ChatGPT Search Implementation Checklist: From Pilot to Production

OpenAI documents several user agents, and they do different jobs. Confusing them is the most common mistake in robots.txt files.

• OAI-SearchBot: used to surface websites in ChatGPT search results. Blocking it can keep your pages out of search answers.
• GPTBot: used to crawl content that may be used to train OpenAI’s models. You can block it without blocking OAI-SearchBot.
• ChatGPT-User: used when a user’s request makes ChatGPT visit a page directly, such as following a link.

Many publishers block GPTBot for training reasons while allowing OAI-SearchBot so they can still appear in search. Decide this deliberately and write the decision down, because firewalls and CDN bot protection can block these agents even when robots.txt allows them.

• Check robots.txt for rules affecting OAI-SearchBot, GPTBot and ChatGPT-User.
• Check CDN or firewall settings, such as Cloudflare bot rules, for blocks on AI crawlers.
• Confirm key pages render their main content in HTML without needing heavy JavaScript.
• Choose ten pages that answer important buyer questions.
• Write twenty test prompts based on real questions from sales calls and support tickets.
• Run each prompt in ChatGPT with search enabled and record which sources are cited.
• Set up an analytics segment for referrals from chatgpt.com.

OpenAI adds a utm_source=chatgpt.com parameter to many links from ChatGPT search, which makes the traffic easier to spot in GA4 or similar tools. Not every visit carries it, so combine referrer and UTM data.

Answers in ChatGPT search draw on pages that clearly address the question. That rewards the same habits that help featured snippets and Google’s AI features: lead with the answer, keep sections self-contained and include details a model can quote with confidence.

• First paragraph: a direct answer in two to four sentences.
• Headings phrased as questions buyers really ask.
• Specifics such as steps, settings, comparisons and example ranges labelled as examples.
• Visible author, date and update information.
• Product and pricing pages with plain-text facts, not only images.
• Consistent facts across your site and third-party profiles, so sources do not contradict each other.

Being mentioned elsewhere matters too. Answers often cite review sites, comparison articles, forums and news coverage. A brand absent from those sources is less likely to appear, regardless of its own site quality.

After four to six weeks, the pilot should show three things: whether your site is reachable, whether your pages are cited for any test prompts and whether referral traffic converts. Production means turning the manual checks into a routine with owners.

Production checklist

• Assign an owner for crawler access, usually the technical SEO or web team.
• Add AI crawler checks to the release checklist, so a new firewall rule does not silently block them.
• Expand the prompt set to fifty or more, grouped by funnel stage.
• Run the prompt set monthly and log citations in a spreadsheet.
• Report chatgpt.com referrals, landing pages and conversions next to organic search.
• Update the pages that are close to being cited but are not, based on what cited competitors include.

Referral sessions and conversions are measurable. Citation frequency in your prompt set is observable but not a ranking: answers vary with wording, conversation history and time. Treat monthly sampling as a trend indicator. Avoid tools or vendors that promise fixed positions in ChatGPT, since no such position exists.

The first is blocking the wrong crawler. Teams that decide to block AI training sometimes add a broad rule that also blocks OAI-SearchBot, then wonder why their pages never appear. The second is testing with prompts written by the marketing team rather than real buyer questions; those prompts tend to include the brand name and give a flattering picture.

• Using logged-in chats with long history, which personalise results; test in fresh conversations.
• Counting one appearance as success instead of tracking a prompt set over months.
• Rewriting pages without a baseline of citations and traffic.
• Ignoring third-party sources, which often drive more citations than your own pages.
• Forgetting product feeds and structured data for shopping-related questions.

Fixing these early keeps the pilot honest and makes the production phase much easier to justify.

One more point for retail brands: shopping-style answers in ChatGPT can draw on structured product information, so keep product schema, prices and availability accurate and consistent with what your feeds and marketplaces show.