How to optimise your website for ChatGPT

The answer in brief

ChatGPT has two gates. A page first has to be retrievable, and then it has to be chosen.

• Getting retrieved is mostly technical. Allow OAI-SearchBot, check that Bing can index you, and serve your answers in plain HTML.
• Getting chosen is where most pages fail. In one large study, ChatGPT cited only 15% of the pages it retrieved.
• Pages that state the answer early, back it with checkable figures and carry something competitors lack have the best chance.
• Much of the evidence ChatGPT trusts sits off your site, on Wikipedia, Reddit, review platforms, LinkedIn and in the press.
• Adding schema showed no clear citation gain in the best test so far, and nothing links llms.txt to ChatGPT citations.

ChatGPT does not hand people ten blue links. It searches, reads a pile of candidate pages, then writes an answer that quotes or paraphrases a few of them. AirOps studied 548,534 pages that ChatGPT retrieved across 15,000 prompts and found that 85% were never cited in the final answer. Search Engine Land covered the study in March 2026. That gap between being retrieved and being cited is what this guide is about.

Why does ChatGPT matter for your website?

Because a large share of research now happens there. OpenAI said in October 2026 that more than 1.2 billion people use ChatGPT each week, up from 900 million in February. ChatGPT also sends most of the traffic that chatbots pass to websites. StatCounter measured its share of AI chatbot referrals at 76.85% in April 2026. That was an all-time low as rivals grew, and by September its share was back at 80.48%, far ahead of every rival.

Google rankings will not get you cited by themselves. When Ahrefs examined the 1,000 pages ChatGPT cited most in September 2025, it found no organic keywords at all for 283 of them. Analyze AI’s August 2026 study of 4,824 already-cited pages, drawn from B2B prompts, found that domain authority explained less than 3% of the variation in how often a page was cited. Rankings still help. In the AirOps data, pages in Google’s first position were cited 3.5 times as often as pages outside the top 20. A high position raises your odds, and the rest of this guide covers what settles the outcome. We look at why a well-ranked site can still be missing in why AI search may not show a site that ranks on Google.

How does ChatGPT find and choose its sources?

ChatGPT search draws on outside search providers and on OpenAI’s own crawling. OpenAI’s help centre says ChatGPT search sometimes partners with other search providers and links to Microsoft’s and Shopify’s privacy statements, without saying what each one does. Independent tests point to Bing. In February 2025, Seer Interactive ran about 100 questions through ChatGPT search and found that 87% of more than 500 citations matched Bing’s top 20 results for the same query, against 56% for Google. Seer called its sample small and its findings directional. Bing visibility is still worth checking, alongside access for OpenAI’s own crawler.

OpenAI also runs its own crawlers, and its crawler documentation lists four:

• OAI-SearchBot finds pages for ChatGPT’s search features. It decides whether you can appear in search answers, although a blocked site can still show up as a plain navigational link.
• GPTBot collects content that may be used to train OpenAI’s models. Blocking it does not remove you from search.
• ChatGPT-User fetches a page when someone asks ChatGPT, a custom GPT or a connected app to look at it. OpenAI notes that robots.txt rules may not apply, because a person started the request.
• OAI-AdsBot checks landing pages submitted as ChatGPT ads. It matters only if you advertise, which our guide to ChatGPT Ads in the UK covers.

OpenAI’s crawling has grown fast. Chris Long’s analysis of about seven billion log files, published on Botify’s blog in April 2026, estimated that OpenAI’s crawl had tripled since August 2025. Long suggested OpenAI may be building a fuller index of its own and fetching fewer pages live. OpenAI does describe an index and cache of web pages that some business workspaces can search instead of the live web, but it has not explained the crawl growth. Either way, a site that crawlers can read cleanly gains from the trend.

What is the technical floor for ChatGPT visibility?

This is worth finishing before you touch content. How long it takes depends on your site and who controls its hosting.

Get into Bing. Verify the site in Bing Webmaster Tools and submit your XML sitemap. Given the Seer findings, it is the cheapest insurance on this list.

Let OAI-SearchBot through. OpenAI asks sites that want to appear in ChatGPT search to allow OAI-SearchBot in robots.txt and to make sure their host or CDN lets traffic from its published IP addresses through. Check the second part with whoever runs your CDN, because bot-blocking rules often catch it. Do not block Bingbot either. Allowing or blocking GPTBot is a separate decision about training and has no effect on search. OpenAI says a robots.txt change can take about 24 hours to reach its search systems.

Serve answers in HTML. Text that only appears after client-side JavaScript runs may never reach a crawler. In searchVIU’s tests, ChatGPT missed a price that only JavaScript displayed. If your pricing, specifications or key answers load that way, open the page with JavaScript switched off and see what is left. Server-render whatever you want cited.

Tell Bing when pages change. IndexNow tells Bing as soon as a URL is added or updated, instead of waiting for the next crawl. Microsoft does not promise a crawl in return, but it costs nothing, and on WordPress a plugin sets it up.

Keep schema in proportion. Article, Organization, Person and, on commercial pages, Product markup remain sensible baselines for rich results and for stating who you are. Do not expect citations from them. In a May 2026 test, Ahrefs compared 1,885 pages that added schema with matched pages that did not, over the 30 days either side. ChatGPT citations rose 2.2%, a change Ahrefs called statistically indistinguishable from zero. The pages were already heavily cited, so the test says little about pages AI rarely surfaces. Google’s own guide to AI features in Search says structured data is not required for its generative AI results and that no special markup exists for them.

Get the floor wrong and no content work rescues you. Get it right and the content decides the rest.

What gets a page chosen once ChatGPT has read it?

Pages that put the answer first. In Kevin Indig’s analysis of 18,012 verified ChatGPT citations, 44.2% came from the first 30% of the text, a pattern he called a ski ramp. An answer buried under three paragraphs of preamble is harder for a model to find and lift than one stated in the opening lines.

Five habits help.

Open every section with its answer. Write each H2 as the question a reader would ask and answer it in the first sentence or two beneath. Some practitioners call this an answer capsule: roughly 40 to 60 words that make sense on their own. Context, caveats and examples come after it. These are editorial habits that make the answer easy to find, and ChatGPT does not require them.

Write sentences a model can quote. “Allowing OAI-SearchBot is required to appear in ChatGPT search” can be lifted as it stands. “Crawler settings may play some role in visibility” gives the model nothing to attribute.

Name your numbers and your sources. The Princeton-led GEO: Generative Engine Optimization paper, published at KDD 2024, found that rewriting pages could raise their visibility in AI answers by up to 40%. In Search Engine Land’s summary, adding statistics, quotations and citations of sources did best, and citing sources helped lower-ranked pages most. The researchers mainly tested a controlled answer engine built on GPT-3.5, with a smaller check on Perplexity, a year before ChatGPT search launched, so treat the percentages as a direction of travel. The lesson stands: a dated figure with a named source is easier to cite than “many businesses”.

Publish what only you have. Pricing breakdowns, test results, client survey data, worked examples and your own methodology give a model something it cannot get from the other pages it retrieved. A roundup of what everyone already says can be swapped for another roundup.

Make each list item stand alone. Passages get lifted out of context, so a fragment such as “Better results” is useless to a model. Every bullet should read as a full statement.

Two formats are worth testing where they fit. A comparison table of options, prices and specifications gives ChatGPT something concrete when a user asks which option to choose. A short block of follow-up questions with answers of two to five sentences does the same for narrow questions. Google stopped showing FAQ rich results on 7 May 2026, so the reason to add one is that readers have those questions.

Length is not the goal. Google’s AI features guide says there is no ideal page length. Write what the question needs and make the answer easy to find.

Does ChatGPT prefer fresh content?

Citation studies show a lean towards recently updated pages. Most of the pages in Ahrefs’ study of ChatGPT’s most-cited pages that showed an update date had been updated that year, though the evidence is correlational. A page with 2024 figures on a fast-moving topic gives a model less reason to trust it than one updated this year. Freshness means substance: new figures, current examples and corrected claims. Changing the date on an unchanged page does nothing for the reader and costs you their trust when they notice.

Pick the pages that matter commercially, put them on a review cycle, and show a visible updated date when the content actually changes.

Why does the rest of the web decide whether ChatGPT trusts you?

Because ChatGPT leans heavily on sources you do not control. Ahrefs classed about two-thirds of ChatGPT’s 1,000 most-cited pages as pages marketers would struggle to influence through ordinary outreach. Your own pages need to say clearly who you are, and other people’s pages need to back it up.

On your own site, name your authors, give each one a bio page with real credentials, and keep them writing on the subjects they know. Yoast’s guidance on AI citations recommends the same, along with consistent terminology across your content. Use one form of your company name, one address and one description everywhere, so a model can tie every mention to the same business.

Off your site, the mix of sources ChatGPT cites changes quickly, so be wary of any single league table. Similarweb’s US data for January and February 2026 put Wikipedia at 13.15% and Reddit at 11.97% of ChatGPT’s citations, well ahead of any other domain. Semrush measures something different: the share of responses that cite a domain at least once. In its tracking of more than 230,000 prompts, ChatGPT cited Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September. Similarweb’s figures are for the US, and UK figures are scarce. If your market is British, see our write-up of Semrush Spotlight’s UK AI visibility benchmark.

The platforms that keep appearing in these studies are the ones to work on:

• Reddit. Give useful answers in the subreddits your buyers read. Many communities ban self-promotion, so follow each one’s rules.
• Wikipedia. Where your company or category meets Wikipedia’s notability rules, a well-sourced article is worth having, given how often ChatGPT cites Wikipedia. You earn it through independent coverage first, so treat it as a medium-term job.
• Review platforms. Check which review sites, such as G2, Capterra or Trustpilot, appear when ChatGPT answers questions about your category, and keep those profiles accurate and current.
• LinkedIn. Profound’s data from real ChatGPT prompts had LinkedIn rising from about 11th to about 5th among cited domains between November 2025 and February 2026. Posts and long-form articles took a growing share of those citations.

Review this work every quarter. The ground moves too fast for a one-off campaign.

What about the questions your buyers never type?

ChatGPT often runs its own supporting searches before it answers, a step called query fan-out. In the AirOps study, 89.6% of prompts triggered two or more of these follow-up searches, taking 15,000 prompts to 43,233 queries. Ask for the best accountants for contractors and ChatGPT may also look up professional body membership, typical fees and sector specialisms, then fold the results into its answer.

Fan-out does not happen every time. Ahrefs studied 66.6 million free, logged-out GPT-5.6 Luna responses from August and September 2026 and found that only 14.4% contained fan-out queries, averaging about 1.6 in those that did. The figure counts recorded fan-out only, so it cannot say whether the other answers used the web. It does show that ChatGPT does not always go looking. Your presence across the web matters alongside any single page.

Where fan-out does happen, AirOps found that 95% of the queries had no monthly search volume in standard keyword tools, so keyword research will not surface them. You can cover the substance a buyer needs, such as prices, methods, specifications, accreditations, case evidence and comparisons, on pages ChatGPT can read. Those are the details it goes looking for.

Should you add an llms.txt file?

Not for ChatGPT citations. OpenAI does not list llms.txt among the things that affect ChatGPT search. Its crawler page links to one only as an index of its own developer docs. Google’s AI features guide says Google Search does not use the file, though creating one does no harm. Crawler logs tell the same story. Limy reported that of more than 500 million AI bot visits to the brands it monitored over 90 days, only 408 requested llms.txt.

The file has a real use on documentation sites, where coding assistants and agents can use it to find the right pages. If developers are a large part of your audience, publish one. Otherwise spend the time on the content and off-site work above.

How do you measure ChatGPT visibility?

OpenAI offers no Search Console for ChatGPT, so measurement takes a mix of manual checks, analytics and third-party tools.

Start with prompt testing. List the 15 to 20 questions a buyer in your category would ask, run them in ChatGPT on a fixed schedule, and record whether you are cited, how you are described and who appears instead. Note the date, the model and whether search was on, because answers vary. It is crude, but it gives you a baseline. When we asked ChatGPT for the best accounting software for a small UK business on 25 October 2026, in a temporary chat, it named six products and put a source beside each one (Figure 4). Those sources are the pages you are competing to become.

Then look at analytics. ChatGPT adds utm_source=chatgpt.com to many of the links it sends, so much of the traffic appears in Google Analytics, though some arrives as direct. If your ChatGPT traffic dropped in October 2026, read our guide to auditing GA4 after the October 2026 UTM bug before you conclude anything. Expect small numbers. Conductor’s benchmarks, from 1,215 enterprise sites in ten industries using US data from May to September 2025, put AI referrals at 1.08% of all website traffic, with ChatGPT sending 87.4% of that share.

The quality of that traffic is contested. Visibility Labs analysed GA4 data from 94 ecommerce brands across 2025 and found ChatGPT visitors converted at 1.81%, against 1.39% for non-branded organic search, 31% higher. The same brands saw a lower average order value from ChatGPT, and non-branded organic search sent about 70 times as many visits. Other research has found no consistent edge: Amsive’s study of 54 sites found no significant difference between AI and organic conversion. Track it on your own site and judge by your own numbers.

Last, watch unprompted mentions of your brand on Reddit, LinkedIn, review sites and in the press. Track them alongside citations, referrals and conversions. Profound, Peec and Similarweb all sell citation-tracking tools that cover several AI assistants, and Microsoft Clarity added a Citations report in May 2026. Check what each one covers before you rely on it.

What changes when agents start shopping for your customers?

Some buyers will never see your site, because an AI agent will read it for them. OpenAI launched Instant Checkout in September 2025, starting with US shoppers buying from Etsy sellers without leaving ChatGPT, and the ChatGPT Atlas browser in October 2025, with an agent mode that could work through tasks on a user’s behalf. OpenAI has since retired Atlas.

The early results were mixed. Walmart’s Daniel Danker told WIRED that products bought directly inside ChatGPT converted three times worse than those reached through a link to Walmart’s own site. In March 2026 OpenAI told Digital Commerce 360 that Instant Checkout was moving into merchants’ apps inside ChatGPT. Later that month it said it would let merchants use their own checkout while it concentrated on product discovery. Our main ChatGPT SEO guide covers how product listings and feeds work in ChatGPT.

Much of the groundwork is the same as for citation. Accurate prices, clear specifications, product data a machine can read and a consistent identity help an agent describe and compare you correctly. Whether an agent can actually buy from you is a separate question of feeds, stock, variants and checkout, so test that on its own.

Where should you start?

Start with the technical floor, because it is quick and nothing else works without it. Check Bing, allow OAI-SearchBot and its IP addresses, and check that your key answers survive with JavaScript off. Then take your five most commercially important pages and rewrite their openings so each section answers its question in the first two sentences, with a named figure where you have one. Run your prompt test before and after, and keep running it. A single comparison cannot separate your edits from changes in the model.

The off-site work takes longer, and it is where much of ChatGPT’s trust comes from. Our guide to GEO for business goes further on building that presence. To find out where your site stands in ChatGPT today, talk to AIWIZ.