GEO and AI search

How to Earn Brand Mentions in AI Chatbot Answers for Small Business Clients

Chatbots name the small businesses that many independent sources describe the same way. Here's the step-by-step work to get a client named, starting with an honest baseline.

person using black laptop computer
Photo by Firmbee.com

A chatbot names a small business when lots of independent sources describe it the same way. You can see this pattern for yourself. Run one question through five tools and read what they cite. So if you want to know how to get mentioned in AI chatbot answers for a plumber in Harrogate or an accountant in Leamington Spa, the job is corroboration, not optimisation. You get the facts straight. You state them plainly on the client's site. Then you get trusted sites the client doesn't control to say the same thing.

Nobody outside these companies knows exactly how a given model picks the brands it names, and none of the tools publish a formula for it. What you can watch is how they behave. Tools that browse lean on the pages they find. Tools that answer from training data repeat what was widely written before their cutoff. Both favour a business that's described consistently in many places. Neither favours a website that repeats its own keywords.

This article gives you six steps, a re-test schedule and a way to sell the work as a retainer. Each step tells you what to do, what you should see and the mistake that costs people time. The tools named are the versions available in the last week of March 2024.

By the end you will have a tested baseline, a matched set of facts and a 12-week plan to get a client named

In twelve weeks you'll hold five things you can show a client. First, a prompt audit spreadsheet that records where they appear today. Second, one canonical fact sheet. Third, a rewritten entity page. Fourth, two fair comparison pieces. Fifth, a list of local citations, each one either live or chased. Every one is a file or a URL, which makes the work easy to report and hard to argue with.

Get these in place before you start:

  • Written sign-off from the owner on the business facts: name, address, phone, services, hours and founding year
  • Logins for the website CMS, Google Business Profile, Bing Places for Business and Apple Business Connect
  • Accounts for ChatGPT (free and Plus), Gemini, Microsoft Copilot, Perplexity and Claude
  • One shared spreadsheet for the audit, with a tab for each re-test
  • Access to the client's customer email list, for the review step later
  • About three hours a week for twelve weeks

Step 1: Run a baseline prompt audit before you change anything

Write 15 to 20 prompts a real customer would type. Skip the jargon. Write "best emergency plumber in Harrogate", "who does wedding flowers near Clifton on a budget" and "accountant in Leamington Spa for sole traders doing self assessment". Mix urgent prompts, price-led prompts and comparison prompts. Then add two brand prompts, such as "what do you know about [client name]", so you can check accuracy as well as presence.

Run every prompt in each tool. Free ChatGPT runs GPT-3.5 with no browsing, so it answers from training data. ChatGPT Plus runs GPT-4 and can browse. Gemini is Google's chatbot, renamed from Bard in February. Microsoft Copilot and Perplexity both search the web and show sources. Claude 3 came out in early March and answers from its training data without searching. Google SGE is still a Search Labs experiment, so log it if you have access, but don't build the client report around it.

Answers change from run to run, and that's the honest answer to how you measure this. Use a fresh chat every time, with memory and history off, and keep the wording identical. Run each prompt three times per tool and record how many of the three name the client. "Two of three in Perplexity" is a figure you can compare in week 12. A single screenshot isn't.

Log these columns for each run: prompt, tool, date, mentioned yes or no, position in the answer, whether the description was accurate, which competitors appeared (by category, such as "national boiler chain", kept for the client's eyes only) and which sources were cited.

What you should see: the client missing from most answers. Often they appear in Perplexity or Copilot and nowhere in free ChatGPT. That's expected. Browsing tools read the live web, so a recent directory page or review gets them in. Free ChatGPT only knows what was written before its training cutoff, and a small local firm rarely appears there often enough to be repeated.

The common mistake: fixing things before you've recorded a baseline. If the About page changes in week one, you can't show what the work did, and the retainer conversation in month three gets much harder.

Read the cited sources, not just the answer

Perplexity and Copilot list the pages they drew on. Open every one. You'll see the same few types come up again and again, usually a local news story, a trade association member list, a review platform page and the odd "best of" roundup from a regional site. That list is the most useful thing the audit produces. It shows which sources the browsing tools already trust for this town and this trade, and it becomes your target list for citation outreach in Step 5. If a competitor appears because a chamber of commerce directory lists them, your client's next job is getting into that same directory.

Step 2: Write one canonical fact sheet and make every listing match it

Models trust facts that agree. If the client's Google listing says one address, the Apple listing says another and a trade directory still shows the name they traded under in 2019, you've given the tools three businesses to choose between. So write one document and treat it as the only source of truth:

  • Legal name and trading name
  • Full address, or service area only if they work from home
  • Main phone number
  • Service area, listed by town
  • Primary category, one only
  • Services, in the words customers use
  • Opening hours, including emergency cover
  • Year founded
  • A one-sentence description, such as "Family-run plumbing and heating firm in Harrogate since 2009, fixing boilers and leaks across Harrogate, Knaresborough and Ripon"

Then correct every listing to match it. Start with Google Business Profile, then Bing Places (which can import straight from the Google profile), then Apple Business Connect, which feeds Apple Maps and Siri. After those, move on to the social profiles, trade directories and any review platforms the audit turned up.

What you should see: a short list of conflicts, not a long one. It's usually an old address from before a move, a former trading name or a wrong primary category, such as a florist listed as "gift shop".

The common mistake: spending an afternoon changing "St" to "Street" across forty listings while a well-ranked directory still shows the old address. Fix the conflicts that change meaning first. Formatting is last.

Step 3: Build an entity page that says plainly what the business is

Rewrite the About page so that the first paragraph answers who, what, where, since when and for whom in plain sentences. Write it so a model can lift it word for word. For the Harrogate plumber, that's "Harrogate Heating and Plumbing is a family-run plumbing and heating firm based in Starbeck, Harrogate, founded in 2009. It repairs boilers, leaks and blocked drains for homeowners and landlords across Harrogate, Knaresborough and Ripon, with a 24-hour emergency line." The story, the team photos and the values go further down the page.

Then add structured data. Schema markup is a short block of code in the page that restates the same facts in a format machines read without guessing. Use LocalBusiness, or a more specific type such as Plumber or Florist, for a firm with a location. Use Organization for one that doesn't serve customers locally. The sameAs field lists the URLs of every profile you fixed in Step 2, which tells search engines that all of those profiles are the same business.

{
  "@context": "https://schema.org",
  "@type": "Plumber",
  "name": "Harrogate Heating and Plumbing",
  "foundingDate": "2009",
  "telephone": "+44 1423 000000",
  "areaServed": ["Harrogate", "Knaresborough", "Ripon"],
  "sameAs": [
    "https://www.example-trade-association.org.uk/members/harrogate-heating",
    "https://www.example-chamber.co.uk/directory/harrogate-heating"
  ]
}

Is schema worth adding for chatbots? Yes, but be straight about what it does. It helps search engines link the business to its profiles, and the browsing tools draw on search results. Nothing shows that schema on its own makes a chatbot name anyone. It's an hour of work that makes everything else easier to connect. It doesn't replace the other steps. If you want the wider picture, we've covered how generative engine optimisation differs from traditional SEO separately.

What you should see: a page a stranger could sum up in one line after reading the first paragraph. Test it. Paste the paragraph into Claude and ask for a one-sentence summary. If the summary leaves out the town or the trade, rewrite the paragraph.

The common mistake: brand-voice copy. "We craft experiences that bring people together" sounds fine on a mood board and tells a model nothing. It could be a florist, a caterer or a wedding venue in any country.

Repeating keywords proves nothing, because a site agreeing with itself is not corroboration

Ten mentions of "best plumber in Leeds" on the client's own page are one source saying one thing ten times. A single line in the local paper saying the firm fixed a school's heating over half-term is a second, independent source. If the browsing tools behave the way the audit suggests, the newspaper line is worth more than the whole page of repeats.

Keyword density was a weak signal even in classic search. It's weaker now. Google's March 2024 core update began rolling out on 5 March, together with new spam policies, and one of them covers scaled content abuse. That means pages made in bulk mainly to rank, whoever or whatever wrote them. The pattern it targets is familiar to anyone who's run local SEO, such as forty near-identical service pages with only the town name swapped, or an FAQ that asks "Do you cover Leeds?" in twelve different ways. We've set out what's changing in our piece on Google's March 2024 core update and new spam policies.

So does an FAQ stuffed with the town name help AI mention the client? No. In the best case it adds nothing a model didn't already have from the entity page. In the worst case it's the kind of thin, repetitive page the update is demoting this month. Since the browsing chatbots work from search results, a page that drops in search drops out of what they can read too.

This is why the rest of the plan points outward. The client's site states the facts once, clearly. Comparison content answers the questions customers really ask. Then the slow work begins, getting chambers, trade bodies, local press and real customers to describe the business in their own words, in places the client doesn't own.

Step 4: Publish fair comparison content that answers the questions customers actually ask

Customers ask chatbots to compare things all the time, and the tool needs a page to draw the answer from. Write two pieces that compare options, not named competitors. For the Harrogate plumber, that's "Repair or replace your boiler, and how to tell which you need". For a florist in Clifton, it's "Flower subscription or one-off bouquets". For the Leamington Spa accountant, try a page called "Who we're not the right accountant for".

The audit prompts are your brief. Before you write a word, tag each prompt by intent as urgent, price-led, comparison or brand check. Our guide to keyword intent mapping walks through the method. The comparison and price-led prompts become these two pages. Urgent prompts belong on the service pages and the Google profile, where someone with a leaking pipe lands.

Be fair, even when it costs the client a sale. A plumber who writes "if your boiler is under eight years old and the fault is a single part, repair it" sounds like someone giving advice. That's the kind of sentence a model quotes. Say who the service doesn't suit, too. The accountant might write that they don't take on limited companies running payroll, and that a sole trader with one income stream can often file their own self assessment.

Compare categories, and leave rivals unnamed

Should you write pages that name competitors? No. You can't keep someone else's prices and services accurate, the page reads as an attack, and you're handing a model the rival's name to repeat next to your client's. Compare types of option instead, such as a national boiler chain, an independent engineer and a DIY fix. The reader gets the same help, and you stay on safe ground.

What you should see: two pages, each answering one audit prompt in its first two paragraphs, with the fact-sheet description in the footer or author line. The common mistake is a sales page in disguise, where every option ends with "so call us". If the other option never wins, it isn't a comparison, and readers can tell.

Step 5: Earn citations from local sources that models already trust

Your target list already exists. It's the column of cited sources from the Step 1 audit. Those are the pages the browsing tools picked for this town and this trade, so getting the client onto them is the most direct route there is. After those, work through the local sources a resident would trust:

  • Chamber of commerce member directories
  • Trade associations and registration bodies for the client's trade
  • Local press, through a story with a real hook such as a community job, an apprentice or an anniversary
  • Sponsorship pages for local events, sports clubs and school fairs
  • Supplier "where to buy" or "approved installer" pages
  • Council, parish and community group listings

Which third-party sites matter most for a local small business? Start with the ones the audit shows being cited, because you've seen them in use. After that, put a source a Harrogate homeowner would trust ahead of one with a big audience and no local link. Chambers pay off twice, because membership brings referrals as well as a listing. We've covered how agencies win work through local business groups and chambers separately.

Aim for 8 to 12 real mentions in 12 weeks. Send every source the fact-sheet description and ask them to use it word for word. When ten independent pages say "family-run plumbing and heating firm in Harrogate since 2009", you've got corroboration. Ten different descriptions are just noise.

What you should see: slow replies and roughly one ask in three turning into a live page. Log each one as asked, agreed, live or indexed. The common mistake is buying a bulk directory package. Hundreds of low-quality listings created at once bring in new spellings and old addresses, and nobody can log in to fix them later. That undoes Step 2.

Step 6: Use the client's email list to turn customers into specific, independent reviews

A customer writing "fixed our combi boiler in Knaresborough the same afternoon" is independent, specific and local, all in one sentence. Reviews like that are the corroboration a client can't write for themselves. Getting them is mostly a matter of asking the right people the right way.

Send a short email to customers from the last six months. Ask them to describe the job in their own words, covering what was done and where. Don't give them a script and don't offer anything in return. Google's review policy prohibits incentivised reviews, and scripted reviews come out identical, which defeats the point of having many independent voices.

In SendTo, build a segment of customers who clicked a past service email, since they're the ones still paying attention. Write a two-line brief, such as "Thank recent boiler customers and ask for an honest Google review describing the job", and the AI campaign designer turns it into a finished HTML email in seconds. Before it goes, the AI reviewer checks the subject line, structure, wording and links, and runs basic compliance checks. SendTo verifies every address before sending, so the request goes to real inboxes and bounces don't drag down the client's sending reputation. Be clear about what SendTo doesn't do here. It sends the request. It doesn't collect, filter or publish reviews.

What you should see: a handful of reviews over two or three weeks, several naming the town and the job. The common mistake is mailing the whole list, including people who bought in 2017. They won't remember the job, and an old, unengaged list hurts delivery for every email after it.

Check it worked: rerun the same audit at weeks 4, 8 and 12

Use the same prompts, the same tools, fresh sessions with memory off, three runs each and a new tab in the same spreadsheet. Change any of those and the comparison stops meaning anything. This is how you measure AI mentions when the answers change every time. You count across repeated runs instead of trusting one screenshot.

Success shows up in a set order. Description accuracy improves first, so brand prompts stop getting the old address or the former trading name. Mentions then start in the browsing tools, which are Perplexity, Copilot, ChatGPT Plus with browsing and Gemini. Free ChatGPT and Claude 3 answer from training data, so this quarter's work won't reach them until a newer model is trained on a newer crawl. That can take many months, and nobody outside those companies knows when it'll happen.

Watch the brand prompts for invented details. Ask a model for specifics it doesn't have, such as the founder's name or a call-out price, and it often fills the gap with a confident guess. Log every error. Each one shows you a fact that needs stating more plainly in more places.

If nothing has moved by week 8, check these in order:

  1. Listings have drifted back. Google accepts suggested edits from the public, so recheck the profile against the fact sheet.
  2. Citations agreed aren't citations live. Open each URL and search for the page to confirm it's indexed.
  3. Comparison pages can't be crawled. Look for a stray noindex tag, a robots.txt block or a page with no internal links pointing to it.
  4. Sources use different wording. If each one describes the client differently, send the fact-sheet sentence again.

Sell it as a quarterly retainer with honest reporting, not a guaranteed ranking

You can influence what ChatGPT or Gemini says about a business. You can't control it, and you shouldn't sell the work as though you can. What you can sell is a fixed set of jobs done every quarter, with a measured before and after. A retainer quarter looks like this:

  • Baseline prompt audit and quarterly re-tests
  • Fact-sheet clean-up across every listing
  • Entity page rewrite and schema
  • Two fair comparison pieces each quarter
  • Citation outreach toward 8 to 12 live mentions
  • Review request campaigns sent from the agency's own branded email platform

How to get mentioned in AI chatbot answers without promising it

Price the hours and the outputs, never the outcome. The quarterly report is the audit grid itself, showing prompts down the side, tools across the top and runs mentioning the client in each cell, with week 0 next to week 12. Add the live citation list and the new reviews. A Leamington Spa accountant can read "Perplexity named you in 2 of 3 runs for sole-trader prompts, up from 0" and decide for themselves whether the work paid off.

The review emails are where an agency platform earns its place. When you run them on SendTo under your own brand, the client sees your product doing the work, and it becomes one more line in the retainer instead of a tool they have to buy elsewhere. The partner fee is listed on our pricing page.

Start this week with the audit. Write the 20 prompts, run them before you change anything, and you'll have the baseline every later report depends on.

Write about this work for agencies? SendTo's affiliate programme pays you 20% of the partner subscription each agency you refer pays us, every month they pay it. It's never a share of what their own clients pay them. Attribution lasts 90 days from the first click, the first touch wins, and we pay commission monthly by PayPal.

Join the SendTo affiliate programme

Frequently asked questions

Which chatbots should you test a client in?

Test five: ChatGPT (free and Plus), Gemini, Microsoft Copilot, Perplexity and Claude. Log Google SGE if you have access, but don't build the report around it. The tool versions here date from late March 2024, so check what's current before you run your audit.

How does a chatbot decide which business to name?

Nobody outside these companies knows the exact formula. Tools that browse lean on the pages they find. Tools that don't browse repeat what was widely written before their training cutoff. Both favour a business that lots of independent sources describe the same way.

Why does my client show up in Perplexity but not in free ChatGPT?

Perplexity reads the live web, so a recent directory page or review is enough to get your client in. Free ChatGPT answers from training data, and a small local firm rarely appeared often enough before the cutoff to be repeated. That gap is normal in a baseline.

How do I measure AI mentions when the answers change every time?

Use a fresh chat with memory off, keep the wording identical and run each prompt three times per tool. Then record how many of the three runs name the client. Two of three in Perplexity is a figure you can compare in week 12. One screenshot isn't.

Does stuffing the town name into pages and FAQs help AI mention us?

No. Ten repeats of a keyword on your client's site are one source saying one thing ten times. A single line in the local paper is independent corroboration. Google's March 2024 spam policies also target pages made in bulk mainly to rank.

Is schema markup worth adding for AI chatbots?

Yes, as an hour's work, but it won't fix anything by itself. LocalBusiness markup with a sameAs list helps search engines tie the business to its profiles, and browsing tools draw on search results. Nothing shows that schema alone gets a chatbot to name anyone.

Should comparison pages name competitors?

No. Compare categories, like a family-run local firm against a national boiler chain, and answer the questions customers actually ask. That covers comparison prompts without you making claims about named rivals.

How do I price and report this without promising results?

Sell it as a quarterly retainer and report the audit honestly: the baseline, then reruns at weeks 4, 8 and 12. Show the files and URLs you delivered. Never guarantee a ranking, because nobody controls what a model says.

Everything this blog teaches, on a platform with your name on it.

Email campaigns, websites, landing pages and print for your clients, made from a short brief and sent under your brand. Create your account and look around with a sample workspace loaded.

No contract · Cancel any time · Explore everything before you pay