Researching local businesses with your AI agent, with sources
How to have your AI agent research a local business before a pitch, the first job of AI agents for client work: what to ask, sources and dates, what to check.
A solo founder building a one-person AI agency wants to pitch the bakery two streets away. She asks her AI agent to tell her about it and gets back a tidy paragraph: family run since 2009, known for sourdough, a gluten-free range, wedding cakes to order. It reads well. On the call she mentions the wedding cakes. The owner looks puzzled. They stopped doing those three years ago, and the family sold the shop last spring.
Nothing in that paragraph said where each fact came from, so she had no way to tell the fresh ones from the stale ones.
Your agent can do the legwork of researching a local business, and it can do it well. The trick is to ask for research you can check: every fact with its source and the date it was seen, a plain "not found" where it found nothing, and no guesses dressed up as facts. Then you check the few facts your pitch will lean on.
Can an AI agent research a local business accurately?
It can read a business's website, its map listing and its reviews faster than you can, and summarise them in plain words. When Anthropic added web search to Claude in March 2025, it said Claude gives direct citations when it uses information from the web, so you can check the sources yourself. That's the part that matters for you.
The risk is the gaps. OpenAI's own research note from September 2025 describes hallucinations as plausible but false statements, and explains that models are often rewarded for guessing rather than saying they don't know. A local bakery is exactly the sort of small, rarely written-about subject where a model may fill a gap with something that sounds right.
Anthropic's guide to reducing hallucinations suggests three habits that suit this work well: give the model permission to say it doesn't know, ask it to back each claim with a quote or source, and have it drop any claim it can't support. The same guide is clear that these habits reduce errors without removing them. So the answer is yes, as long as you treat the agent's research as a first draft with its working shown.
What should you ask your agent to find?
Keep the brief small. The same three questions that make a good research a business before pitching routine work for an agent too:
| Question | Where the agent should look | What counts as a source |
|---|---|---|
| What do they do, and for whom? | The business's own website, menu or price list | A link to the page, with the date seen |
| How do customers find them today? | Their Google or Apple Maps listing, their social pages | The listing link, and the hours and phone number as shown |
| What do customers say? | Recent reviews on their listing | Two or three short quotes, each with its review date |
Leave out the things an agent is most tempted to invent: how much the business earns, how many staff it has, what the owner is planning, and any score for how likely they are to buy. None of it has a source you could show the owner, and none of it helps the first conversation.
Name the business precisely when you ask. "The bakery on the high street" invites a mix-up; the business name, its suburb and its website address don't. If the business has no website, give the agent the map listing instead, and say so. That one line at the start of the request prevents a common mistake in agent research: facts from a similarly named business down the road ending up in your notes.
How do you make the agent show its sources?
Say it in the request, in plain words. Something like this works:
- For each fact, give the link you found it on and today's date.
- If you can't find something, write "not found". That is a good answer.
- Keep facts and guesses apart. Put anything you're unsure about under "questions for the owner".
- Don't score or rank the business.
What comes back should read like notes, not a sales page. For the bakery, a good result might say: "Sourdough and pastries, no wedding cakes listed (website menu page, seen 1 June)", "Opens 7 am weekdays on the map listing, 7.30 am on the website (both seen 1 June)", and "Owner's name: not found".
That second line is gold. A mismatch between the website and the map listing is a real, fixable problem you can raise without overselling, and you found it because the agent showed both sources side by side instead of picking one.
What should you check yourself?
You don't need to redo the research. Check the facts you'll say out loud, and the ones a mistake would embarrass you on.
- Open every link you plan to mention. Make sure the page says what the agent says it says, and that it's the right business. Two cafés with similar names in the same suburb are more common than you'd think.
- Compare hours, phone and address with the map listing. These change most often and matter most to customers.
- Read the review quotes in place. A quote out of context can flip its meaning.
- Look at the dates. A menu page last changed two years ago may not reflect the shop today.
- Check it's still trading under the same name. A recent review or a recent post is enough. Businesses change hands, and old pages stay up long after.
Anything you won't mention can stay as the agent found it, marked unchecked. When a fact does turn out wrong, fix it in your notes and write down how you know.
How do you research a short list without losing track?
Most people research five or ten businesses at once, not one. That's where agent research either saves hours or makes a mess, depending on where the results land.
Give each business its own folder or record, and keep the agent's notes, the sources and your corrections in it. If you've set up how to organise client files for an AI agent, prospects fit the same pattern: one folder per business, named the same way every time. If your notes live in an app, what MCP means is worth five minutes: it's the standard way for an agent to read and update the notes you already keep, instead of working from a copy you pasted into a chat.
Work through the list in the same order each time. Research all of them, check the facts you'll use, then decide who to contact. Deciding while you research tends to turn the notes into a sales pitch to yourself.
What happens after the research?
Research is only useful if it changes the first conversation. Pair each checked fact with one open question: the hours mismatch becomes "Which opening time is right? I wasn't sure which to trust." The missing owner's name becomes a question you ask, not a gap you fill.
If you use the agent to write the first message too, the same rule applies: it drafts, you read and send. Letting AI draft client emails that you send yourself works well when the draft only uses facts you've checked. It's also worth deciding in advance to approve what your agent sends before anything reaches a business owner, so a stale fact never goes out under your name.
Good research doesn't mean the bakery will say yes. It means the call is about their shop, with facts they can recognise, and nothing you'll have to walk back.
Where Volant fits
Volant is built for researching local businesses to pitch. Your agent researches a business and shows its sources. You correct the facts before anything uses them. No made-up scores. Whatever you use, the habit is the same: ask for sources, accept "not found", and check what you'll say out loud.