How to track your brand in ChatGPT

A ChatGPT answer that draws on the web carries more than its text. It can name your brand, link one of your pages as a source, put that link first or last, and name your competitors beside you. Each is a separate signal, and a tracker that records only one of them misreads the others.

OpenAI’s documentation of its web search tool shows why the difference matters. An answer carries inline citations, each with the address and title of the source and the place in the text where it was used; a separate list holds every address the model consulted, and the number of sources consulted is often greater than the number cited^[1]. A page can inform an answer without being shown in it.

Over many prompts, these signals add up to your AI search visibility in ChatGPT. An AI citation is its most concrete unit, because it is the one signal that can send a visit.

Tracking a site that ChatGPT search cannot read measures a robots.txt rule, not your content. OpenAI uses OAI-SearchBot to surface sites in ChatGPT’s search features. A site that shuts it out is not shown in ChatGPT search answers, though it can still appear as a navigational link, and a change to robots.txt takes about 24 hours to reach OpenAI’s systems^[2].

Look for a group for OAI-SearchBot, or the catch-all group it falls back to, that disallows the pages you want cited. Such a rule is less common than many assume: in the October 2026 edition of QueryRev Research’s State of AI crawler access study, 2.7% of the 6,294 top sites read shut OAI-SearchBot out by name, against 9.2% for GPTBot, OpenAI’s training crawler^[3]. Blocking GPTBot does not take a site out of ChatGPT search; how each of OpenAI’s crawlers reads robots.txt is the subject of its own guide.

An llms.txt file does not change any of this: it is a reading list, not a permission, as the guide to the file explains.

Your own logs show a side of ChatGPT that no prompt set can. OpenAI names three user agents that matter here. OAI-SearchBot crawls for ChatGPT search, and GPTBot crawls for training. ChatGPT-User is different: it may visit a page when a person asks ChatGPT or a custom GPT a question, and it does not crawl the web on its own^[2]. A request from ChatGPT-User is a sign that someone’s conversation needed one of your pages.

# Pages ChatGPT fetched for a person’s question, most requested firstgrep "ChatGPT-User/" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20 # Search crawling by status code, with the robots.txt checks left outgrep "OAI-SearchBot/" access.log | grep -v "robots.txt" | awk '{print $9}' | sort | uniq -c

1. 1Count ChatGPT-User requests by page each month, beside the tracking sheet. A page fetched often for questions your prompt set never asks points to prompts worth adding.
2. 2Read the status code of each request. A 200 means the page was served; a 403 or a challenge page means a firewall or CDN turned the agent away, whatever robots.txt says.
3. 3Leave robots.txt checks out of the crawl count. OpenAI says its crawlers may add a robots.txt marker to the user agent when they fetch that file, so those requests can be told apart from the rest^[2].
4. 4Check that a request is really OpenAI’s. Each agent has its own published list of IP addresses, such as searchbot.json for OAI-SearchBot and chatgpt-user.json for ChatGPT-User^[2]; a request that only borrows the name is not evidence of anything.

Logs do not show the answer, and they miss every answer built from pages ChatGPT had read before. What they show is which pages are being read on demand, and the prompt set then helps you explain why.

The prompt set is the tracker. Every figure you report later is a share of it, so write prompts the way a customer would type or say them, not the way a keyword list phrases them, and cover each moment at which someone might meet your brand in a ChatGPT conversation.

Tag each prompt with its kind and its search intent, and keep your brand out of the category and problem prompts: a prompt that names you measures what a model remembers, not whether it finds you. How to collect the prompts, freeze the set and record a baseline before any work starts is set out in the guide to measuring results in AI answers.

A tool can do the asking, through ChatGPT itself or through OpenAI’s API, and the two are different products. The API’s web search tool returns each citation as an annotation, with its position in the text^[1], which makes placement easy to record. It also lets the caller set an approximate user location, as a country, city, region or time zone^[1], so a prompt can be asked as if from the market you sell in.

Its answers may differ from the ones a person sees in the ChatGPT app. Whichever you use, say in the report which one produced the figures, keep the location fixed, and never mix the two in one series: a change of method looks exactly like a change in the answers.

Monthly is enough for most brands: it matches the reporting cycle and leaves time for a change to show. Run the brand-fact prompts again after you change a price, a policy or the range, and run the whole set again after any change to robots.txt or to a page the set depends on.

Record each answer as one row: the date, the prompt, the run, the answer’s text and sources, and four scores. The definitions below are one workable way to make each score something two people would record the same way.

date,prompt,kind,run,named,cited_url,placement,competitors,wrong_fact2026-10-05,P07,category,1,yes,https://shop.example/mugs,2,Other Shop,2026-10-05,P07,category,2,yes,,3,Other Shop;Third Shop,2026-10-05,P12,brand facts,1,yes,https://shop.example/shipping,1,,no shipping to Norway2026-10-05,P19,comparison,1,no,,,Other Shop;Third Shop,

From these rows come the shares you report: the share of answers that name you, the share that cite you, and the same figures for each competitor. Which of them deserve to be KPIs, and how much a share can move by chance, is the subject of the guide to AI search visibility metrics.

Being named is not always good news. An answer can name you with an old price, a closed branch, a product you no longer sell or a policy you changed last year. Record every wrong statement in a column of its own, and treat it as a defect to fix, not a score to average.

1. 1Open the sources cited next to the wrong statement. The error often comes from one page, yours or another site’s.
2. 2If it is your page, correct it first, in plain sentences, where a reader would look for the fact.
3. 3If it is another site, such as a directory, a review site or an old press article, ask its owner for a correction, and make sure your own page states the fact plainly enough to be the better source.
4. 4Keep the prompt in the set and check it on the next run. The fix is done when the answer changes, not when the page does.

Visits that follow an answer arrive in Google Analytics 4 in a channel of their own, read as set out in the metrics guide; how to count the answers that send no visit is part of the measurement guide.