Ask an AI assistant “who's a good electrician in Cardiff” or “where can I get my cat groomed near Harrogate” and it will usually name two to five businesses, with a sentence on each. For the businesses named, that is an easy customer. For everyone else, it is invisible competition.
Nobody outside the AI companies knows exactly how assistants choose businesses, and the companies do not publish a formula. But enough is documented, and enough can be observed by testing, to give a clear picture of what they look at and what a business can influence.
Step one: the assistant decides it needs to search
For a question about local businesses, a modern assistant will normally search rather than answer from memory, because the model's training data is out of date and thin on small local firms. Whether it searches can depend on the product, the settings and how the question is phrased. Adding a place name and a word like “now”, “open” or “best” makes a search more likely.
Each assistant searches in its own way. The table below summarises what the companies have said publicly. Where they have not said, we have marked it.
| Assistant | Main web source | Local and map data | Confidence |
|---|---|---|---|
| Google AI Overviews, AI Mode, Gemini | Google’s search index | Google Maps and Business Profiles | Documented by Google |
| Microsoft Copilot | Bing search | Bing Places and Bing Maps data | Well established |
| ChatGPT | OpenAI’s own crawler plus third-party search providers | Third-party map and listings data; providers not fully disclosed | Partly documented |
| Perplexity | Its own index plus live page fetching | Mixed: listings, review sites and business pages | Partly documented |
| Claude | Web search through a search provider, plus Anthropic’s crawlers | Via whichever sites search returns | Partly documented |
Step two: it gathers candidates from sources it trusts
Assistants do not usually visit fifty plumbers' websites. They read a handful of results, and for local questions those results are dominated by a few kinds of source:
- Map and listing data: Google Business Profile, Bing Places, Apple Business Connect. These give the name, category, location, hours and ratings in a structured form.
- Review and recommendation platforms relevant to the sector, such as Tripadvisor for hospitality, Checkatrade or TrustATrader for trades, Doctify for healthcare, and so on.
- Directories and “best of” lists, including local news sites, council business directories and trade body “find a member” pages.
- Business websites themselves, especially pages that match the question closely.
- Community discussion such as Reddit threads and local forums, which some assistants cite frequently.
Step three: it weighs the evidence
Once an assistant has candidates, it has to decide which to mention and in what order. Based on what the companies have said about quality and on the patterns visible in answers, the signals that seem to matter most are these.
Relevance. The business clearly does the thing asked about, in the place asked about. A salon that lists “balayage” on a dedicated page will match “balayage specialist in Leeds” better than one whose website just says “hair”.
Reputation. Ratings and, more importantly, what reviews say. Assistants often summarise review text, so reviews that mention specific services, staff and places are more useful than a wall of five-star ratings with no words.
Agreement between sources. If your website, listing and three directories all say the same thing, the assistant can state it confidently. Conflicting hours, names or phone numbers are a reason to leave you out or hedge.
Prominence. Being mentioned by sources the assistant already trusts: local press, trade bodies, well-known platforms, other businesses.
Distance and availability, where the question asks for it. “Open now” or “near me” questions depend on accurate hours and a known location.
What assistants cannot see
It helps to know the blind spots, because they explain many puzzling answers.
- Content behind logins, such as a private Facebook group or a members-only directory.
- Text inside images, such as a price list uploaded as a photo or opening hours on a poster graphic.
- Sites that block AI crawlers, often unintentionally through a hosting or security setting.
- Information that exists only on social media apps and is not publicly indexed.
- Word of mouth that never made it online, however strong your reputation offline.
What a business can influence
Almost everything that makes an assistant more likely to recommend you is ordinary reputation work, made machine-readable.
- Allow the main search and AI crawlers (Googlebot, Bingbot, OAI-SearchBot, PerplexityBot, Claude-SearchBot and the user-fetch agents) and check nothing in front of your site blocks them.
- Write a specific page for each main service, stating the service, the place and the key facts in text.
- Complete and match your listings on Google, Bing and Apple, and the platforms your sector’s customers use.
- Ask for reviews in a way that invites detail: “what did we do for you, and how did it go?”
- Get mentioned where your community already looks: the local paper, the parish magazine website, your trade body, suppliers and partners.
- Keep facts current. When hours or services change, update every source the same week.
A worked example
Imagine someone asks an assistant for “a good independent bookshop in Norwich that does author events”. The assistant searches, and the results include a listings page for bookshops in Norwich, a local news article about a festival, two bookshop websites and a Reddit thread.
One shop's website has a page called “Author events” listing upcoming talks, with dates and a booking link. Its listing on Google and Bing shows the same address and hours, and several reviews mention readings and signings. The festival article names it as a venue. Every source agrees and several answer the exact question.
A second shop may be just as good, and may even host more events, but its website only mentions them in an Instagram embed the assistant cannot read, and its reviews talk about coffee rather than events. It is far less likely to be named, not because it is worse, but because the evidence is harder to find.
That is the pattern for almost every sector. The assistant recommends what it can confirm.
Why answers differ between assistants and between days
You may find ChatGPT names you on Monday and not on Wednesday, or that Copilot and Gemini give completely different lists. That is normal. Each assistant searches a different index, reads a slightly different set of pages each time, and phrases the question to itself differently. Answers are also influenced by the user's location, their settings and sometimes their past conversations.
The practical lesson is to judge over many questions and several weeks, not one screenshot. If you are consistently absent across assistants, there is usually a clear cause: blocked crawlers, thin pages, inconsistent listings or few reviews. If you appear sometimes, you are in the pool and the same work will make you appear more often.
Nerdsharks builds websites designed to be read easily by people, search engines and AI assistants alike, with the facts stated in plain text. Whoever builds yours, the list above is where the effort pays off.