How to Show Up in ChatGPT and AI Overviews

Getting a business into ChatGPT search starts with the right crawler. That is not the same one OpenAI uses for training, so a mix-up here can undo all the other work. If we block the wrong one, or block all of them, the writing never gets a chance to matter.

Cited in this note

8%

OpenAI says it runs separate crawlers.¹ OAI-SearchBot fetches pages for search results. If you turn it off, your pages stay out of ChatGPT search answers. GPTBot is the crawler for training. ChatGPT-User fetches a page when a person asks for a look-up, and robots.txt may not apply to that one. These are three different jobs, and we repeat that until the owner can say it back to us. Most robots files treat every bot as the same thing. So a rule written in a hurry mixes them up. The mix-up is quiet, because the page just never shows up and someone blames the writing.

The visibility guide reads the site the way a practical check should. It starts with whether a crawler can see the pages at all. The answer-engine note and the generative-engine note cover the names people use for Google’s summaries. This note stays on the switches, and on the page those switches are meant to show.

What Google says to do

Google’s guide for its generative features says the work is still SEO.² It says answer-engine work and generative-engine work are both SEO from where Google sits. It says Google does not use llms.txt. It also says there is no special rule for how to chunk a page. The page has to be indexed, and it has to be able to show up as a snippet. No extra schema is needed for these features either. If you came here only for ChatGPT, please do not skip this part, because many owners want both. The Google half of the job is still an ordinary page. A special file will not stand in for it on either side.

Say a page is blocked, thin, or not allowed to show a snippet. It will not turn into an overview because someone added a new file. So we let the search crawler in and keep the page open to indexing. Then we write the answer where a person can read it too.

We see this go wrong in a pattern. First, an agency blocks every bot “for safety,” including the search crawler. The owner hears that the site is now protected. Then someone adds a new file because a thread said that file is how you show up. The page underneath is still a list of cities and a phone number. Nothing in that story gives a homeowner an answer they can use. The crawler has to be allowed to fetch. The page has to be open to indexing, and the sentences have to be worth repeating. The file is not the sentence.

Pew found, in July 2025, that visits with an AI summary led to a click 8% of the time. When no summary showed, it was 15%.³ That means showing up is not the same as being clicked. So we write the page so that the citation is worth the click.

The Pew study is about Google’s summaries, but the habit is the same, so we bring it in here. A generated answer can cover the easy part of a question and leave you unvisited. If you win the mention and lose the call, you have only done half of the job. The page has to carry the part a summary cannot finish. That means who sends the truck, what you will not touch, and how a person books.

A worked example

Picture an electrical contractor with us. This is a made-up scene, and it is not a measured result. An office manager runs the office, the inbox, and the site password, and two techs do the work. Homeowners ask whether someone can look at a panel that keeps tripping. They also want to know if that person is licensed and insured. And they want someone honest enough to say when the problem is on the utility side.

Their site today is a header, three service words, and a contact form that sometimes fails on a phone. The office manager once pasted a robots file from a forum. A post there said AI companies were copying the web. He did not know he had turned away the crawler that fetches pages for search. Then the owner paid for a rewrite. It added a file at the root and left the service page as thin as before. They were told it would help them show up, but it did not. The search crawler was still shut out, and the page still had nothing a citation could honestly repeat.

So we would split the switches the way OpenAI splits them. OAI-SearchBot has to be allowed if ChatGPT search is going to fetch the panel page. Blocking it keeps the page out of those answers. That is a fair choice if the owner does not want to be there, and a mistake if she does. GPTBot is the training crawler, and that is a separate decision. The owner can say no to training and still want search. She cannot say that with one “disallow all” line she does not understand. ChatGPT-User is the third case, because it fetches when a person asks and robots.txt may not apply. The office manager should know that rule exists. Then he will not tell the owner that a robots file controls every look-up a user can start.

The page itself has to work for a person standing in a dark kitchen. The opening lines answer the panel question. Then they say what you do on a trip and what you will not do. That includes anything on the utility side of the meter. They also say how to reach the office manager. And they say what to tell her so the right tech brings the right tester. The techs should know the description when they read it. They are the ones who will hear “but your website said” in a hallway. The owner signs off on the claims about license and insurance. The office manager signs off on the crawler rules, in writing. The three bots are named in words he can point at.

A generated answer might then mention the firm for a tripping panel in their towns. That mention is not a booking, and Pew is the reminder, even though it measured a different summary. When a summary is on the screen, a click is less common. So the paragraph that the answer cannot replace is the one about the utility line. It also covers calling the office manager with a photo of the panel label. The citation makes them curious, and the clear refusal gets them to the phone. A thin page only gets them a sentence that could have been about any electrician in the county.

How we’d run it

We would fix the robots file and the service page in the same week. We would not call a new root file the project.

First, write the three crawlers in the office manager’s notebook in plain words. OAI-SearchBot fetches pages for search, and leaving it blocked keeps you out of ChatGPT search answers. GPTBot is for training. ChatGPT-User fetches when a person asks for a look-up, and a robots file may not control it. If the office manager cannot tell them apart yet, nobody edits the file.

Second, decide the training question on purpose. Some owners want no training crawl, and that is allowed, but it is not the same lever as search. Then make the search choice out loud. Maybe the owner wants to be found in ChatGPT search. Then OAI-SearchBot has to be allowed to fetch the pages she hopes will be cited. We write the decision next to the file. That way the next freelancer does not “clean it up.”

Third, look at the Google basics on the same pages. Each page should be indexed, able to show a snippet, not blocked, and not thin. The only useful sentence should not hide behind a form. Google’s guide says the generative features are still SEO. It does not use llms.txt, and there is no special rule for chunking. No extra schema is needed. A page that fails these basics will not become an overview because a file appeared.

Fourth, rewrite the service page in the voice of the kitchen, and not in the voice of a keyword list. It should cover the panel question and the utility line. It should also say what to photograph and how booking with the office manager works. A person should be able to read it without feeling sold to. An answer should be able to lift a sentence without making up the rest.

Fifth, add a reason to click. We assume the summary took the easy meaning of a tripping breaker. So we leave the shop-specific part on the page. That means who comes, what you turn down, and what a homeowner should have ready. Pew saw a click 8% of the time when a summary showed, against 15% when none did. Showing up is the citation, and the click needs its own writing.

Sixth, test it the boring way. Ask someone who does not work there to find the page from the question. They should use a phone, and then tell you what they would do next. If they cannot say, the citation will not save you, so fix the page. After that we leave the crawler rules alone unless the owner changes her mind. A steady file matters more than a monthly tweak from whoever read a thread last.

Common mistakes

The first mistake is one switch for every bot. A copied robots file that blocks everything will block OAI-SearchBot along with GPTBot. You have then opted out of search answers and called it privacy. If privacy was the goal, that is fine. If the goal was to show up, you closed the door and then paid someone to decorate it.

The second mistake is the magic file. Google says it does not use llms.txt. A page that is blocked, thin, or not able to show a snippet will not become an overview because a new file was added. We have heard the file sold as the ChatGPT trick as well. The real trick is the right crawler plus a page worth citing. The file is just a list at best.

The third mistake is a page that only a model could love. It has fragments, repeated town names, and no sentence a tech would say out loud in a hallway. There is no special chunking rule, and headings are for people who scan. Would a tech be embarrassed to read the page to a customer? Then it is the wrong page, however well a crawler can fetch it.

The fourth mistake is mixing up a mention with a call. The July 2025 Pew finding is the warning. Visits with a summary led to a click 8% of the time, against 15% without one. We write the line the summary cannot finish. That is the edge of the work, the photo to send, and the person who answers.

The fifth mistake is making claims nobody will back. One example is “same hour” when the truck is across the county. Another is “we repair utility equipment” when you do not. ChatGPT-User may fetch the page because a person asked, and robots.txt may not apply. So a careless sentence can travel even when you thought you had a gate. We write only what the owner will say in the driveway.

The sixth mistake is handing the file to someone who cannot name the bots. The office manager should own it, or a person he trusts who will explain each change in plain words. An anonymous plugin update that “optimizes robots” is how the search crawler gets shut out again on a Friday.

What we’d do

We would allow the search crawler and make a separate, clear choice about training. Then we would spend the real effort on the panel page that a homeowner can use. We would not start with a root file. We would not let a single disallow line stand in for a decision. Our view is that most “show up in ChatGPT” projects are a robots accident. Add a thin page with a strategy name on top, and that is the whole project.

Google still wants an indexed page that a snippet can lift. It still turns down the special file and the special schema. OpenAI still uses different crawlers for search, for training, and for a fetch a person starts. Pew still says clicks get thinner when a summary is already on the screen. Those three facts are the whole job. Let the search crawler in if you want to be cited. Write the answer where the office manager, a tech, and the homeowner would all know it. Then make the next step so clear that a citation has somewhere to go.