More people now ask an AI assistant for a recommendation before they ask Google. “Who fits kitchens near Leeds?” “What is a good waterproof jacket for walking in Scotland?” The assistant answers with names, and the business it names gets the call.
You cannot pay to be one of those names, and anyone selling a guaranteed place in ChatGPT is selling something they do not control. But you can make it much easier for an assistant to find you, read you and trust what it reads. Most of that work is the same work that ranks you on Google. A few parts of it are new, and one of them caught us out on our own site.
How an assistant decides who to recommend
An assistant answers from two places. Knowing which one it is using tells you what to fix.
- What it learned in training. The model has read a huge snapshot of the web, months or years old. If your business was described clearly and often in that snapshot, it may know you without looking anything up.
- What it finds when it searches. For anything local, recent or specific, assistants such as ChatGPT, Claude, Perplexity, Copilot and Gemini run a live web search, read the top results and summarise them. This is where most recommendations come from, and it is the part you can influence this month.
Both routes depend on the same thing: the assistant has to be able to read your site, and what it reads has to be clear.
1. Check AI crawlers are not blocked
In September an AI assistant told us that our own website “actively blocks AI scraping tools”. It was right. Our robots.txt welcomed everyone, but the network our site runs on had a setting switched on that refused the crawlers OpenAI and Anthropic use to read the web. We had never touched it. Cloudflare has blocked AI crawlers by default on new domains since July 2025, so a great many sites are in the same position without knowing it.
Because the assistant could not read our site, it described us from old directory listings instead, and got us wrong. That is the real cost of a block: you are not left out of the answer, you are described by whoever else wrote about you.
The AI companies each run more than one crawler, and they do different jobs:
| Company | Reads the web for training | Searches and fetches for answers |
|---|---|---|
| OpenAI | GPTBot | OAI-SearchBot, ChatGPT-User |
| Anthropic | ClaudeBot | Claude-SearchBot, Claude-User |
| Perplexity | None listed | PerplexityBot, Perplexity-User |
| Google-Extended (a setting, not a separate crawler) | Googlebot, the same one as search |
You can make a reasonable case for blocking the training crawlers. Blocking the search ones means you cannot be recommended from a live search at all. Whatever you decide, decide it on purpose. Check three places:
- Your robots.txt. Type your address followed by /robots.txt. Look for any of the names above with Disallow: / underneath.
- Your hosting or CDN settings. On Cloudflare, it sits under Security, in the bot settings, as a policy for AI bots. Other hosts and security plugins have their own versions.
- The site itself. Our free website audit now fetches your start page as an AI crawler and tells you whether it was turned away.
2. Be in the indexes the assistants search
When an assistant searches, it uses a search engine’s index. Google’s AI answers come from Google’s index, and Copilot’s come from Bing’s. ChatGPT’s search has drawn on Bing too. A page that is not indexed cannot be found by the search, so it cannot be recommended.
- Set up Google Search Console and submit your sitemap. Check the Pages report for pages Google has found but not indexed.
- Set up Bing Webmaster Tools as well. It takes ten minutes, it can import your settings straight from Search Console, and most businesses have never opened it.
3. Write the facts down in plain text
An assistant recommending a business is answering a question like “who does this, where, for how much, and are they any good?” It can only repeat facts it can find. So put them on the page, as text, in plain sentences:
- What you do, in the words a customer would use. “Boiler repair and servicing” rather than “thermal comfort solutions”.
- Where you do it. The towns or areas you cover, or that you deliver worldwide.
- Who it is for, and who it is not for.
- Prices, or at least a starting price. An assistant asked for something affordable cannot recommend a business whose prices it cannot see.
- Opening hours, contact details and how to book.
Text inside images, sliders and PDFs is much harder for a crawler to read than a paragraph. A short FAQ that answers the questions customers actually ask is one of the most useful pages you can have, because its format matches how people question an assistant.
4. Add schema markup
Schema markup is a short block of code that states your key facts in a format machines read without guessing: business name, type, address, phone number, opening hours, prices and reviews. It does not guarantee anything, but it removes ambiguity, and ambiguity is where assistants make mistakes. A local business wants LocalBusiness markup, or a more specific type such as Plumber. An online shop wants Product markup on every product page. Include sameAs links to your social profiles and listings, so the assistant can connect them.
5. Make every listing say the same thing
Assistants cross-check. If your website, Google Business Profile, Facebook page and three directories give two addresses and describe two different businesses, the assistant either picks one or leaves you out.
The assistant that got us wrong was reading two directory listings that still described us as the marketing agency we used to be. Our own site had been rewritten months earlier, but nobody had updated the listings. Go through every profile you have: your Google Business Profile first, then directories, review sites and social profiles. Make the name, address, phone number and description of what you do match your website.
6. Be mentioned by other people
When an assistant answers “who is the best” or “who is recommended”, it leans heavily on what other sites say: review platforms, local news, industry roundups and “best of” lists. That is the same thing that has always built a reputation, now with a machine reading it.
- Reviews, on Google and on whichever review site matters in your trade, with replies from you.
- Roundups and lists. If there is a “best cafés in” article for your town, or a “best independent outdoor brands” list for your trade, ask to be considered.
- Local press and trade bodies. A membership listing or a short news piece is a mention an assistant can find.
What to skip
- Packages that promise a ranking in ChatGPT. Nobody controls what an assistant says. What can be sold honestly is the work above.
- An llms.txt file as a priority. It is a proposed standard for giving AI tools a summary of your site. It does no harm, but none of the big assistants has said it uses one to decide what to recommend. Do the six steps above first.
- Pages written for machines. Keyword-stuffed text reads badly to a customer and no better to an assistant, which is built to read like a person.
How to tell if it is working
Ask the assistants yourself, the way a customer would, in a fresh chat with search switched on: “Who does [your service] in [your town]?” or “Where can I buy [your product]?” Note whether you appear, what the assistant says about you, and which sources it links to. Those sources are the pages it trusts, and the places to make sure your details are right.
Then check your analytics for visits from chatgpt.com, perplexity.ai and copilot.microsoft.com. ChatGPT tags many of its links with utm_source=chatgpt.com, so they show up as their own source. The numbers will be small at first, but they are worth watching.