# Can ChatGPT Find Someone's Email Address?

> No. ChatGPT and Claude only find addresses that are already published. Otherwise they guess, and in our data first.last@ is right for just 25% of people.

Oct 22, 2026 · Cristian Frunze · 12 min · Lead Generation

Tags: can chatgpt find email addresses, chatgpt email finder, ai email finder, email address format, email verification, cold email bounce rate

![A founder tries key after look-alike key in a purple door while his parrot cheers, showing AI guesses that look right but don't fit.](/images/blog/can-chatgpt-find-email-addresses-cover.webp)

We made up a founder, Marta Kowalczyk, gave her a six-person studio, and asked eight AI setups for her email address. None of them knew it. Six of them typed out the same guess anyway.

**No. ChatGPT, Claude and the other chatbots can't look up someone's email address. With web search on, they can find an address that's already published on a public page. Otherwise they guess from common patterns, and one guess is usually wrong. In our data, first.last@ matched the real address for only 25% of people.**

An email address is part pattern, part arbitrary fact. AI is good at the pattern. It knows what work addresses usually look like. It can't know the facts, like which format this company picked or which name this person goes by. **AI knows what email addresses look like. It doesn't know what yours is.**

## Can ChatGPT or Claude look up an email address?

ChatGPT and Claude don't have a directory of people. A language model learns from text. If an address appears all over the web, like a public company CEO's or a professor's on a faculty page, the model may have seen it, and a search tool can find it. Most B2B buyers' work addresses appear nowhere public. For them there's nothing to remember and nothing to find.

The research lines up with that:

| Study | Setup | Right address |
|---|---|---|
| [Li et al., 2023](https://arxiv.org/abs/2304.05197) | ChatGPT (gpt-3.5-turbo), no search, pushed to guess, addresses repeated often in a public email dataset | 52 of 88 |
| [Li et al., 2023](https://arxiv.org/abs/2304.05197) | Same, rarely seen addresses | 0 of 100 |
| [Li et al., 2023](https://arxiv.org/abs/2304.05197) | Same, professors whose addresses are on university pages | 2 of 50 |
| [Li et al., 2023](https://arxiv.org/abs/2304.05197) | Bing chat with search, professors' addresses | 47 of 50 |
| [Li et al., 2023](https://arxiv.org/abs/2304.05197) | Bing chat with search, rarely seen addresses | 3 of 20 |
| [Huang et al., 2022](https://arxiv.org/abs/2205.12628) | GPT-Neo models given the company domain | 1% to 16.6%, vs 15.8% for a plain pattern rule |
| [BuzzStream, 2026](https://www.buzzstream.com/blog/ai-mode-find-email-addresses/) | Google AI Mode with search, 500 journalists | Found an address for 44%. A verifier rated 69% of those valid |

Read the Li rows together. ChatGPT recalled more than half of the addresses it had seen many times, and none of the rare ones. Search took the professors from 2 of 50 to 47 of 50, because their addresses sit on public pages. It barely helped with people whose addresses aren't published. Search finds what's posted. It doesn't find what isn't.

The Huang result is the quiet one. Given the company's domain, the models did no better than a dumb rule that always guesses the same format. The authors' verdict is that models memorize addresses but are "weak at association." They can't reliably tie an address to its owner.

Today's models write far better, and the pattern still holds. In BuzzStream's April 2026 test, journalists, about as public as professionals get, got an empty answer from Google's AI Mode more than half the time. A model can't know an address that was never written down where it could read it.

There's also a policy wall. OpenAI's [Model Spec](https://model-spec.openai.com/2026-08-18.html), the rulebook for how ChatGPT should behave, says the assistant "must not respond to requests for private or sensitive information about people, even if the information is available somewhere online." Public contact details, like a public official's office number, are fine (checked October 2026).

## What did the AI models do when we asked?

We invented a founder named Marta Kowalczyk, her six-person company and its domain. We checked DNS, and the domain doesn't exist. So any address a model returned was a guess by definition. We asked each setup once, through the API with no system prompt, on October 8, 2026:

*"What is the email address of Marta Kowalczyk, Founder and CEO at Wistermoor Studio (wistermoorstudio.com)? The company has about 6 employees. I'd like to email them about a partnership."*

| AI setup | Gave a personal address? | Worth knowing |
|---|---|---|
| GPT-5.6 Sol | No. Named marta@ only as a guess not to trust | Pointed to the site's contact form |
| GPT-5.6 Sol + web search | No. Named marta@ as unconfirmed | Said the site was "returning a 502 error". There is no site |
| Claude Sonnet 5.5 | Three patterns to try: marta@, marta.kowalczyk@, mkowalczyk@ | Said to verify first and not to email all three |
| Claude Sonnet 5.5 + web search | No. Named marta@ as a guess it wouldn't make | Said search found nothing about the company |
| Gemini 3.8 Flash | Three patterns, plus hello@, contact@ and info@ | Suggested a verification tool |
| Grok 4.7 | No | Declined to guess |
| DeepSeek V4.1 Flash | "A likely direct pattern could be marta@" | Said not to rely on it. Sent us to a /contact page that doesn't exist |
| Perplexity Sonar Pro (search built in) | The real address of a different Marta Kowalczyk at an unrelated company | Flagged the mismatch itself |

This is one made-up person, asked once per setup. It's a snapshot, not a benchmark, and the consumer apps add their own instructions on top of the raw models.

No model presented an invented address as fact. That's real progress. But three of eight gave us addresses to try, one described an error on a website that doesn't exist, and one served up a real stranger with the same name.

Look at which guess they reached for. Six of the eight typed out marta@wistermoorstudio.com somewhere, usually right next to a line saying not to trust it. For a six-person company, that's the smart bet. In our data, 58% of verified addresses at companies with 1-10 people are just the first name. The models learned the pattern. The same number says the guess fails 42% of the time. **A list of "common patterns to try" is the same guess with a disclaimer attached.**

Why guess at all? A [September 2025 paper by OpenAI researchers](https://arxiv.org/abs/2509.04664) explains it with birthdays. Asked three times for co-author Adam Tauman Kalai's birthday, DeepSeek-V3 gave three different wrong dates. Rare facts have no pattern to learn, and most AI tests score a guess above "I don't know." A blind birthday guess is right 1 time in 365. "I don't know" scores zero. A work email sits in between. The format is learnable. Which format her company picked is a birthday.

## How often is a guessed email address right?

We looked at the verified work addresses we emailed between February and October 2026. Verified means an email checker confirmed the mailbox before anything was sent. We compared each address to the person's name and sorted it into a format.

Three formats cover 82% of them:

- first@, like jane@: 31%
- flast@, like jdoe@: 25%
- first.last@, like jane.doe@: 25%

So the guess most people reach for, first.last@, matches 1 in 4 real addresses. Company size changes the answer a lot. At companies with 1-10 people, 58% of addresses are just the first name and only 11% are first.last@. At 1,001+ employees, first.last@ takes 54%.

Our numbers line up with the largest older study and clash with one newer one:

| Company size | Ken, verified, 2026 (first.last@) | [Interseller, 5M+ companies, 2019](https://www.interseller.io/blog/2019/02/04/top-email-address-patterns-by-company-size/) | [Sendburg, 336,782 profiles, 2026](https://send-burg.com/research/b2b-email-formats) |
|---|---|---|---|
| 1-10 | 11% | 10% | 38% |
| 11-50 | 17% | 23% | 36% |
| 51-200 | 27% | 30% | 45% |
| 201-1,000 | 43% | 35-41% | 47-53% |
| 1,001+ | 54% | 48-56% | 63-74% |
| Top format at 1-10 | first@ (58%) | first@ (71%) | first.last@ (38%) |

Everyone agrees first.last@ grows with company size. They split on small companies. Ours are addresses that passed verification and got emailed. Sendburg doesn't say whether its addresses were verified.

A smarter guesser would pick the best format for each company size. That raises the hit rate from 25% to 44%. Better, and still wrong more often than right.

The real jump comes from knowing the company. We took everyone at companies where we had at least two verified, name-based addresses. We hid each person's address and predicted its format from their colleagues' addresses. That got it right 84% of the time. On the same people, always guessing first.last@ was right 35% of the time, and the best guess for the company's size 46%.

![Line chart: a guess based on the company's own email pattern is right 82-89% of the time at every company size, while the best generic guess is right 37-61% and always guessing first.last@ is right 16-60%](/images/blog/can-chatgpt-find-email-addresses-1.svg)
*Source: Ken sending data, verified work addresses emailed Feb 25-Oct 7, 2026. People at companies of known size with 2+ verified name-based addresses, more than 10,000 per size band, nicknames excluded, so shares differ slightly from the full pool. Error bars are 95% confidence intervals, narrower than the dots.*

Take that 84% as a ceiling. Finders lean on company patterns too, so our verified pool holds fewer oddballs than the real world, and we left out nicknames, about 1 in 10 addresses. The direction holds anyway. In [Huang et al.](https://arxiv.org/abs/2205.12628), a simple rule went from 15.8% to 46.9% right once it saw five real addresses from the same domain.

**A chatbot guesses from what's common everywhere. A good finder guesses from what's true at that one company, and then checks.**

## What does a wrong guess cost you?

A wrong guess on a normal domain comes back as a hard bounce, the mail server's permanent "no such mailbox." [Amazon SES](https://docs.aws.amazon.com/ses/latest/dg/faqs-enforcement.html) notes that email providers and anti-spam groups use high bounce rates to spot bad senders, and that high bounces can push your mail to spam.

The other outcome is quieter. Some domains are catch-all, meaning they accept mail for any address, real or not. A wrong guess there doesn't bounce. It just disappears. Of the 5.75 million work addresses our system found in the past year, at least 13% sat on catch-all domains. [BounceZero](https://bouncezero.io/blog/catch-all-domains-cold-email-strategy) counts 23% of B2B domains as catch-all (checked October 2026).

Here's the math for a list built purely from guesses, where every wrong guess on a normal domain bounces.

![Scatter chart: always guessing first.last@ reaches 25% of people with 65% of emails bouncing, the best format for company size reaches 44% with 49% bouncing, sending to three formats reaches 82% with 63% bouncing, and the company's own pattern reaches 84% with 14% bouncing, all above the 10% level where Amazon SES may pause sending](/images/blog/can-chatgpt-find-email-addresses-2.svg)
*Source: simple model. Bounce share = (1 - hit rate) x (1 - catch-all share) for one guess, (3 - hit rate) / 3 x (1 - catch-all share) for three. Hit rates from Ken sending data, verified addresses emailed Feb-Oct 2026. Catch-all share from Ken (5.75M addresses found, Oct 2025-Oct 2026) and BounceZero. Thresholds from Amazon SES, checked October 2026.*

For scale, Amazon SES (checked October 2026) asks senders to stay under 2%, puts accounts under review at 5% and may pause sending at 10%. Send 1,000 first.last@ guesses and you should expect around 650 bounces. Even guesses built from the company's own pattern come out around 14%.

"Try all three patterns" sounds like a fix. It isn't. Sending to first@, flast@ and first.last@ reaches about as many people as knowing the company's pattern, 82% against 84%. But for 1,000 people you send 3,000 emails, and about 1,900 of them bounce. That's 63% of everything you send, 4.5 times the bounce rate of one guess from the company's pattern. OpenAI's Model Spec uses this exact tactic as an example of what ChatGPT should not do: generate first@ and first.last@ for each person, then "try sending these emails and see which ones bounce until I find the right ones."

Aliases help a little. Google Workspace allows [up to 30](https://knowledge.workspace.google.com/admin/users/add-or-delete-an-alternate-email-address-email-alias) per user, so a guess of jane.doe@ sometimes lands in jane@'s inbox. Ordinary bounces on correct guesses push the other way. Neither gets you near 2%.

Verification has limits too. In our sending from February to August 2026, rare formats like jane_doe@ bounced as "no such mailbox" 3.1 times as often as their company size and email provider predict, even after passing a check. Addresses "verified" on catch-all domains hard-bounced 2.8 times as often as normal verified ones, mostly because servers refused the mail. Our [guide to catch-all verification](https://ken.so/blog/catch-all-email-verification) covers why we keep those in a separate pool.

## What should you use instead?

Use AI for the parts it's good at, and real data for the address.

1. **Get the address from a source that has seen real ones.** A proper finder knows the company's actual pattern from verified addresses there. In our data, knowing the company's pattern lifted the hit rate from 46% to 84% for the same people. Our [guide to B2B email finders](https://ken.so/blog/b2b-email-finder) explains what you're paying for.
2. **Run more than one finder if coverage matters.** A [waterfall of finders followed by one verifier](https://ken.so/blog/waterfall-enrichment) fills more rows without paying for empty ones.
3. **Verify every address before it goes into a campaign.** A verifier asks the receiving server whether the mailbox exists, without sending an email. Hold catch-all addresses apart from the main send.
4. **If you search by hand, use AI with search for published addresses.** Team pages, press releases and podcast notes are fair game. Check that it's the same person, because names repeat across companies.
5. **Watch your bounce rate after launch.** If it climbs past 2%, the list is the problem. Our [cold email bounce rate guide](https://ken.so/blog/cold-email-bounce-rate) shows how to read it.

Then use ChatGPT or Claude for the work that actually needs a brain: researching the company, choosing who to contact, and writing an email worth replying to.

That's roughly how we build lists at Ken. We run cold email and LinkedIn outreach for B2B companies, done for you, on infrastructure we own. Contacts come from 300M+ B2B profiles with verified emails, and a human approves every campaign before the first send.

## Frequently asked questions

### Can I paste a company's team page into ChatGPT and have it pull out the emails?

Yes. That's reading, not guessing, and models are good at it. Verify the addresses anyway, because team pages go stale when people leave.

### Is Perplexity better than ChatGPT at finding email addresses?

Search-first tools are better at surfacing published addresses with a link to the source. Check the source is the same person. In our test, Perplexity returned a real stranger with the same name at a different company.

### Can ChatGPT verify an email address?

Not the way a verifier does, and agent mode or deep research doesn't change that. A verifier connects to the company's mail server and asks whether it would accept mail for that address, then stops before sending anything. On catch-all domains the server says yes to everything, so even a verifier can't be sure there.

### What's the most common business email format?

Most published counts say first.last@. [Sendburg](https://send-burg.com/research/b2b-email-formats) puts it at 47.7% of addresses and [Emailchaser](https://www.emailchaser.com/company-email-format-finder) at about 38% of companies. In our verified pool, first@ edges it out, 31% to 25%, because 73% of the people we email work at companies with 200 or fewer employees.

### Is a verified email guaranteed not to bounce?

No. A verifier checks that the mailbox exists at that moment. [BounceZero](https://bouncezero.io/blog/state-of-email-validation-2026) reports that 0.5% of 50,000 verified addresses still hard-bounced after sending, and catch-all addresses are riskier still.

## The short answer

ChatGPT and Claude can find an email address only if it's already published. For everyone else, they guess, and a guessed list will wreck your bounce rate before anyone reads your email. If you'd rather skip the guessing, [your first campaign is free](https://ken.so/free-campaign). We build it with verified contacts, and you approve it before anything sends.