ChatGPT Prompt Tracking: A Reproducible Method for Monitoring Brand Answers

ChatGPT has no single fixed answer to a buyer question. Four documented variables change what it says, and a tracking method that ignores them measures noise.

• Retrieval state. As of October 2026, ChatGPT decides per query whether to search the web. With search off, the answer comes from what the model learned in training. With search on, it draws on pages fetched at answer time.
• Model path. ChatGPT routes between model families. OpenAI's developer docs list a February 16, 2026 knowledge cutoff for the GPT-5.6 family and an August 31, 2025 cutoff for chat-latest, which OpenAI describes as the latest Instant model used in ChatGPT. That is roughly five and a half months of difference in what the model knows without search.
• Memory and instructions. OpenAI's help centre says that if memory is enabled, ChatGPT may use saved memories when rewriting a search query. Product results also consider context such as Memory or Custom instructions.
• Query rewriting. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries before sending them to search providers.

The next move: hold fixed every variable you can control, and write down the ones you can't.

Build the method on what OpenAI states in writing. Treat everything else as unknown.

OpenAI's whole ranking disclosure is one sentence: "ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed." Plan-to-model mapping comes only from secondary reports. As of September 2026 those reports said Free and Go users get gpt-5.6-luna and Plus users get gpt-5.6-sol. OpenAI has not confirmed this, so record your plan tier rather than assume a model.

Run every prompt the same way each time. The full protocol is in our reproducible prompt testing playbook. For ChatGPT specifically:

1. Freeze the wording. Write prompts the way buyers ask them. Never edit a prompt mid-series. New wording is a new prompt with its own history. Decide early whether to track prompts that name competitors.
2. Control the session. Test logged out, or in a dedicated account with memory off and custom instructions empty. Start a new chat for each prompt. Keep country and language fixed, and keep separate prompt sets per market.
3. Record the model path. Log the plan tier and whatever the model picker shows. In the API, pin a dated model ID, not an alias, because OpenAI warns alias snapshots are regularly updated.
4. Record retrieval state. Log on, off or unknown. If the answer shows no citations or sources, log unknown, not off.
5. Repeat each run. Ask each prompt more than once per session. Nothing published sets the right number, so pick one and keep it.
6. Log the answer, not a feeling. For every run, record:
   • whether your brand is named
   • each specific fact stated about you (price, product names, location)
   • which competitors are named
   • every cited URL

Then check each stated fact against your own site. A wrong price is a finding you can act on.

The consumer app picks retrieval state for you. The API lets you choose it, which is why a serious method runs both. Our comparison of API answers and consumer app answers covers where the two differ.

Search off

The answer reflects training data only. Anything you published after the model's cutoff cannot appear. On the Instant path, that cutoff is August 31, 2025. Nothing you change this quarter reaches these answers.

Search on

OpenAI's Responses API returns a sources list of every URL the model consulted, separate from the citations. OpenAI says "the number of sources is often greater than the number of citations." Log both lists. Being retrieved but not cited is a different problem from not being retrieved.

Citation is also not the same as influence. A mechanistic study (arXiv:2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources.

How often the app searches

Published estimates disagree. Profound reports about 18% of conversations trigger a web search, Semrush's US clickstream put it at 34.5% of queries in February 2026, and cloro reported 80.5% on a 100-prompt set. They measure different things on different populations. Don't assume a share; record what each run shows.

You test the prompt. ChatGPT searches with queries it writes itself. That gap is permanent from the outside, because OpenAI does not expose the rewritten queries.

The only first-party report of real grounding queries is Bing Webmaster Tools' AI Performance report, and it covers Copilot and Bing, not ChatGPT. So any ChatGPT prompt set is a sample of likely buyer questions, not a view of the actual retrieval input.

What that means for reporting:

• Report presence, absence and the facts stated, per prompt and per retrieval state.
• Never report a rank or position. No ground truth exists to check one against.
• Keep ChatGPT separate from other engines rather than blending everything into one number. See whether one dashboard can fairly cover every engine.

For a quick baseline, the free checker on this site runs up to three buyer questions and scores the answers. It is a snapshot, not continuous tracking. Our guide to free ChatGPT brand tracking shows how to extend it by hand.

If search-on answers never cite your site, check access first. These are the levers with documented mechanisms.

• OAI-SearchBot controls ChatGPT search. OpenAI says sites opted out "will not be shown in ChatGPT search answers, though can still appear as navigational links."
• GPTBot governs training only. OpenAI says each setting is independent, so blocking GPTBot does not remove you from search, and allowing it does not get you cited.
• ChatGPT-User handles user-initiated fetches, and OpenAI says robots.txt rules may not apply to it.

A minimal robots.txt rule for search eligibility:

User-agent: OAI-SearchBot
Allow: /

robots.txt is not enough on its own. OpenAI's help centre says to allow OAI-SearchBot and confirm your host or CDN allows traffic from its published searchbot IP addresses. A firewall that returns 403 removes you silently. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

Finally, put key facts in the HTML. Vercel found ChatGPT's crawlers fetch JavaScript files but don't execute them. None of this guarantees a citation. It removes the blocks you control.