Findings · 2026-10-09 · By VSNARY | Emmanuel Orta · 0 views
Four AI models, three local businesses: who got it right
We asked Gemini with Google Search and three open-weights models for the phone number, address and hours of three home-service businesses we run in Denver. The phone numbers were right once there was search. The addresses came from directories, and one model invented everything.
On 9 October 2026 we asked four AI models for the phone number, address and hours of three Denver businesses we run. Gemini with Google Search gave the right phone number for all three, but gave a street address to two service-area businesses that publish none: one copied from a YellowPages listing, one belonging to a different business with the same owner. gpt-oss-120b, answering without search, invented phone numbers, streets and weekday hours for two of them. Llama 3.3 and Mistral Small 3.1 declined every time.
When a customer asks an assistant for a business’s phone number, the assistant’s answer is the business’s listing, whether or not the business ever saw it. We run three home-service businesses in the Denver area, so we asked about them. Each business is one we know the true answers for, which is the only way to grade an answer properly.
The question was the one a customer types: “What is the phone number, address and opening hours of Supreme Fencing Denver in Denver, CO?”, and the same for Supreme Fencing Arvada and Tree Service Denver. All three are service-area businesses: crews go to the customer, and none publishes a street address. Supreme Fencing Denver and Tree Service Denver answer the phone around the clock.
Four models answered between 9:06 and 9:28 AM Mountain Time on 9 October 2026: Gemini with Google Search (Google), which searches and cites its sources, and three open-weights models answering from training alone: gpt-oss-120b (OpenAI), Llama 3.3 70B (Meta) and Mistral Small 3.1 (Mistral AI).
| Model | Phone | Address | Hours |
|---|---|---|---|
| Gemini with Google Search | Right, 3 of 3 | A street for 2 businesses that have none; 1 correctly “mobile service” | Right for both 24/7 businesses |
| gpt-oss-120b (no search) | Invented for 2; declined 1 | Invented streets for 2 | Invented weekday hours for 2 |
| Llama 3.3 70B (no search) | Declined 3 of 3 | Declined | Declined |
| Mistral Small 3.1 (no search) | Declined 3 of 3 | Declined | Declined |
Search fixed the phone numbers #
Gemini with Google Search gave the correct number for all three businesses, and for all three it cited the business’s own website among its sources. For the two businesses that answer around the clock, it said so. Asked about Tree Service Denver, it also did something useful: it noticed that “Tree Service Denver” describes several companies, said which one it meant, and described the business as a mobile service with no storefront, which is exactly right.
The addresses came from somewhere else #
For the other two, Gemini supplied a street address the businesses do not have:
- Supreme Fencing Denver: 3015 Welton St, Denver. That is the street YellowPages prints on its listing for the business. Gemini cited the business’s own site, Trustpilot and Thumbtack, not YellowPages, so the street reached the answer by a route the citations do not show.
- Supreme Fencing Arvada: a street that belongs to a different business with the same owner. Nothing on the fencing site mentions it. The model joined two businesses that share an owner and gave one the other’s address.
Both errors have the same shape: a service-area business publishes no street, and the model fills the gap from whatever street it can find nearby. For a customer the cost is small: the phone was right. For the business it means a map pin at an address it does not operate from, repeated by every assistant that reads the same sources.
One model invented a business card #
Without search, gpt-oss-120b produced confident tables for two of the three businesses:
- Supreme Fencing Denver:
(303) 777-7777, 1200 S Broadway, Suite 100, open 9 to 6 on weekdays and closed Sunday. - Supreme Fencing Arvada:
(720) 555-1234, “the number most often shown in online directories”, at 1150 W 44th Ave, open 8 to 5 on weekdays.
None of it is real. The 555 exchange is the one films use, and seven sevens is a placeholder. Both answers closed with advice to call ahead and check the details, which is the right advice and cannot undo a table that looks like a listing. For Tree Service Denver the same model declined, which shows the invention is not systematic: it depends on how familiar the name sounds.
Declining is the honest answer #
Llama 3.3 70B and Mistral Small 3.1 declined all three questions and pointed the customer to Google, Maps, Yelp and other directories. With no way to look anything up, that is correct. We count a decline as no answer, not as a wrong one, and we count an invented value as wrong even when the model adds a caveat. A hedged answer that still prints a number is still printing a number.
What a business can do about it #
- Say plainly that you have no storefront. State on the site, in your structured data and on your Google Business Profile that you serve customers at their location. Gemini got this right for the one business whose site says it clearly.
- Fix the directories, not just the site. An answer engine that cites your site can still take its street from a directory it does not cite. Our listing audit reads what each directory prints.
- Keep sibling businesses apart. Businesses with one owner should not share addresses, numbers or profiles they do not actually share, or a model will merge them.
- Ask again. Answers change from one asking to the next. A single reading is a snapshot; a daily one is a record.
Method #
Each model was sent the question once, with nothing about the business beyond the question itself. Each answer was compared field by field with the business’s own published record: phone right, wrong or not stated; street right, wrong, given to a business that has none, or city only; hours compared only where the comparison is unambiguous. Values a model offers as an example are set aside; values it presents as the business’s are judged. Every wrong value is traced to the directories that print it and to other businesses on the same account. The readings are signed and public: Supreme Fencing Denver, Supreme Fencing Arvada, Tree Service Denver. Each business is asked again every day; this post describes the readings of 9 October 2026. Disclosure: CrawlCheck and the three businesses share an owner.
Every figure above came out of CrawlCheck.
Point it at your own domain and see the same measurements, free.
The main product
Found this on your own site? We fix it for $749.
Scan free to see where you stand. The fix is one site, every finding implemented and re-measured, with a sealed before and after.
Questions this post answers
Do AI assistants give the right phone number for local businesses?
In our test of three Denver businesses on 9 October 2026, Gemini with Google Search gave the right number for all three. Models answering without search either declined or, in gpt-oss-120b’s case, invented numbers for two of the three.
Why would an AI give my business the wrong address?
Most often because a directory prints one. For businesses that publish no street, models fill the gap: in our test one street came from a YellowPages listing and another from a different business with the same owner.
Is it a problem if an AI model says it cannot find my business?
No. A model with no search that declines is giving the honest answer. The problem is a model that invents a plausible number or street and presents it as your listing.
How do I stop AI from giving my service-area business a street address?
State clearly on your site, in your structured data and on your Google Business Profile that you serve customers at their location, then correct every directory that still prints a street. Check again after the changes, because models update from different sources at different times.
Which AI models were tested?
Gemini with Google Search (Google), and three open-weights models without search: gpt-oss-120b (OpenAI), Llama 3.3 70B (Meta) and Mistral Small 3.1 (Mistral AI). Each was asked the same customer question on 9 October 2026.
Can I check what AI says about my own business?
Yes. Ask each assistant the question a customer would, and compare each field with your own record. CrawlCheck runs this daily for businesses on a licence and publishes a signed reading per business.
Related findings
Comments
Comments are read before they appear. Nothing is published automatically, and no account is needed.
Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.