Ask a chat model such as ChatGPT, Claude, Gemini or DeepSeek for an available .com and it will usually hand back a confident list. In our 10-model run, 1,702 of the 2,449 names offered that way were already registered. A model answering from memory has no registry lookup, so the list reflects names that sound open. The leaderboard below shows how often each model missed.
This page is the AI Domain Hallucination Index: 2,449 suggestions from 10 models, run on 2026-09-01 (source: leaderboard CSV), each checked live via RDAP against the authoritative .com registry. “Available” means the registry returned no active registration at check time. Verify a name the AI just gave you, or generate .coms that passed an RDAP check. The AI domain name generator and how to check if a domain is taken use that check too.
The leaderboard
Each model was asked for available .com names; the n column shows how many verified suggestions each produced. The percentage is how many of those picks were free to register. Overall, 30.5% of the 2,449 suggestions were actually available (source: leaderboard CSV). The best model in this run, anthropic/claude-sonnet-5, was 50.5% available (n=287). The weakest, anthropic/claude-haiku-4.5, was 8.2%: 259 of its 282 picks were already taken (91.8%). Even the top row left 49.5% of its picks taken. google/gemini-2.5-pro is shown for completeness but was rate-limited to a sample too small to rank.
| Model | Actually available | n | |
|---|---|---|---|
| anthropic/claude-sonnet-5 | 50.5% | 287 | |
| anthropic/claude-opus-4.8 | 46% | 287 | |
| openai/gpt-5.6-luna | 38.2% | 288 | |
| deepseek/deepseek-v4-pro-0813 | 32.6% | 291 | |
| google/gemini-2.5-pro low sample | 31.7% | 41 | |
| mistralai/mistral-nemo | 31% | 168 | |
| google/gemini-3-flash-preview | 22.9% | 288 | |
| nousresearch/hermes-3-llama-3.1-70b | 21.9% | 247 | |
| meta-llama/llama-3.1-8b-instruct | 21.1% | 270 | |
| anthropic/claude-haiku-4.5 | 8.2% | 282 |
Raw leaderboard CSV: ai-hallucination-index.csv (CC BY 4.0). Per-model write-ups: why DeepSeek V4 Pro suggests taken domains and why GPT suggests taken domains.
November 2026 check: four newer Claude models
On 2026-11-01 we asked four newer Claude models for 12 available .com names in each of 24 business types and checked all 1,152 picks on Verisign's RDAP service: 47.1% were available and 52.9% were already taken (source: November 2026 results CSV).
| Model | Actually available | n | |
|---|---|---|---|
| claude-opus-5-5 | 79.2% | 288 | |
| claude-fable-5-1 | 57.3% | 288 | |
| claude-sonnet-5-5 | 34.4% | 288 | |
| claude-haiku-5-5 | 17.7% | 288 |
Opus 5.5 led at 79.2%, and 60 of its 288 picks were still taken. A free name can still be a weak one: 85% of Opus 5.5's names ended in one of 12 four-letter endings, as in roastlyne.com and steamorra.com, and 81% of Haiku 5.5's did the same. SaaS startup names were free least often (27.1%) and catering company names most often (68.8%).
Each model answered as a Claude Code agent told to answer from its own knowledge with no lookups (its session log shows none), covering all 24 business types in one session. September's models answered through the API with one call per type, so these rates are kept out of the September leaderboard. None of these models checked a name before suggesting it; the DSK MCP server runs an RDAP check on each name before returning it.
Paste the name the model gave you
Check it live, or generate names that pass the registry before they reach the screen.
Verify a domain → Generate available names →Availability by business type
The miss rate is not even. Across all 10 models, retail and professional-services names rank lowest: online store 10.6% available (n=94), SaaS startup 17.4% (n=86), consulting firm 19% (n=84). All 24 business types are in the table below, from the per-industry CSV. No category reached 50% available (the best, yoga studio, was 49.4%), and 14 of the 24 were under 30%. The two lowest rows, online store and SaaS startup, are the niches the Shopify name generator and SaaS name generator serve, so a name from either needs its own registry check before you build a logo around it.
| Business type | AI picks free to register | n |
|---|---|---|
| Online store | 10.6% | 94 |
| SaaS startup | 17.4% | 86 |
| Consulting firm | 19.0% | 84 |
| Restaurant | 21.8% | 133 |
| Tutoring service | 22.4% | 107 |
| Real estate agency | 22.7% | 97 |
| Gym | 24.2% | 120 |
| Marketing agency | 25.8% | 97 |
| Auto repair shop | 26.2% | 107 |
| Law firm | 26.4% | 106 |
| Accounting firm | 26.6% | 109 |
| Catering company | 27.9% | 86 |
| Photography studio | 28.3% | 99 |
| Cleaning service | 29.8% | 84 |
| Dental practice | 33.3% | 96 |
| HVAC company | 35.1% | 97 |
| Coffee shop | 39.0% | 123 |
| Pet grooming | 39.0% | 123 |
| Roofing company | 39.1% | 110 |
| Plumbing company | 39.2% | 97 |
| Landscaping business | 39.2% | 97 |
| Bakery | 40.9% | 132 |
| Hair salon | 47.5% | 80 |
| Yoga studio | 49.4% | 85 |
Some “available” names were registered in the 1990s
These rows come from our published registration-date sample of names that models called “available” but that the registry showed as registered:
| Name the AI called “available” | Registered since |
|---|---|
| haven.com | 1993 |
| reprise.com | 1994 |
| comfort.com | 1995 |
| musclemania.com | 1996 |
| javajunction.com | 1997 |
| statutory.com | 1998 |
| brightsmiledental.com | 1999 |
Source: ai-hallucination-index-registration-dates.csv (CC BY 4.0). The file has 20 rows but 19 distinct names, because haven.com is listed twice; the table shows 7 of the 19. This small illustrative sample skews old (median registration year 1998, with or without the duplicate). The larger 240-name sample of taken suggestions from the same run, cited in the FAQ, has a median of 2018 and is the better guide to typical age.
Why a chat model cannot check the registry
A language model predicts the next plausible token, and without a lookup tool it cannot see the registry, so nothing in that prediction step tells it whether a .com is registered. It has no way to know whether a name was bought last week or decades ago, so what it reports is how available a name sounds. One possible reason, which this run did not test: short, brandable .coms may sound free to a model because similar strings appear often in its training text, such as blog posts, old directories and “great name” lists.
Availability and trademark clearance are two different searches. An AI-suggested name can also collide with a live mark the model never checked. Confirm both before you commit; our domain trademark check guide covers the trademark half.
Separate test: Domain Search King MCP (our own product, not part of this index)
In a June 2026 test, 76 names returned by the Domain Search King remote MCP were rechecked by RDAP: 75 were available and 1 was taken (98.7%). The MCP RDAP-checks each name before returning it, which is the step a chat model answering from memory skips. 76 names is a small sample, so treat 98.7% as indicative. The test ran in June, outside the September run, so the leaderboard and its CSV have no Domain Search King row. The MCP works with Claude, Cursor and any MCP-capable agent.
Limits of the September 2026 run
Every rate on this page, outside the November 2026 section, describes the models as they answered on 2026-09-01 to one prompt set; a later run could rank them differently. A registration record found at check time means the suggestion was already registered, and a registry error counts as inconclusive. The public files give per-model, per-industry and per-TLD rates; they do not include per-domain check timestamps. The registration-date sample illustrates older registered names and is not the complete run.
What the rates mean for a shortlist
Take a shortlist of 10 names from a chat model. At the overall rate of 30.5% available, expect about 3 to be free and 7 to be registered already. From the worst row, claude-haiku-4.5 at 8.2%, expect fewer than 1 in 10. From the best row, claude-sonnet-5 at 50.5%, expect about 5. Those are averages from one run, so a single list can land well above or below them, but only one of the 10 rows got even half its picks right.
The order you check in matters. Check names before you design anything around them. Run every name on the list through the checker above before ranking them, drop the registered ones, and only then choose. If your favourite is gone, the domain taken guide lists the fallback moves, and why good AI domain names are taken explains why the short dictionary words go first.
Rechecked on 2026-10-15
To confirm the old registrations still hold, we queried Verisign's RDAP service for haven.com on 2026-10-15. It returned a registration event dated 1993-11-08 and an expiry of 2027-11-07. The same day, this page's checker (/api/rdap) returned "available":false for both haven.com and javajunction.com. Both names were still registered, so a model that lists either as available is as wrong today as it was on 2026-09-01.
How we measured it
Each model received the same prompt: available .com names across dozens of business niches and naming styles (freeform, one-word, prefix/suffix). Every suggestion was checked live against the .com registry. Available means no active registration at check time.
2,449 suggestions sit on the leaderboard, from a single run stamped 2026-09-01; every row is in the public leaderboard CSV, so the per-model rates can be recomputed from source. To reproduce it, ask a model for “available” .coms and check each name by RDAP (for .com, Verisign's RDAP service at rdap.verisign.com). A registrar search is a close substitute but not the same check.
Raw data is free to reuse under CC BY 4.0 with a link back: leaderboard CSV · registration-year sample · per-industry CSV · per-TLD CSV.
FAQ
Why do AI models suggest taken domains?
A language model predicts plausible names. Unless it is given a lookup tool, it does not query the domain registry, so it cannot tell a free name from a registered one. In the 2026-09-01 run, 69.5% of the 2,449 names models offered as “available” were already registered.
How was availability measured?
Every suggested .com was checked live against the authoritative .com registry. “Available” means the registry returned no registration at check time. 2,449 suggestions were verified in the 2026-09-01 run (source: leaderboard CSV).
How long have the taken domains been registered?
In a 240-name sample of taken suggestions from the 2026-09-01 run, the median registration year was 2018 and the oldest was 1993.
What should I do with a name an AI just gave me?
Check it against the registry before you get attached to it. If it is taken, generate alternatives that passed an RDAP check, or follow the AI-taken rescue flow. A domain a model invents for a real brand, if unregistered, can be bought by anyone, including attackers; Palo Alto Networks’ Unit 42 documented such campaigns (The Hacker News). We cover the risk under phantom squatting.
Which AI model suggested the most available domains?
In the 2026-09-01 run, anthropic/claude-sonnet-5 had the highest rate: 50.5% of its 287 picks were available, so 49.5% were still taken. The other nine models scored lower, between 8.2% and 46%, so every model's list needs a registry check.
Are Domain Search King results always current?
Names are RDAP-checked before display, but results for popular keywords can be served from a cache up to about 30 hours old. Any check is a snapshot, and a name can be registered minutes later. The paste-a-name checker on this page does an uncached lookup each time.
Are the register links affiliate links?
Yes. Register links may earn Domain Search King a commission at no extra cost to you. The registrar sets the price, so confirm it at checkout.
