How AI reads Sweetgreen

sweetgreen.com Jun 26, 2026 6 min read Emerging

Short

43/ 100

AEO Level 3Emerging

The short answer

Sweetgreen scores 43 out of 100 on our AEO scan, landing at level 3. The farm-to-table salad chain has solid content structure but leaves AI agents without crawl permissions, machine-readable interfaces, or an identity layer — meaning answer engines can surface menu items but struggle to act on or attribute them.

What AI sees

When an AI agent hits Sweetgreen's homepage today, here is the signal it actually collects.

Structured data and clean semantic HTML give agents a reasonable grip on menu categories and restaurant locations — content_structure scores 80 out of 100. Agent discovery sits at a mid-range 60, meaning crawlers find some breadcrumbs but no explicit permission grants. What agents cannot do is confirm who controls the site, access a machine-readable API surface, or request a Markdown-formatted summary of the page. The homepage returns only HTML regardless of the Accept header sent, and no Link response headers point agents toward programmatic resources. For a brand selling customizable bowls to nutrition-minded consumers who increasingly discover restaurants through AI-powered queries, that gap translates directly into missed citations.

Sweetgreen homepage as an AI agent sees it — structured menu content visible, but no Markdown response or Link headers present

Sweetgreen robots.txt showing no AI crawler user-agent blocks, leaving GPTBot and peers without explicit permission

Where it loses points

Agent interfaces and identity authentication both score zero — the sharpest drag on Sweetgreen's overall 43-point result.

How to fix it

Three targeted changes address the most critical gaps the scan uncovered and would move the score fastest.

AI Crawler Rules in robots.txt

Goal

Declare explicit per-agent rules so GPTBot, ClaudeBot, PerplexityBot, and peers know exactly which paths they may index and cite.

Issue

The scan found no user-agent blocks for any major AI crawler in Sweetgreen's robots.txt, leaving their behavior entirely undefined.

Fix

Add named user-agent sections for GPTBot, ClaudeBot, and PerplexityBot. Under each, Allow the menu, locations, and nutrition pages you want cited, and Disallow checkout or account paths. This single file edit can unlock consistent AI citation across ChatGPT and Perplexity overnight.

MCP Server Card

Goal

Publish a machine-readable server card at /.well-known/mcp/server-card.json so autonomous agents can discover Sweetgreen's capabilities without guessing.

Issue

No MCP Server Card exists at the standard well-known path, directly contributing to the agent_interfaces category scoring zero.

Fix

Create /.well-known/mcp/server-card.json with a serverInfo object (name, version, description), a transport endpoint, and a capabilities block. For Sweetgreen this could expose menu-lookup and store-locator tools, making the brand directly actionable inside AI assistants rather than merely citable.

Markdown-on-Demand Responses

Goal

Return a clean Markdown version of pages when agents send Accept: text/markdown, giving AI parsers a compact, citation-ready format.

Issue

The site returns only HTML regardless of the Accept header, forcing every AI crawler to parse full browser-targeted markup.

Fix

Intercept requests carrying Accept: text/markdown at the CDN or application layer and respond with Content-Type: text/markdown. For menu and nutrition pages, a stripped Markdown response with ingredient lists and calorie counts is precisely what voice assistants and AI answer engines quote back to health-conscious users.

Common questions

Why does Sweetgreen score 43 out of 100 on AEO?

The score reflects strong content structure (80/100) and moderate agent discovery (60/100) offset by zero points in both agent interfaces and identity authentication. AI agents can parse menu pages yet cannot confirm site ownership, access an API surface, or request structured output — three gaps that collectively cap the result at 43.

Does missing AI crawler rules in robots.txt hurt Sweetgreen in AI search results?

Yes, meaningfully. When no explicit user-agent rules exist for GPTBot, ClaudeBot, or PerplexityBot, each crawler falls back to its own defaults — some become conservative and under-index, others skip the site entirely. For a restaurant brand where AI-powered food queries are growing fast, undefined crawl permissions produce inconsistent and often incomplete citations across major AI platforms.

What would an MCP Server Card do for Sweetgreen's AI visibility?

An MCP Server Card at /.well-known/mcp/server-card.json tells AI agents what the site can do programmatically — menu lookups, store locators, or nutritional queries. Rather than indexing static content alone, agents could call Sweetgreen's tools directly, turning a passive citation into an active interaction and making the brand far more present inside agent-driven dining discovery.