Can you trust the phone number AI gives you? we checked 80,657 of them
When an AI assistant recommends a local business, it often hands over a phone number or a web address too. We checked every phone number 12 assistants gave in 4,100 US markets against the business's own website, and all 114,275 web addresses against public DNS. Who gets it right depends almost entirely on how the assistant finds its answers.
- ChatGPT
- Claude
- Gemini
- Perplexity
- Google AI Overviews
- Microsoft Copilot
- Grok
- Meta AI
- DeepSeek
- Mistral
- Siri
- Alexa
In AI Syndicate Research's November 2026 study of 4,100 US local markets and 12 AI assistants, phone numbers from live-search assistants (ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Microsoft Copilot) matched a number on the business's own website 86.7% of the time; numbers from memory-only assistants (Grok, Meta AI, DeepSeek, Mistral) matched 23.3%. The memory-only group gave 49% of all the numbers, and 21% of the web addresses it gave were domains that are not registered.
Executive summary
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
Key findings
- Live-search assistants mostly get the number right. When ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews or Microsoft Copilot gave a phone number we could check, it matched a number on the business's own website 86.7% of the time (95% range 85.6–87.8%). ChatGPT was highest, at 92.3%.
- Memory-only assistants mostly get it wrong. Numbers from Grok, Meta AI, DeepSeek and Mistral matched only 23.3% of the time (19.9–26.9%). Mistral matched 23.3% and Meta AI 22.8%. Most of their other numbers appear nowhere on the business's site.
- And they hand out about half the numbers. Mistral put a phone number in 52% of its answers, almost as often as ChatGPT (62%). Of the 80,657 numbers the 12 assistants gave, 49% came from the memory-only group.
- The wrong numbers look real. 99% of all numbers carry an area code from the right state. One giveaway: in 17.4% of Mistral answers that give two or more businesses a number, two of them end in the same four digits, against 0.4% for ChatGPT.
- Some web addresses lead nowhere. 21% of the web addresses the memory-only assistants gave are domains that are not registered at all (21% for Mistral, 17% for Meta AI). For the live-search assistants it is 0.1%.
- Being AI's top pick does not protect you. Memory-only assistants matched 25% of numbers for businesses ranked in a market's top three, no better than for the rest.
What it means. A customer who asks an AI assistant for a plumber and calls the number in the answer may not reach that plumber. The live-search assistants copy numbers from the web and mostly get them right. The assistants that answer from memory write numbers that look plausible and usually are not the business's. A business cannot correct a memory model's answer quickly, but it can make sure the number the live-search assistants find is the same everywhere. These are observed patterns, not proof of cause.
Scope. 4,100 markets: 205 metros in all 50 US states, times 20 types of business. 12 assistants, 3 fixed questions each, asked between 1 and 9 November 2026: 149,500 answers naming 715,019 businesses. The method, its limits and what we hold back are set out at the end.
How did we check the phone numbers and websites?
This paper reads the November 2026 edition of AI Syndicate Local, our public index of which local businesses AI assistants recommend. Each month we ask 12 assistants the same three questions about a type of business in a metro area: the best one, the one to recommend to someone who needs one, and an affordable one nearby. Nobody asked for contact details, so every phone number and web address in this paper is one the assistant chose to give.
- Live web search: ChatGPT, Claude, Gemini and Perplexity answer with web search on. Google AI Overviews is the overview Google shows for the question. Microsoft Copilot answers come mostly from Microsoft's Bing-grounded AI service, and otherwise are reconstructed from Bing results.
- Model memory only: Grok, Meta AI, DeepSeek and Mistral answer from what the model learned.
- Reconstructed voice answers: Siri and Alexa cannot be asked programmatically. We reconstruct their answers from the kinds of sources each draws on, in the short style each speaks in.
Phone numbers. We found every US phone number in every answer (80,657 in all) and tied each to the business named just before it in the answer text. To check a number, we need the business's real numbers. We used the business's own website: the one on its business listing, not one taken from the AI answers, and only where our crawl of that website found a phone number on it. A number is right when it is one of the numbers on that website. That gave us 7,735 checkable numbers for 3,869 businesses. Websites shared by a chain's locations, such as a national brand site, are left out, because they show a national line rather than the local office's.
Web addresses. For every business an answer named, we recorded the web address it gave, if any. Almost all of these appear word for word in the answer text. We looked up every domain in public DNS: a domain that does not exist is not registered. Where we know the business's own website from its listing, we also checked whether the address leads there, to a listing page such as a directory or review site, or to some other website.
Two checks that need no answer key. For every number, we checked whether its area code belongs to the market's state. And for every answer that gave two or more businesses a number, we checked whether any two of those numbers end in the same four digits. Real businesses in the same city rarely share their last four digits.
How often do AI assistants give a phone number?
It depends almost entirely on the assistant. Across all 12, 17.1% of answers included at least one phone number. ChatGPT and Mistral did it most: 62% and 52% of their answers. Grok, Siri, DeepSeek, Gemini, Alexa and Claude did so in fewer than 4% of their answers.
Web addresses follow the same split. ChatGPT linked 99% of the businesses it named, and Mistral gave an address for 66%. Most other assistants rarely wrote one out, and Perplexity and Claude cite their sources instead. Together, ChatGPT and Mistral gave 92% of the 114,275 web addresses in the answers.
So the assistant a customer happens to use decides whether they get a number at all, and, as the next section shows, whether it is likely to work. About half of all the numbers in AI's local answers (49%) come from the four memory-only assistants.
How often is the phone number AI gives the right one?
For the live-search assistants, usually. For the memory-only ones, usually not. Of the 7,735 numbers we could check, 77.3% matched a number on the business's own website. That average hides two very different groups: 86.7% for the six live-search assistants and 23.3% for the four that answer from memory.
Show the numbers
| Type of assistant | Match the business's website |
|---|---|
| Live web search | 86.7% |
| Voice (reconstructed) | 80.0% |
| Memory only | 23.3% |
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
Broken out by assistant, the split is just as clean: every live-search assistant beats every memory-only one.
ChatGPT, the assistant most people use, matched 92.3% of the time, Claude 84.2% and Perplexity 82.7%. Gemini, Google AI Overviews and Microsoft Copilot matched about two times in three. Mistral and Meta AI, the two memory-only assistants that give numbers most often, matched fewer than one time in four. The reconstructed Alexa answers matched 81.4%, on few numbers.
The table shows all 12 assistants.
| Assistant | Answers with a number | Numbers given | Numbers we could check | Match the business's website |
|---|---|---|---|---|
| ChatGPT | 62.0% | 23,872 | 4,146 | 92.3% (91.4–93.3) |
| Claude | 3.6% | 899 | 146 | 84.2% (78.6–89.9) |
| Gemini | 3.4% | 1,125 | 170 | 67.1% (58.5–75.4) |
| Perplexity | 36.6% | 8,769 | 1,348 | 82.7% (80.3–85.0) |
| Google AI Overviews | 8.8% | 2,627 | 225 | 68.4% (61.9–75.3) |
| Microsoft Copilot | 13.0% | 3,176 | 492 | 67.1% (61.6–72.5) |
| Grok | 0.4% | 59 | 0 | too few to check |
| Meta AI | 18.3% | 5,228 | 202 | 22.8% (16.3–29.5) |
| DeepSeek | 2.6% | 362 | 5 | too few to check |
| Mistral | 51.9% | 34,054 | 941 | 23.3% (19.4–27.1) |
| Siri | 0.1% | 23 | 1 | too few to check |
| Alexa | 3.5% | 463 | 59 | 81.4% (69.6–90.8) |
Grok (59 numbers, 0 checkable), DeepSeek (362 numbers, 5 checkable), Siri (23 numbers, 1 checkable) gave too few checkable numbers to rate. Grok and DeepSeek rarely give a number at all, and Siri's reconstructed spoken answers almost never do.
A number that does not match is not always a wrong number. A business may use a separate line for calls from ads, or have a second office. That is why the live-search assistants do not reach 100%. But that cannot explain a gap of 63 points between the two groups, on the same kinds of business in the same markets.
What do the wrong numbers look like?
Mostly like real numbers. When a memory-only assistant's number did not match, it was usually a number we could not find on any business's website: 69.9% of their checked numbers. Toll-free lines that are not on the business's site made up 2.9%, numbers listed on a different business's website 1.3%, and numbers one or two digits away from the real one 2.6%.
Show the numbers
| What the number turned out to be | Share of numbers |
|---|---|
| Not found on any business website we crawled | 11.3% |
| One or two digits off the business's number | 0.5% |
| Listed on a different business's website | 0.8% |
| A toll-free line not on the business's website | 0.7% |
| On the business's own website | 86.7% |
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
Show the numbers
| What the number turned out to be | Share of numbers |
|---|---|
| Not found on any business website we crawled | 69.9% |
| One or two digits off the business's number | 2.6% |
| Listed on a different business's website | 1.3% |
| A toll-free line not on the business's website | 2.9% |
| On the business's own website | 23.3% |
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
The reconstructed Siri and Alexa answers gave only 60 checkable numbers; 80.0% matched.
Nothing about the wrong numbers gives them away at a glance. 99.2% of all the numbers carried an area code from the market's own state, and the memory-only assistants were no worse at this than the others (Mistral 99.5%, ChatGPT 99.0%). The models know which area codes belong to which city. They do not know the rest of the number.
One pattern does stand out. When an answer gave two or more businesses a phone number, we checked whether two of those numbers ended in the same four digits. For real businesses in the same city that is rare. In ChatGPT's answers it happened 0.4% of the time. In Mistral's, 17.4%.
The most extreme case: one Mistral answer about auto repair in Georgia gave nine different businesses three different phone numbers, and every one ends in 1111. Real numbers almost never line up like that.
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
That is the mark of a number assembled from a pattern rather than copied from a source. We do not publish any of the wrong numbers, and we did not call them. Some will belong to other people or businesses.
Do the web addresses AI gives actually work?
Most do, but not the ones the memory-only assistants write out. We looked up all 43,696 domains in the 114,275 web addresses the assistants gave. Of the addresses from Grok, Meta AI, DeepSeek and Mistral, 20.9% (95% range 20.4–21.4%) point to a domain that is not registered: no website has ever been set up there, or it has lapsed. From the live-search assistants, 0.1%.
These are not typos in our data. Virtually all of the addresses appear word for word in the assistant's answer. The memory-only assistants build an address from the business's name, the way a person might guess one, and often guess a domain nobody owns.
For the 16,029 addresses given for businesses whose real website we know from their listing, we followed each one.
Show the numbers
| Where the address led | Share of addresses |
|---|---|
| A domain that is not registered | 0.0% |
| A broken, parked or unreachable site | 1.6% |
| A different working website | 5.2% |
| A listing page: directory, review site, social profile | 25.1% |
| The business's own website | 68.1% |
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
Show the numbers
| Where the address led | Share of addresses |
|---|---|
| A domain that is not registered | 11.1% |
| A broken, parked or unreachable site | 8.7% |
| A different working website | 14.8% |
| A listing page: directory, review site, social profile | 2.2% |
| The business's own website | 63.2% |
Source: AI Syndicate Research, Can You Trust the Phone Number AI Gives You?, November 2026 (aisyndicate.com)
The live-search assistants sent people to the business's own site 68% of the time and to a listing page about it, such as a review site or directory, most of the rest. The memory-only assistants reached the right site 63% of the time, but 15% of their addresses led to some other working website and 20% led nowhere.
Better-known businesses fare better here. Of the addresses memory-only assistants gave for a business ranked in its market's top three, 8.4% were unregistered domains, against 17.2% for businesses ranked lower and 24.8% for businesses no other assistant confirmed.
Does it depend on the type of business or its ranking?
The gap between the two groups holds in every type of business we could check. The table shows the share of checked numbers that match, with the number of checked numbers in brackets.
| Type of business | Live search | Memory only |
|---|---|---|
| Med spas | 93.5% (153) | too few (0) |
| Pest control | 91.7% (336) | too few (20) |
| Roofing | 91.6% (416) | 17.5% (40) |
| Water damage restoration | 91.2% (362) | 5.9% (51) |
| Window and door replacement | 91.2% (339) | 0.0% (33) |
| Solar installation | 90.4% (94) | too few (0) |
| Plumbers | 89.9% (476) | 8.9% (79) |
| HVAC and air conditioning | 89.2% (640) | 11.0% (118) |
| Electricians | 88.9% (425) | 7.0% (43) |
| Auto repair | 88.1% (385) | too few (4) |
| Cosmetic dentists | 87.8% (378) | too few (10) |
| Personal injury lawyers | 87.6% (362) | 37.6% (205) |
| Garage door repair | 86.5% (726) | 0.0% (134) |
| Moving companies | 84.3% (281) | too few (2) |
| Speeding ticket lawyers | 80.4% (445) | 40.7% (135) |
| Criminal defense lawyers | 76.9% (415) | 44.9% (185) |
| Business litigation lawyers | 68.2% (258) | 20.3% (74) |
The memory-only assistants did best with lawyers: 38–45% of their numbers for criminal defense, speeding ticket and personal injury firms matched, perhaps because law firms advertise one number widely and for years. For trades such as garage door repair, electricians and plumbers they almost never got it right. Hotels, real estate teams and restaurants had too few checkable numbers to include: these businesses' websites rarely list a number our crawl could read, or the businesses belong to chains.
A business's place on the ranking makes little difference. The live-search assistants matched 87.8% of numbers for businesses ranked in their market's top three and 85.3% for those ranked 4th to 10th. The memory-only assistants matched 25.1% and 19.5%. Being AI's favorite does not mean AI knows your number.
What should a local business do about it?
For a local business.
- Make your number easy to copy. The live-search assistants get it right because they copy it from your website, your listings and the review sites. Put one main number in plain text on every page of your site, in your site's structured data, and on every listing, and keep them the same.
- If you change your number, change it everywhere. An old number on a directory can be the one an assistant picks up.
- Own the domain your name suggests. Memory-only assistants guess web addresses from business names. If the obvious one is not yours, consider buying it and pointing it at your site.
- Check what the assistants say about you. Ask the main assistants for your trade in your city and look at the number and address they give. You cannot quickly change what a memory-only model learned, but you can know when it is sending customers astray, and tell customers which number to use.
For anyone using AI to find a business. Before you call, open the business's website or listing and use the number there. A number from an assistant that cites a source is more likely to be right than one from an assistant that answers from memory.
These are patterns in what the assistants said, not guarantees.
How this data is used. This research feeds the AI Syndicate platform, so our clients have the leading edge. Clients see what each of the 12 assistants says about their business, including the phone numbers and web addresses it gives, inside their dashboard, and see where an assistant gets a fact about them wrong.
How was this measured, and what can it not tell you?
Data. AI Syndicate Local, November 2026 edition: 4,100 markets (205 metros in 50 states, 20 types of business), 12 assistants, 3 fixed questions each, asked between 1 and 9 November 2026 with the city named in the question and no user location or account history. 149,500 answers named 715,019 businesses. A model extracts the business names and any web address given for each, in order. The edition also has early markets from Canada, the United Kingdom and six new trades; they are left out here.
Phone numbers. A phone number is any US-format number of ten digits, with or without a leading 1 and the usual separators. Each is tied to the business whose name, matched on its first three words, appears closest before it in the answer. 78,589 of 80,657 numbers could be tied to a business; 19 were not valid US numbers.
What counts as right. A number is checked only when its business is on the market's ranking, the ranking's website for the business came from its business listing rather than from the AI answers, that website is not shared by locations in three or more metros, and our crawl of the site found at least one phone number (29,406 sites). A number is right when it is one of those numbers. Toll-free lines, numbers on another business's website and near misses (one or two digits different from a number on the site) are counted separately; the rest are numbers not found on any website we crawled.
Web addresses. Every domain was looked up in DNS on 16 November 2026; a domain that a public resolver reports as non-existent (NXDOMAIN) is not registered. Addresses that differed from the business's listed website and were not listing pages were opened once; a site that did not respond, returned an error or showed a parked or for-sale page is broken. A listing page is a directory, review site, social profile, licensing or bar-association site, or any site the answers gave for five or more differently named businesses that is not the website of three or more ranked businesses.
What this cannot tell you.
- A number that does not match the business's website may still reach the business, through a tracking line or a second office. The match rates are lower bounds on accuracy for every assistant. We did not call any number.
- The check covers businesses whose own website lists a number we could read, and leaves out national brand sites. Hotels and restaurants are mostly left out for that reason.
- Only some businesses could be checked, and which ones differs by assistant: the memory-only assistants name more chains and more businesses no other assistant confirms. Rates compare like with like within the checked set, not across every number given.
- It is one month. The assistants and their sources change.
- Siri and Alexa answers are reconstructions, read as close approximations, not recordings. Siri's spoken answers rarely include a number.
- It measures what assistants say, not what customers do.
What stays private. We do not publish any wrong number or the business it was given for. Per-assistant answers and per-business results are not published.
Statistical notes
- Unit of measurement. Phone rates count each checked number once; a business given the same number in several answers counts each time. Web address rates count each address given.
- Ranges. 95% ranges come from resampling markets 1,000 times. They describe uncertainty across markets.
- Minimum sample. An assistant's rate is shown only when it rests on at least 30 numbers, answers or addresses; below that the table says too few.
- Sensitivity. Counting only numbers within 200 characters of the business's name in answers where every named business was found, the live-search rate is 86.5% and the memory-only rate 27.1%. Checking against any website on the ranking, including ones taken from AI answers, they are 88.8% and 18.5%.
How can I cite or reuse these numbers?
Quote any figure in this report with credit to AI Syndicate Research and a link to this page. You may quote up to 50 words of text, or share a single chart with its source line intact. Reproducing the report or large parts of it needs written permission. A suggested citation:
A PDF copy and a plain-text Markdown version are available. The rankings behind the report are at aisyndicate.com/ai-local/. Journalists, businesses and researchers who need a cut of the data that is not here can write to support@aisyndicate.com.
Key figures
Every key figure in this report, in one table. The same figures are in findings.json and findings.csv, free to quote with credit. Ranges are 95% intervals from resampling markets.
| Finding | Value | 95% range | Sample |
|---|---|---|---|
| Share of phone numbers from live-search AI assistants that match a number on the business's own website | 86.7% | 85.6–87.8% | 6,527 numbers |
| Share of phone numbers from memory-only AI assistants that match a number on the business's own website | 23.3% | 19.9–26.9% | 1,148 numbers |
| Share of ChatGPT's phone numbers that match the business's own website | 92.3% | 91.4–93.3% | 4,146 numbers |
| Share of Mistral's phone numbers that match the business's own website | 23.3% | 19.4–27.1% | 941 numbers |
| Share of Meta AI's phone numbers that match the business's own website | 22.8% | 16.3–29.5% | 202 numbers |
| Share of all phone numbers in AI's local answers that come from memory-only assistants | 49.2% | 80,657 numbers | |
| Share of ChatGPT answers that include a phone number | 62% | 12,513 answers | |
| Share of Mistral answers that include a phone number | 51.9% | 12,513 answers | |
| Share of phone numbers in AI answers with an area code from the market's state | 99.2% | 74,964 numbers | |
| Share of Mistral answers giving two businesses phone numbers that end in the same four digits | 17.4% | 16.4–18.5% | 6,184 answers |
| Share of web addresses from memory-only AI assistants whose domain is not registered | 20.9% | 20.4–21.4% | 59,995 addresses |
| Share of web addresses from live-search AI assistants whose domain is not registered | 0.1% | 0.1–0.1% | 54,245 addresses |
Sources
- AI Syndicate Local: the November 2026 edition, AI Syndicate. Read Nov 16, 2026.
- AI Syndicate Local methodology: how the index is measured, AI Syndicate. Read Nov 16, 2026.
Questions about phone numbers and websites in AI answers
Are the phone numbers ChatGPT gives for local businesses accurate?
Mostly. In AI Syndicate Research's November 2026 study of 4,100 US markets, 92.3% of the phone numbers ChatGPT gave for local businesses that we could check matched a number on the business's own website. Across the six live-search assistants the figure was 86.7%.
Which AI assistants make up phone numbers?
The ones that answer from memory. Numbers from Grok, Meta AI, DeepSeek and Mistral matched the business's own website only 23.3% of the time in November 2026 (Mistral 23.3%, Meta AI 22.8%). Mistral put a number in 52% of its answers.
Do AI assistants give fake website addresses?
Sometimes. 20.9% of the web addresses the memory-only assistants gave were domains that are not registered at all, against 0.1% for live-search assistants such as ChatGPT.
How can I tell if a phone number from AI is wrong?
Usually you cannot from the number alone: 99% carry an area code from the right state. Check it against the business's own website or listing before you call, and prefer an assistant that shows its sources.
How can a business make AI give the right phone number?
Publish one main number in plain text on every page of your website, in its structured data and on every listing, and keep them identical. Live-search assistants copy numbers from those sources. This is an association, not proof of cause.
Can I quote these numbers?
Yes. Quote any figure with credit to AI Syndicate Research and a link to this report. The rankings behind them are at aisyndicate.com/ai-local/.
Related pages.
See what AI says about your business.
Run the free check to see whether the AI assistants recommend your business, and what they say about it. Or open the AI Local index to see every ranked business in your market.