What Is ChatGPT Search and How Do You Get Cited?

Three things: ChatGPT has to be able to read your website, understand what you do and see that others mention you.

ChatGPT is no longer a side project for UK searchers. Ofcom’s Online Nation report (December 2025) describes it as already the second-largest search service in the UK, with 252 million UK web visits in August 2025, up 156 per cent on a year earlier. If your customers put their buying questions to ChatGPT and you are missing from the answer, the enquiry goes to a competitor.

Unlike a page of blue links, ChatGPT search follows the whole conversation, so the second and third questions matter as much as the first. This guide covers how it finds its sources, what it needs from your website, which content it cites and how to measure whether you appear.

If you want to be found where your customers ask their buying questions, see how we approach AI search, or book a free search audit.

ChatGPT search is the part of ChatGPT that looks things up on the web. It finds current sources and summarises them in an answer, with citations that link back to the pages it used.

The technique behind it is called retrieval-augmented generation (RAG). When a question needs up-to-date information, ChatGPT rewrites it into one or more search queries, sends them to its search providers and reads the pages that come back. The passages that answer the question best make it into the answer; the rest are left out.

Web search is available on every ChatGPT plan, from Free and Plus to Business and Enterprise, and to people who are not signed in. ChatGPT decides for itself when a question needs the web, and users can also switch search on from the tools menu in the message box.

Searching in ChatGPT works more like a conversation with a knowledgeable shop assistant than a scan down a results page.

Take walking boots. You ask for five suggestions for black walking boots, and ChatGPT returns five pairs with a short description of each. You reply that they need a high ankle, should be as light as possible and should show the price. It narrows the list. Then you give it a budget, and it narrows the list again. Each step builds on the one before, in the same chat.

On Google’s classic search results page, each of those steps is a new search or a filter you set yourself. Google’s AI Mode now takes follow-up questions in the same way, which shows where search as a whole is heading.

Being cited in ChatGPT search depends on two things: ChatGPT has to find your page through the sources it searches, and the page needs a passage that answers the question better than your competitors’ pages do.

Ranking well helps with the first, but less than most people expect. In an Ahrefs study of 15,000 prompts published in August 2025, about 12 per cent of the links cited by AI assistants appeared in Google’s top ten for the same prompt, and for ChatGPT, Gemini and Copilot it was nearer 8 per cent; Perplexity was the outlier at almost a third. The overlap with Bing’s top ten was around 10 per cent on average.

Bing still deserves attention, for reasons covered in the OAI-SearchBot section below. Our view is that Bing rewards exact keyword matches more than Google does and leans less on semantic understanding, much as Google did a few years ago. Being indexed and ranking there is part of the groundwork.

A top position on Google is therefore no guarantee. ChatGPT does not pick the domain with the most links. It picks the passage that answers best.

OAI-SearchBot is the crawler OpenAI uses to find pages for ChatGPT search. If it cannot reach your website, you will not appear in the answers, however good your content is. OpenAI states that sites which block OAI-SearchBot will not be shown in ChatGPT search answers, although they may still appear as plain navigational links.

Many businesses block everything from OpenAI to protect their content. There is no need, because OpenAI uses separate crawlers for separate jobs:

• OAI-SearchBot finds pages for ChatGPT search.
• GPTBot collects content that may be used to train OpenAI’s models.
• ChatGPT-User visits a page when someone asks ChatGPT to open it. Because a person starts the visit, OpenAI says robots.txt rules may not apply.

So you can say no to training and yes to search. They are two separate decisions. To allow search and block training, add a rule for each crawler to your robots.txt file:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

OpenAI says it can take around 24 hours for its systems to pick up a change to robots.txt.

The firewall is the next thing to check. A web application firewall (WAF) or content delivery network in front of your website, such as Cloudflare, Akamai or Amazon CloudFront, can turn OAI-SearchBot away even when robots.txt lets it in. The crawler then gets an HTTP 403 Forbidden response and cannot fetch your pages for ChatGPT search. OpenAI’s help centre asks site owners to confirm that their host or CDN allows traffic from its published search-crawler IP addresses. Pass that to whoever runs your hosting or IT.

Then there is Bing. OpenAI says ChatGPT search sometimes works with other search providers, and Microsoft is among those it names, so Google Search Console is not enough on its own. Verify your site in Bing Webmaster Tools and use IndexNow, a protocol that tells Bing and the other participating search engines as soon as a page is added, updated or deleted, instead of waiting for the next crawl.

This is usually the least effort of anything in this article. But if one line in robots.txt or one firewall rule is wrong, none of the rest matters.

Entity authority is how clearly and consistently your company is described, both on your own website and elsewhere. If sources disagree about what you do, where you are or who runs the business, ChatGPT has little reason to name you with confidence.

Building it takes more than naming the entities that matter to you, such as your brand, products, places and people. You also need to show how they relate to one another.

1. Define your core entities: which brands, products, places and people are central to the business?
2. Map how they relate: write the relationships down in plain sentences, such as “Sunbird is a search agency that specialises in…”.
3. Mark them up: Organization, LocalBusiness and Person schema make the relationships explicit, and sameAs connects your website to your LinkedIn page and other official profiles.
4. Link related pages: link deliberately between pages that cover related entities.
5. Earn independent mentions: work towards being named by sources that AI services trust and retrieve.

The more consistent and structured the information about you is, the easier it is for a model to state facts about your business without hedging.

ChatGPT gives more weight to what others write about your company than to what you write about yourselves. For buying questions, comparison articles of the “Best X for Y” kind are a common source. Across more than 230,000 prompts tracked by Semrush between July and October 2025, Reddit and Wikipedia were the two domains ChatGPT cited most. Buying links will not get you into those sources.

Picture two companies in the same sector. One calls itself the market leader on its own website. The other appears in two independent comparisons, has reviews and is mentioned in a Reddit thread. Which one will ChatGPT recommend? We would put our money on the second.

If nobody outside your own website mentions you, a competitor who is mentioned will be recommended ahead of you. Put your most important buying questions to ChatGPT, note which pages it cites and work on being included on exactly those. On forums such as Reddit, automated posts will get you nowhere, but you can answer questions factually, correct errors and show that you know your subject.

Content that ChatGPT cites answers one question per paragraph, makes sense on its own and keeps a factual tone.

ChatGPT does not read your page from top to bottom as a person would. It works with smaller passages of the text (often called chunking), compares them with the question and uses the ones that answer best. A paragraph that opens with “This also means that…” means nothing once it is lifted out of context.

Tone matters too. Phrases such as “market-leading” and “end-to-end solution” give a model nothing it can repeat as fact. Plain text that reads like a reference work is far easier to lift into an answer, which is an awkward lesson for any agency that likes to talk about itself.

Use names. “Our solution” and “the platform” mean nothing to a model that sees one paragraph in isolation; “Sunbird” and “WordPress” do.

Sourced figures help. In a 2024 study by researchers at Princeton University and IIT Delhi, adding statistics, quotations or source citations raised a text’s visibility in generative search answers by 30–40 per cent on the study’s main measure. In the same study’s test on Perplexity, keyword stuffing did 10 per cent worse than the original text.

A common objection is “but we have the keyword in every heading”. That is not how it works: the model is looking for the answer, not the keyword.

Go through your most important pages with this checklist:

• The heading is the question the customer asks.
• The first sentence under the heading answers it.
• The paragraph makes sense without the rest of the page.
• Names appear where it would otherwise say “we”, “the system” or “the solution”.
• Every figure has a source with a name and a year.
• Comparisons sit in a table, not in running text.

In a table, it is hard for a model to mix up which price belongs to which product. The same goes for your customers.

Optimising for follow-up questions means a page answers the whole chain of questions a customer asks in ChatGPT, not just the first one. The same applies to deep research, where ChatGPT runs several rounds of searches before it writes its report.

To structure a page for follow-up questions:

1. Identify the question sequences: which follow-up questions do your customers usually ask after the first one?
2. Build comparison tables: make it easy to compare products or services on several points at once.
3. Add an FAQ section: answer the most common follow-up questions directly on the page.
4. Define technical terms: explain the terms of your trade so the answer can use them correctly.
5. Treat FAQ markup as optional: FAQPage schema labels questions and answers for machines, but Google stopped showing FAQ rich results in May 2026, so the visible questions and answers are what count.

Someone who asks for the best walking boots for the Lake District in autumn rarely stops there. The sequence might look like this:

• First question: “Best walking boots for the Lake District in autumn”
• Follow-up 1: “Which of these suit narrow feet?”
• Follow-up 2: “How does Gore-Tex differ from eVent?”
• Follow-up 3: “Which is best under £150?”

That is four questions, and one page can answer all of them. ChatGPT search carries the context of earlier questions through the conversation, so a page built for the whole sequence has the answers ready when the follow-ups come.

Since April 2025, ChatGPT search has shown product results when a question suggests that someone wants to buy. A typical shopping answer includes:

• Product images
• Product details, prices and summaries of reviews
• Links to the retailers that sell the product

OpenAI says the product results are chosen independently and are not ads; ads in ChatGPT are shown separately.

As with any other answer, the buyer can keep asking: narrower requirements, a different budget, a comparison of two models. More of the buying journey now happens inside the answer, so expect fewer visits from shoppers, and expect those who do arrive to have decided more of what they want.

According to OpenAI’s help centre, ChatGPT builds its product results from structured metadata such as price and product description, supplied by third-party providers or by merchants themselves. Shopify merchants are already covered through Shopify Catalog. Other retailers can apply to send OpenAI a direct product feed, which keeps prices and availability current in ChatGPT. When several merchants sell the same product, OpenAI says it ranks them on availability, price, quality and whether they are the maker or main seller.

Your own product pages still matter, and basic product schema is only the start.

These are the schema.org types and properties to get right:

1. “Product”: the basic product information
2. “aggregateRating”: the combined rating from your reviews
3. “review”: individual reviews with detailed feedback
4. “offers”: price, availability and delivery options
5. “brand”: a clear brand identity, which is also an entity signal

Then write product descriptions that work in a conversation:

• Answer comparison questions: “This model differs from X in that…”
• Cover use cases: “Best suited to you if…”
• Include dimensions and specifications: they are what a buyer’s requirements get matched against
• Say what the product is not suited to: it helps rule the product out for the wrong use

Reviews on every platform where you sell, whether your own website, Amazon or eBay, shape the picture ChatGPT forms of a product. In your own structured data, though, mark up only reviews collected on your own website. Google’s review guidelines are explicit: “Don’t aggregate reviews or ratings from other websites.”

• Long-tail keywords, found through keyword research
• Technical SEO
• Structured data
• Mentions on sources that AI services trust and retrieve
• Content that goes deeper and states exactly what you do
• Entity optimisation

Two steps you can take this week:

• Verify your website in Bing Webmaster Tools, if you have not already.
• If you track keywords on Google, track them on Bing too.

SEO is becoming part of everything you do on your website rather than a separate channel.

ChatGPT search has become a traffic source in its own right, mainly for complex questions and product comparisons, though for most businesses it still sends far fewer visits than traditional search. Informational searches will lose the most clicks; how many depends on how far people come to trust the AI answers.

But someone who needs a joiner or new walking boots still ends up on a website, so we do not expect enquiries to fall as fast as traffic. And if AI answers spread their citations across more sources than Google’s top results do, the traffic that remains may be shared among more websites.

Most businesses measure rankings and traffic. For informational searches, a better measure may now be how often your domain appears as a cited source in ChatGPT’s answers, often called AI visibility or citation share. AI visits without a clear referrer can also land in Direct or Unassigned in GA4, so analytics alone will not show the full picture.

You find out whether your business shows up in ChatGPT by asking the same customer questions at regular intervals and counting how often you are named, linked as a source or recommended.

Start simple. Write down 20 questions your customers ask before they buy, such as “which accountancy firm in Leeds suits a growing limited company?”. Put them to ChatGPT once a month and note three things:

1. Are you named in the answer?
2. Is your website linked as a source?
3. Which competitors are named instead?

If you are named in, say, 4 of 20 answers and a competitor in 11, you know where you stand and which questions to work on first. The answers vary from one run to the next, so ask each question more than once before drawing conclusions.

Once the list grows, there are platforms that do the same job automatically. Ahrefs Brand Radar and Authoritas are two examples: they collect AI answers and show how often a brand is named compared with its competitors. Before you pay, check that the tool can track the UK market and ChatGPT specifically, not only Google’s AI Overviews.

Then there is the traffic. Visits from chatgpt.com usually show up as referrals in your analytics, and they are worth following all the way to a lead.

What counts is not the number of mentions but what they bring in. Ten visits from ChatGPT that turn into two leads are worth more than 200 visits that turn into none.

AI search gives businesses three challenges: fewer clicks, generic keywords and the need for up-to-date information. Each one has a matching opportunity.

When the AI answers the question itself, fewer people click. Build landing pages that help the visitor act, not just learn: they have already had the basics from the answer.

For broad questions, the AI pulls together information from several sources, so no single page owns the answer. Focus instead on narrow, specific use cases where you can be the only authoritative source: “best shoes” is hard to win, but “best walking boots for narrow feet with bunions” is your chance.

When a question needs current details, ChatGPT looks for fresh information, and a page with old statistics risks being passed over or producing a wrong answer. Keep a content calendar for updates: small changes, such as new examples, fresh statistics and current references, can be enough to give you the edge.

You measure visibility in ChatGPT by tracking a fixed list of customer questions over time and seeing what share of the answers name or link to your business. Add referral traffic from chatgpt.com, and a tracking tool once there are too many questions to run by hand.

ChatGPT rewrites the question into one or more searches, draws on pages found by OpenAI’s own crawler and by partner search providers, and picks out the passages that best answer the question. Passages with a clear answer, factual language and sourced facts make it into the answer. When ChatGPT recommends a supplier, what independent sources say about the company also carries a lot of weight.

A search agency that works on the technical side, on content and on mentions beyond your own website can help you get cited in ChatGPT. Sunbird does exactly that, so we have an interest here. Whoever you choose, ask to see how they measure mentions in ChatGPT before you sign. An agency that only reports Google rankings is measuring the wrong thing.

Ahrefs Brand Radar and Authoritas are among the tools that track how often a brand appears in AI answers. For most businesses, a spreadsheet with a fixed set of questions goes a long way at first; a paid tool earns its keep when you want to compare yourselves with many competitors.

To appear in AI search, your website has to be open to the AI crawlers, your content needs passages that can be lifted out as answers, and independent sources need to mention you. The foundations are the same for ChatGPT, Perplexity and Gemini, but the indexes differ, and Google’s AI Overviews draw on Google’s own index. Taken together, this work is usually called generative engine optimisation, or GEO.

Put your effort where ChatGPT search is strongest: in-depth guides that answer the whole chain of follow-up questions, and detailed product data if you sell products. Ranking on Bing is only part of it; structure your content so that it can become a primary source in an AI-generated answer. Our guide to generative engine optimisation covers the wider groundwork. For this article, work through these steps in order:

1. Let OAI-SearchBot in: check robots.txt and the firewall
2. Review your Bing rankings: look for big differences from Google
3. Add product data: if you sell products, get “aggregateRating” and “review” right and look at a product feed
4. Map question sequences: which follow-up questions do your customers ask?
5. Build entity authority: describe your business consistently, around clear entities and relationships
6. Set up measurement: start tracking citations, with branded searches as a proxy

ChatGPT cites whoever answers most clearly and whoever others talk about. Mentions take time to build, which is also why a competitor cannot simply outspend you on them.

Sunbird is a search marketing agency. If you want ChatGPT to name you ahead of your competitors, see how we approach AI search.

Or book a free search audit with me, Nils. It takes 30 minutes and there is nothing to prepare.