How to get your business recommended by AI assistants

Before changing anything, find out what the assistants say now. A baseline turns a vague worry into three concrete lists: the prompts where you are named, the competitors named where you are not, and the sources the assistants cite. The third list is the most useful thing this stage produces, because it tells you where stage five should spend its time.

Write prompts the way customers actually ask

People do not type “plumber Leeds” into an assistant. They describe the situation: the boiler has stopped, it is Sunday, they rent, they are in Headingley, and they want someone today. Your prompts need to look like that, because the details in a prompt change the searches an assistant runs and the sources it reads. Write four kinds, and keep the branded ones apart when you report, because a prompt that names you mostly tests whether the assistant can describe you, not whether it recommends you. The measurement guide covers building a larger prompt set once this routine is running.

The best raw material is your own enquiries. Read twenty recent emails, form submissions or call notes and copy the phrases customers use: the problem, the urgency, the type of property, the area, the budget. Reviews are the second source, because they describe the job in the customer's words. Long queries in Search Console are the third. Your own guesses about what people ask come last.

An assistant is good at turning a pile of enquiries into clean test prompts, as long as you strip out names, addresses, phone numbers and email addresses before you paste anything in. Use the prompt below, then edit what comes back by hand.

Test each assistant cleanly

Assistants personalise, so the conditions you test under matter as much as the prompts. OpenAI says ChatGPT works out an approximate location from your IP address, may use saved memories when it rewrites your prompt into searches, and can be used for web search without signing in.² Its own advice for local results that show the wrong area is to include the city, neighbourhood or postcode in the question.² AI Overviews and AI Mode may use different models and techniques, so the responses and links they show will vary.³ Microsoft documents that Copilot, as used with work accounts, turns a prompt into a short query of a few words, sends it to the Bing search service, and shows the exact queries it sent in its citations.⁴ So write the town into every unbranded prompt, test signed out or in a fresh account, and report each assistant separately.

Record the cited sources without retyping them

Copying every link out of every answer by hand is slow and error-prone. In ChatGPT, citations sit inline and under a Sources control when one is shown; OpenAI says selecting Sources shows the cited sources and other relevant links.² Perplexity and Copilot list their sources with each answer as well. Copy the whole answer, sources included, into a separate assistant with the prompt below, and paste the table it returns into your log. Use a different assistant from the one you are testing, so the extraction does not leak into the test.

Read the baseline as a pattern

One answer is an anecdote. The same prompt can name different businesses on two runs a minute apart, so read across runs and across prompts. Three patterns matter. If you are named in most runs of most prompts, your job is to protect that and fix any wrong facts. If you are named only on branded prompts, the assistants know you exist but do not think of you for the job, which usually points at stages four to six. If you are never named and the brand-check prompts return thin or wrong answers, start with stages two and three.

Turn the cited links into a source list: each site, how many times it was cited, for which prompts, and whether you appear on it. The sources that cite competitors but never you are the shortest route to being named, and they recur in every stage that follows.

Score how ready you are

Use the scorecard below to rate the business honestly against the things the rest of this tutorial fixes. Your weakest criteria tell you which stage to give most time to.

If an assistant's crawler cannot fetch your site, the assistant can only describe you from other people's pages. This stage is the cheapest in the tutorial and catches a surprisingly common problem: a business with good reviews and a decent site that is blocked at the robots.txt file or at the CDN without anyone knowing.

Know which crawler feeds which assistant

Most providers run separate crawlers for search answers and for model training. For being recommended, the search crawlers are the ones that matter. The table follows each provider's own documentation in October 2026.

Google-Extended is the one that trips people up. Blocking it does not take you out of AI Overviews or AI Mode, but it does opt your content out of grounding in the Gemini app.⁹ A business that wants to be named in Gemini should leave it allowed, and accept that the same token also covers training. Training crawlers such as GPTBot are a separate choice: OpenAI says each of its settings is independent, so you can block GPTBot and still allow OAI-SearchBot.⁵ Training crawling also sends back very few visitors. Across the Cloudflare customer sites it measured, Cloudflare's Radar data for July 2025 showed 38,065 Anthropic crawls and 1,091 OpenAI crawls for every visitor referred, against 5.4 for Google, and training made up nearly 80% of AI bot crawling.¹¹ I have set out worked robots.txt configurations for each goal, and the two below cover what most local businesses need.

Two robots.txt files that work for a local business

Two parsing rules decide whether a robots.txt does what you meant. Under the robots.txt standard, a crawler obeys the group that matches its user agent and falls back to the User-agent: * group only when none matches, so once a crawler has its own group it ignores the User-agent: * rules. If the same user agent appears in more than one group, the rules are combined, and matching is not case-sensitive.¹² Google applies the same rules and, where several groups could match, follows only the most specific one.¹³ The safest pattern for a business that wants to be recommended is to leave the search crawlers unnamed, so they follow your ordinary rules, and to name only the crawlers you want to block.

Notice what version two leaves out. It does not name Google-Extended, because blocking that token also opts you out of grounding in the Gemini app.⁹ It does not name the search crawlers in an Allow: / group, because that would make them ignore your User-agent: * disallows. And it does not try to block the user-triggered fetchers, ChatGPT-User and Perplexity-User, which fetch a page when someone asks an assistant about it: OpenAI and Perplexity both say robots.txt may not apply to those anyway.⁵⁶ Anthropic is the exception: it says its bots, Claude-User included, honour robots.txt, so do not name Claude-User in a block group unless you mean it.⁷

Check the CDN and the website builder are not overruling the file

robots.txt is a request. A CDN, firewall or website builder can block a crawler before it ever reads the file, and OpenAI says the host or CDN must allow traffic from its published search bot IP addresses for a site to be eligible for ChatGPT search.² Start with a test that shows what each crawler user agent receives, then look at the settings of whatever sits in front of your site. The CDN guide goes further into verifying real crawler IPs and reading edge logs.

If the site sits behind Cloudflare, four settings matter, and Cloudflare documents each of them. AI Crawl Control lists each AI crawler with its allowed and unsuccessful requests and its robots.txt violations, with an Allow or Block action per crawler; a block is enforced by creating or updating a WAF custom rule, and on the Free plan crawlers are identified by user agent.¹⁴ AI bot traffic policies, available on every plan, let you block Search, Agent and Training traffic separately, on all pages or only on pages with ads; for new domains from 15 September 2026 the defaults block Training and Agent traffic on pages with ads and leave Search allowed.¹⁵ Agent is the category that covers chat fetch bots, the kind of fetcher ChatGPT-User is. Managed robots.txt prepends Cloudflare's own rules to your file, disallowing crawlers that include GPTBot, ClaudeBot and Google-Extended.¹⁶ And Bot Fight Mode challenges traffic that matches known bot patterns, runs outside the rules engine so Skip, Bypass and Allow rules have no effect on it, and logs what it challenges in the Security Analytics events.¹⁷

Website builders have their own switches, and their defaults change, so check yours. At the time of writing, Squarespace has a Crawlers setting with a checkbox labelled “Block known artificial intelligence crawlers”. Its list includes GPTBot, ClaudeBot and Google-Extended, does not include OAI-SearchBot or PerplexityBot, and Squarespace warns that excluding AI crawlers might affect site traffic.¹⁸ Wix has a Robots.txt Editor under SEO & GEO, then Tools and settings, and its help article on blocking AI crawlers gives example lines that include ChatGPT-User.¹⁹ Copying that line would stop ChatGPT reading your page when a customer asks it to, which is the opposite of what a business that wants recommending needs. On WordPress, robots.txt is usually controlled by an SEO plugin or a file on the server, and security plugins can add their own bot blocks, so run the curl test after any plugin change.

See the page the way a non-rendering crawler does

A crawler that fetches your page but does not run its JavaScript sees only what is in the HTML the server sends. Googlebot can process JavaScript as long as it is not blocked,²² and Bing's crawler uses a recent version of the Microsoft Edge rendering engine.¹⁰ OpenAI, Anthropic and Perplexity do not say whether their crawlers run JavaScript in the documentation I checked, but Vercel's analysis of crawler traffic across its network in late 2024 found that none of the major AI crawlers it saw, including OAI-SearchBot, GPTBot, ClaudeBot and PerplexityBot, executed JavaScript; some fetched script files without running them.²³ So the safe assumption for every assistant is that anything you want quoted must be in the page source. There are three quick ways to see it, and I have written up how to check what crawlers see on a JavaScript site in more detail.

An assistant naming a local business reads several sources at once: your site, your profiles, directories, review platforms. Nobody outside the providers knows exactly how they weigh agreement between those sources, but my working assumption, set out in how an assistant decides which businesses to name, is that a business described the same way everywhere is a safer one to name than one with three phone numbers in circulation. The search engines underneath point the same way. On Google, the profile name should be your real-world name, as used consistently on your storefront, website and stationery.²⁵ Bing Places says it estimates a business's popularity from web signals, and that the more a business is mentioned across the web, the more likely it is to appear in results.²⁶

Write one master record

Everything in this stage copies from a single document. Write it once, agree it with whoever answers the phone and whoever runs the website, and treat it as the only source of truth.

Google Business Profile, step by step

Three profiles come first, because each one feeds a family of search products that the assistants use. Google's is the most important. Business Profiles can help your products and services be visible in Google's AI responses as well as in other Search results,²² and keeping the profile up to date is one of the documented practices for its AI features.³ I have written about why the profile often does more work than the website.

To edit the profile on a computer, sign in to the Google account linked to it, go to business.google.com and select Edit profile. Changes are reviewed before they go live, some businesses must verify before they can edit at all, and a business that moves must verify again.²⁷ Service areas are under Edit profile, then Location: next to Service area, select Edit, choose or type each area, and Save. You can set up to 20 service areas, and the overall area should not extend beyond about two hours' drive from where the business is based.²⁸

Bing Places, including the import from Google

Bing Places matters because Microsoft documents Copilot searching Bing to ground its answers, and because Bing is open about how its listings work. It says it ranks local results on relevance, distance and popularity, and that complete and accurate Bing Places data is the first thing to fix.²⁶ It also says Bing Places is one of several sources for a listing: other sources can take priority if they are fresher or if they conflict with Bing Places, many of them are shown in a “Data from” section at the bottom of the listing, and the basic details (name, address, phone, website, hours, category and status) are checked automatically against trusted external data, so a change that disagrees with other sources may not publish or may be overwritten.²⁹ That is consistency enforced in software, and it is why this stage fixes the directories as well as the profiles.

Apple Business

Apple Business Connect was folded into a new free service, Apple Business, from 14 April 2026. Existing claimed locations and place card details migrated automatically, and place cards show across Apple Maps, Safari, Spotlight and other Apple services.³¹ If you claimed your location under the old name, check it survived the move; if you never claimed it, do it now.

Every other listing

Situational prompts carry details: the problem, the property, the area, the day, the budget. An assistant can only match those details to your business if some page states them. There is no special markup, AI text file or Markdown you need to add to appear in Google's AI features, and the content it values has a unique, experienced view, unlike commodity content that restates what is available elsewhere.²² Academic work points the same way. The GEO paper (Aggarwal and others, KDD 2024) tested ways of rewriting web content against a benchmark of queries across many subjects and found they could raise a source's visibility in generative engine answers by up to 40%, with results that varied by domain.³³ It was a lab benchmark, not a study of local businesses, but it supports the idea that how a page states its facts changes whether it is used. For a service business, that means pages that say exactly what you do, where, for how much and how, in a way only you could write.

What a quotable service page states

• The service named the way a customer says it, in the heading and the first sentence: “Boiler repair in north and west Leeds”, not “Heating solutions”.
• The area as real place names and postcode districts. A list such as “Headingley, Kirkstall, Horsforth, Adel, Meanwood (LS5, LS6, LS16, LS18)” can be matched to a prompt. “Leeds and surrounding areas” cannot.
• A price, a starting price, or a range, with what is included. If every job is quoted, explain how quotes work and what typically drives the price.
• How quickly you can attend, your hours, and what happens out of hours or at weekends.
• The process in a few steps: what happens on the first call, the visit, the quote, the work, the paperwork.
• Credentials by name and number, with a note on how a customer can check them.
• Who the service suits and who it does not. Saying “we do not work on commercial boilers” helps the right customers find you and the wrong ones go elsewhere.
• Two or three recent jobs described in a sentence each: what was wrong, where (by district), what was done.

Credentials deserve a specific word. Many trades and professions have a register that customers are told to check. For gas work in the UK, for example, anyone employed to work on gas appliances in homes must be Gas Safe registered,³⁴ so Hartley's boiler pages state the registration number and say how to check it on the register. Whatever your sector's equivalent is, from a professional body to an accreditation scheme, name it and give the number on the page. It is the kind of fact an assistant answering “which plumbers are properly qualified?” has to find somewhere.

Most service pages start vague because they were written to sound good. An assistant can help you turn one into specific, quotable passages, provided it is told not to invent anything. The prompt below asks it to mark every fact it needs from you, so you fill the gaps with real numbers instead of letting it guess.

Area pages, and when not to write them

A page for every town you cover, with the town name swapped in, is the fastest way to make your site worse. Creating content for every variation of how people search, primarily to manipulate rankings or AI responses, breaches Google's scaled content abuse policy.²² Write an area page only where you have something specific to say about that area: jobs you have done there, a local office, housing stock that changes the work, parking or access arrangements. The guide to location pages that are not doorway pages sets out the test I use. For many single-site businesses, one well-written area list on each service page does the job, and most local businesses do not need more content at all.

Structured data, where it is legitimate

Structured data is not required for Google's generative AI features,²² and any structured data you use should match the visible text on the page.³ It is still worth adding LocalBusiness markup to the home or contact page, because it states your details in a form machines read without guessing. For Google's LocalBusiness markup, name and address are required, properties including telephone, url, openingHoursSpecification, priceRange and geo are recommended, and you should use the most specific LocalBusiness subtype available.³⁵ The subtypes are defined by schema.org, whose LocalBusiness page lists more specific types from Dentist and LegalService to HomeAndConstructionBusiness.³⁶ aggregateRating and review are only recommended for sites that capture reviews about other local businesses,³⁵ so do not mark up your own testimonials as a star rating.

Microsoft gives similar advice for Copilot and Bing's AI answers: clear headings, tables and FAQ sections help surface key information, and regular updates help AI systems reference the current version of a page. It also suggests using IndexNow to tell search engines when content changes.²¹ If your platform supports IndexNow, switch it on.

If you work on the site with an AI coding agent such as Claude Code or Cursor, two skills in my free skill pack are relevant here. The Understand skill adds JSON-LD that matches the visible page and refuses to mark up things that are not there, and the Cite skill checks a page for self-contained answer blocks, consistent entities and AI crawler access. Neither can guarantee a citation, and the Cite skill says so plainly.

Your own site is one source among several, and for a local recommendation it is often not the most influential one. The search engines' own local documentation backs this up. Links to a business from other websites, and its reviews, count towards prominence in Google's local ranking.³⁷ Bing says listings with more positive reviews on sites like Yelp and Tripadvisor count as more popular, and advises listing your business on trusted websites.²⁶ Neither has said how its assistant weighs those sources, and Google warns that seeking inauthentic mentions across the web is not as helpful as it might seem, because its AI features depend on both its ranking systems and its spam systems.²² So the job is to be present, accurately, where your sector is genuinely discussed.

Find the directories, platforms and registers for your sector

Your baseline shows which sources the assistants cite today. The other half of the list is the sector's own: the registers customers are told to check, the trade bodies whose member lists are public, and the platforms that vet businesses. Some examples, to show the kind of thing to look for. For gas work in the UK, the Gas Safe Register, because anyone employed to work on gas appliances in homes must be on it.³⁴ For solicitors in England and Wales, the Law Society's Find a Solicitor, a free service listing organisations and people regulated by the Solicitors Regulation Authority.³⁸ For home trades, endorsement schemes such as Which? Trusted Traders, which says every trader listed has passed an assessment set by trading standards professionals and that its reviews are moderated.³⁹ For restaurants, shops and hospitality, Bing itself names Yelp and Tripadvisor as the kind of site where reviews feed its view of a business's popularity.²⁶ An assistant can help draft the list for your sector, as long as you check every entry.

Comparison pages and paid placements

Many “best plumbers in town” pages are run by directories and lead-generation sites, and some charge for inclusion. Be careful with any that does not say so. In the UK, the CMA's guidance on the fake reviews ban notes that commissioning banned reviews can also breach other banned practices, including using editorial content to promote a product the trader has paid for without making that clear.⁴⁰ Under Google's spam policies, buying links for ranking purposes is link spam unless the links are qualified as sponsored or nofollow, and third-party content placed on a site mainly to exploit that site's ranking signals is site reputation abuse.⁴¹ A clearly labelled paid listing on a directory people actually use is advertising, and you can judge it as advertising. A pay-to-appear “independent” top ten is a risk to you and the publisher, and I would decline it.

Trader recommendation sites are a special case. The CMA has said these platforms should check traders before letting them advertise, monitor their performance, and run an impartial process for consumer reviews.⁴² A legitimate one will ask for proof of insurance, qualifications or identity. Treat that as a reason to trust it, and keep your documents ready.

Reviews do two jobs here. Search engines use them in local ranking: more reviews and positive ratings can help a business's local ranking on Google,³⁷ and Bing counts positive reviews and ratings on review sites towards popularity.²⁶ They are also one of the few places where customers describe your work in their own words, which tend to be the words people use when they ask an assistant. No provider has published how its assistant weighs reviews, so the aim is a steady flow of genuine, recent, specific reviews on the platforms that matter to your sector. Customers check them too: in BrightLocal's 2026 survey, 88% of the US consumers who had used AI for local recommendations said they checked whether a review the AI cited was legitimate or looked at its source.¹

The rules, for Google and in the UK

Two sets of rules apply, and where they differ, follow the stricter. Google's policy for Maps content applies to reviews on your Business Profile. In the UK, the Digital Markets, Competition and Consumers Act 2024 made fake reviews and concealed incentivised reviews banned practices from April 2025.⁴³ The guidance below is for the UK; other countries have their own consumer rules, so check yours.

A review routine that holds up

The monthly test is what tells you whether the work is landing. Run the same prompts, in the same assistants, under the same conditions as the baseline, and compare. Monthly is enough for most local businesses. Weekly tests mostly measure the variation between runs, unless you have just fixed something specific like a crawler block. I have written up the method behind this in more detail, including how to pool the counts.

A one-page monthly summary

Keep the monthly write-up short enough that somebody actually reads it. The table below is the layout I would use; the figures in it are invented to show the shape.

Reading what you see

What nobody can tell you

Some changes have known timings. OpenAI says robots.txt changes take about 24 hours to reach its systems,⁵ Perplexity says up to 24 hours,⁶ and Bing Places changes take up to two weeks to publish outside North America.²⁹ Beyond that, nobody outside the providers knows how sources are weighted or how long a new review or directory listing takes to show in an answer. Google says no third-party tool has access to its internal ranking or AI systems,²² and OpenAI says placement is not guaranteed.² My expectation, which is judgement and not a promise, is that access and detail fixes show within weeks, and that source and review work takes several months to move the share of runs that name you.

A spreadsheet is enough to start. If the number of prompts, towns and assistants grows past what you can run by hand in an afternoon, a tracking tool can do the running. I built SearchOps, a platform that tracks how businesses appear in local and AI search, and it is one option. Whatever you use, ask to see the prompts, the number of runs and the location each test is run from, so you can judge the method behind the numbers.