We audited 18 GEO/AEO companies' own sites. Most skip their own advice.

Method

On June 7, 2026 we requested the homepage of 18 well-known GEO/AEO/AI-search vendors and tools, then probed six signals with simple HTTP requests:

• Crawlable homepage — did it return HTTP 200 to a plain crawler?
• JSON-LD — is application/ld+json present in the homepage HTML?
• llms.txt — does /llms.txt return 200?
• Content-Signal — does robots.txt declare Content-Signal / ai-input / ai-train?
• MCP manifest — does /.well-known/mcp.json return 200?
• Markdown negotiation — does the homepage return text/markdown when asked with Accept: text/markdown?

This is a point-in-time snapshot of a small sample, not a census — treat it as directional. One site (BrightEdge) returned 403 to our crawler, so signal rates below are computed over the 17 that served a homepage.

A note on rigor: for llms.txt and markdown we didn't trust the HTTP status code alone — many sites return 200 for any path (a soft-404 serving the homepage). We required the response to be real plain-text/markdown, not HTML. That distinction matters: one site (GenOptima) returned 200 on /llms.txt but served its HTML homepage, so we counted it as no real file. Status-only checks would have over-reported llms.txt adoption.

Results

Signal	Adoption (of 17)
JSON-LD structured data	82% (14)
llms.txt present (real file)	41% (7)
Markdown to agents	24% (4)
Content-Signal in robots.txt	6% (1)
MCP manifest (/.well-known/mcp.json)	0% (0)
Crawler hard-blocked (403)	1 of 18 (BrightEdge)

Sample: Ahrefs, Semrush, Conductor, Profound, Otterly, Peec, Jasper, Writesonic, Surfer SEO, Clearscope, seoClarity, BrightEdge, First Page Sage, Evertune, LLMrefs, RankScope, GenOptima, Goodie.

What it means

1. The industry over-invests in the signal that doesn't move citations

JSON-LD is the most-adopted signal (82%) — yet controlled testing shows schema has near-zero effect on AI citations. It's valuable for Google rich results, but its dominance here suggests the field is optimizing for the checkbox that's easy to ship and easy to demonstrate, not the one that earns citations.

2. The machine layer is almost entirely unbuilt

Not one site exposed an MCP manifest, and only one declared a Content-Signal. These are the forward-looking, machine-actionable signals — and adoption among the very companies selling AI readiness is effectively zero. The future they describe in their blog posts isn't reflected in their own DNS yet.

3. Even llms.txt — which many of them recommend — is a coin flip

Under half serve the file the category evangelizes. (For what it's worth, that's defensible: llms.txt adoption is ~6% web-wide and its impact is unproven — but it's striking that the advocates themselves are split.)

4. Markdown negotiation is the quiet early mover

About a quarter return clean markdown to agents — higher than we expected, partly because Cloudflare can enable it automatically. It may be the first machine-layer signal to cross the chasm.

The companies selling AI visibility have mostly shipped the cosmetic signal (schema) and skipped the structural ones (Content-Signal, MCP). The gap between what the industry preaches and what it deploys is the story.

Caveats

Small sample (18), single snapshot, homepage-only probes, and HTTP-level signals can't capture everything (a site may serve markdown or govern crawlers by means our probe missed). We publish the method so it can be reproduced and challenged — which is rather the point of honest data in this field.

Frequently asked questions