the shop sells. If the agent hadn't been able to get in, the shop would never have been on the list, and its owner would never have known there was an order to lose.

Most small businesses haven't thought about this side of shopping by agent. The agents now coming out of Meta, OpenAI, Anthropic, and startups like Instinct (the agent I used, which you reach by text) don't stick to big retailers. They shop wherever they can read the menu, the catalog, or the price list. Below I walk through the order, what the agent needed from the florist's site, how that site scored, and what a small business can check this week.

The order, as it happened

I use Instinct, and it's connected to my Gmail and to our family calendar.

  1. It noticed the date. Instinct saw my wife's birthday on the family calendar a week out and asked what I had planned. I had dinner booked and nothing else.
  2. It suggested flowers. It offered to send flowers to her work. It already knew where she works from my email and calendar.
  3. It asked two questions. Which flowers does she like, and what's my budget?
  4. It found a local florist. It built a single page on the Instinct platform showing arrangements from a local shop, all within my budget. That page was the only screen I opened.
  5. I picked one and paid. Instinct sent me a bill, and I paid with Shop, Shopify's payment app, which I had used before.
  6. The flowers went out with a note. Everything else, including the message on the card, was settled by text.

The agent did the legwork and I approved each step. Nothing was bought until I said yes.

The short version

A personal agent found a local florist, showed me its arrangements, and placed the order with a note, all by text. The florist hadn't built anything special for agents. Its site let the agent in, its products were written into the pages themselves, and it kept a full list of those pages for software to find. A shop that blocks agents, or whose catalog only appears after extra code runs in the browser, may never make the list.

What the agent needed from the florist

The florist isn't on Shopify, and it didn't need to be. The payment didn't depend on the florist building anything for agents either. Shopify told Retail Brew on November 3, 2026, that Shop Pay, the payment method that Shop uses, creates a one-time card number for agent purchases at stores off Shopify, and that "the agent never sees your credit card." According to TechInformed, Instinct is one of the agents that can buy from brands outside Shopify, along with Claude Duet and Grok, xAI's assistant.

Everything else the agent needed came from the florist's own website, and none of it is advanced. I ran the site through our Agent Analyzer to see what the agent had to work with:

  • The site's robots.txt (the file that tells bots which pages they may visit) has a single rule. It keeps crawlers (programs that fetch web pages automatically) out of one small part of the site. Google's and Bing's crawlers, and every AI search tool and agent, may read the rest.
  • When the server first sends the home page or a product page, about 6,000 characters of text are already in it. An agent doesn't have to run the page's JavaScript (code that fills in parts of a page after it loads) to see what the shop sells.
  • The site's sitemap (a file listing the pages on a site so software can find them) names more than 13,000 product pages, each with a real date it was last changed. An agent can find an arrangement without guessing at web addresses.
  • The home page carries structured data (labels in a page's code that state facts such as a name, address, or price) of the LocalBusiness type, with the shop's name and address. An agent can use it to confirm the shop is local and real.
  • All 24 content images on the home page have alt text (a short written description of a picture), so an agent that can't see a photo can still read what it shows.

It was not a perfect site

The florist's home page scored 64 out of 100 on our Agent Analyzer, and a product page scored 60. That puts it in our "Fair" band. It lost points in places that matter when someone is shopping.

Product pages carry no Product or Offer structured data, the labels that tell software "this is a product, and here's its price". The price is tucked into the code behind the order button, not in a field an agent is built to look for. The pages also don't say which web address is canonical (the real one, when a page can be reached at more than one address), so the same arrangement can turn up under several addresses.

The pages have no main landmark (a tag that says where the main content of a page starts) either, so nothing in the code separates the product from the menus and footer around it. And one button and one link have no accessible name (the label that screen readers and agents use to say what a control does), so an agent can't tell what they're for.

We scanned the national flower sites on the same day, with the same tool:

SiteHome page score
BloomNation90 (Excellent)
1-800-Flowers81 (Good)
Teleflora73 (Good)
ProFlowers69 (Fair)
The local florist64 (Fair)
FTD64 (Fair)

Any of those sites could have taken the order, and the local shop tied with FTD for the lowest score. Our score measures how easily an agent can use a site, but before scoring matters, the agent has to get in. When a site blocks it, the agent moves on to one it can read.

Our own State of Agent Readiness study shows how often that happens. Of 100 well-known home pages, 27 could not be read by an automated agent at all. For 18 of those 27, bot protection (security software that tries to keep automated visitors out) turned the agent away.

Agents that can buy are arriving now

Since September:

  • Instinct raised $1 billion at a $10 billion valuation on September 28, 2026, TechCrunch reported. It's invite-only and hasn't published user numbers. Shopify announced Shop Pay checkout inside Instinct the same day, according to PYMNTS.
  • Meta announced Muse on September 8, and it works inside WhatsApp. Meta says it has more than 10 million users. On November 12 Meta started letting paying subscribers have Muse phone US businesses for them, TechCrunch reported.
  • OpenAI launched dots (agents that keep working in the background) on September 29. People in the US on ChatGPT's Pro plan have been able to text them since October 28, according to Engadget. On November 13 OpenAI launched ChatGPT Wallet so dots can complete purchases, CNBC reported.
  • Anthropic released Claude Duet on October 21 in an early test version with a waitlist, TechCrunch reported. It asks before any payment or message goes out.

Instinct works almost entirely by text, and you can text dots too. Instinct, dots, and Claude Duet can each pay a business for you.

Muse matters most if you run a small business. When it can't use your site it can phone you instead, but only for paying subscribers, and only if someone at the shop answers, which won't happen at 11pm after you close.

Why a small business should care now

People already ask AI for local recommendations. BrightLocal's 2026 survey of 1,002 US adults found 45% had used ChatGPT or another AI tool for local recommendations, up from 6% a year earlier. The people who arrive from AI tools are good customers, too. Adobe found that in September 2026, visits to US retail sites coming from AI tools turned into a sale 90% more often than other traffic, and brought in 83% more revenue per visit.

Getting recommended is still rare. SOCi's 2026 study of about 350,000 locations belonging to 2,751 brands with many locations each found that ChatGPT recommends 1.2% of them.

The forecasts for agent shopping (often called agentic commerce) are large. McKinsey estimates agents could orchestrate, meaning arrange on a shopper's behalf, $900 billion to $1 trillion of US consumer retail revenue by 2030. It notes that estimate assumes only "moderate" merchant readiness. Morgan Stanley's base case, its main forecast, puts agentic commerce at $190 billion of US online sales by 2030, and its bull case, the optimistic one, at $385 billion.

We couldn't find anyone who had worked out how much of that money would miss the businesses agents can't read, so we estimated it ourselves. Our Revenue at Risk report puts the base case at about $52 billion a year of US spending by 2030 that goes elsewhere because an agent couldn't read the business it would have gone to. It's an estimate with a wide range, and the report shows every number behind it.

What to check this week

To be the florist in this story, you don't need MCP (a way to connect your systems directly to AI tools), an app, or a developer on retainer. Start by asking yourself how far an agent would get if it tried to buy from you tonight.

Five checks for a small business

  • Open your robots.txt and your bot protection settings (in Cloudflare, your host, or your site builder), and make sure AI search tools and agents are allowed. They visit under names such as OAI-SearchBot, ChatGPT-User, Claude-User, and Perplexity-User. See how to choose which AI crawlers get in.
  • Put what you sell on the page as text. Products, menu items, services, and prices should show up without the browser having to run extra code. If your menu is a PDF or a photo, add a text version.
  • Add Product and Offer structured data with the price, and LocalBusiness with your address and hours.
  • Offer online ordering or booking that works without a login or a phone call. If you sell a catalog, see product feeds (a file of your whole catalog that shopping tools can read).
  • Some agents will call when they can't use your site, so make sure the number on your site is one someone picks up.

Then run your home page and one product or service page through the Agent Analyzer. The florist in this story scored 64 and still got the order. You don't need a 90 to get yours, as long as an agent can get in.

What this does not tell you

This is one order from one agent. I don't know which other florists Instinct looked at, or why it chose this one over a national chain, and Instinct hasn't published how it picks a merchant. The scores come from scans run on November 13, 2026, and sites change. None of the four agents above publishes how many purchases it completes, so the figures here are users and funding, not orders.

For more on what agents need from a page, see Markup Is the Interface and Agentic UX Patterns in Real-Life Scenarios.

References