The AI Invisibility Index

Why we ran this study

45% of US consumers now use AI tools to find local services, up from 6% a year earlier, and finding a therapist is about the most personal local search anyone makes. When someone asks an assistant for one, the assistant answers by reading practice websites. We wanted the one number nobody had: how much of the profession can it actually see?

So we pointed our website readiness engine, the same one behind our free scan, at an independent, nationwide sample and let the data speak.

What we found

Historical checklist results. 749 of 962 fully graded practices fell below the original threshold (77.9%). The scan did not test AI answers, citations or recommendations. Our earlier interpretation of this percentage as AI invisibility was not supported by those measurements.

Short-text flags need context. The original scan flagged 87.4% of sites for short text on sampled pages. Word count alone does not establish content quality, ranking potential or whether an assistant can quote a page. Embedded and JavaScript-added content may be excluded.

Optional files and structured data. The original scan did not detect llms.txt on 65.7% of sites and structured data on 15.6%. Missing llms.txt is not evidence of AI unreadability; Google does not require it for Search or its AI features.

Slow phones are a profession-wide problem. Google scores how fast a site loads on a phone from 0 to 100. The middle site in our sample scored 57, and only 5.2% reached Google's "good" band of 90 and above. Separately, 18 practices, about 2%, use their robots.txt file to tell AI crawlers to stay away, which requires checking the purpose of each crawler; training restrictions and search access are not interchangeable.

Historical state comparisons. Colorado is the best in the country and still has 55% of its practices below the line. Alabama and Kansas sit at 95%.

Every state, ranked

Each row shows the share below the original checklist threshold, not the share invisible to AI. These historical percentages use the original methodology and have not been recalculated after the correction.

Methodology

The sample is 984 independent therapy practice websites, up to 20 per state across all 50 states. We drew them from public Google Maps listings for therapist and counseling searches in each state's three largest cities, then picked from that pool at random, with the random draw recorded so the same sample can be rebuilt. We left out national chains, directories, telehealth platforms, and every practice with any connection to Moonraker. Scans ran 14 to 15 September 2026 on Moonraker's production readiness engine, the version then used by our free scan, before the 21 September corrections. It checks whether the site loads, whether it is secure, its page titles and descriptions, its structured data, what its robots.txt allows AI crawlers to do, whether it has an llms.txt, its sitemap, how much real content each page carries, whether the Google Business Profile is connected, and its Google PageSpeed score. 982 of the 984 sites returned usable results, and 962 could be graded on every check.

What this study cannot tell you: the sample leans toward practices that already show up in Google Maps in larger cities, so it is not representative of all therapy practices, particularly online-only practices. The original "AI-readable" label was a checklist category, not a verified test of an assistant reading or recommending the practice. Speed scores come from a test environment rather than from real visitors, and they move a few points between runs.

The dataset

Historical state-level aggregates under the original methodology are public under CC BY 4.0: download the CSV. Cite as "Moonraker AI Invisibility Index, September 2026", linking this page. Journalists: for state-level breakdowns, additional cuts, or comment, contact scott@moonraker.ai. We do not publish per-practice results.

The Australian study

We ran the same measurement across Australia in September 2026, on the corrected rubric: AI SEO for Australian therapy practices, 988 practices across all eight states and territories. The two sets of figures are not comparable with each other, because this page uses the original rubric and that one uses the corrected version, but the Australian study reports what the current methodology finds.

If you run a practice

Check your own site using the corrected methodology with the free AI readiness scan, or see whether the rest of the web agrees about your practice details with the listings check.