Original research

Two in Five Texas and Georgia Businesses Are Invisible to AI: A 574-Business Study

AA

Adev Aarons · Founder, Specularis
September 24, 2026 · 4 min read

Short answer: We tested 574 businesses across Austin, Houston, and Atlanta the way ChatGPT reads them. 42% are invisible to AI right now, and 21% block the AI crawler outright. Real estate is the worst industry of the four we tested, at 60% invisible. This is the same method we used for our 561-business Florida study — the numbers run a few points better outside Florida, but the underlying problem is the same, and it isn’t a Florida problem.

What we did

We pulled real businesses off Google Maps across three metros, Austin, Houston, and Atlanta, across four industries: real estate agents, law firms (personal injury), roofing contractors, and med spas. Roughly 40-60 businesses per metro/industry combination, 574 total after excluding offline or broken sites. Then we requested each homepage the way an answer engine does, as OpenAI’s GPTBot. No cherry-picking. No pre-filtering for sites that were already broken.

A business counts as invisible to AI if either of two things is true:

• The site blocks the crawler (a hard 403 or a Cloudflare “checking your browser” challenge a bot can’t pass), or
• The site publishes no structured data, so even when AI can read the page, it can’t tell what the business is.

The results

By metro:

Real estate was the widest gap in every metro we tested, not just Florida. Across 108 agents in Austin, Houston, and Atlanta, 60% are invisible to AI, worse than personal injury law, roofing, or med spas. Georgia had the highest crawler-block rate of any state we’ve tested so far, Florida included. Med spas were again the best-prepared industry, the same result we found in Florida, likely because they already run the most marketing-forward operations of the four.

Why this keeps happening

The same two failures we found in Florida show up here, in the same proportions.

The crawler block. A 403 to GPTBot is usually a toggle inside a security product or a firewall rule a host set, not a decision the business made. Your robots.txt might say “welcome” while the firewall says “blocked,” and nobody compares the two. It only shows up when you look at the site as the bot does.

Blocking those crawlers buys you nothing on security. Malicious scrapers ignore the rules and rotate IPs. GPTBot, ClaudeBot, and PerplexityBot are the polite ones that actually respect robots.txt. Disallowing them doesn’t stop attackers. It just deletes you from the answer.

No machine-readable identity. Plenty of sites AI can reach still say nothing about what they are. No structured data, no schema. A storefront with no sign. A person figures it out. AI won’t guess.

The fix, and the honest timeline

Same order that worked in Florida.

1. Open the door. Make sure your firewall and robots.txt let the AI crawlers through.
2. Emit structured data. JSON-LD that names your business, says what it is, and where it operates.
3. Get corroborated. Reviews, directories, and real mentions across the web, all saying the same thing about you.

Steps one and two are fast, usually about a week. Step three takes longer, roughly two to three months, because trust builds as the web gets re-read. The businesses that start now own the answer before their market catches on.

Check your own site

The bar is still low outside Florida too. Close to half your competitors in real estate specifically return a 403 to the crawler or publish nothing AI can read.

You can test your own site in a minute: request it with a GPTBot user-agent and see if you get a 403 or a Cloudflare challenge, then view your page source and search for application/ld+json. Or run a free audit that scores all five pillars and shows you exactly what’s blocking you.

Methodology note: businesses sampled from Google Maps across Austin TX, Houston TX, and Atlanta GA, deduped to unique domains, directory and social-platform listings (Facebook pages, shared realtor directories, apartment complexes and hotel spas that surfaced under unrelated searches) excluded as not testable business sites. Homepages requested live as GPTBot. Sites that returned neither a readable browser response nor a bot response were excluded as offline. Data collected September 2026. Same method as the Florida study, so the two are directly comparable.