Availability checked November 4, 2026: ChatGPT collection is available through ProtoScrape in early access for US English desktop requests. It returns answer text and any citations present in the collected response. An answer without citations is not evidence of a web search; record that distinction in your reports.
ChatGPT visibility research starts with a precise question: when someone asks about a problem your product solves, does the answer mention your brand, and does it point to your website? These are separate outcomes. Recording them separately gives a marketing or SEO team evidence it can inspect rather than a single unexplained “AI ranking.”
OpenAI’s web search documentation explains that search results and citations appear in ChatGPT when it uses web search. A study should therefore distinguish a search-enabled observation from an answer without observed search activity. A missing citation alone does not establish why the model produced an answer.
Choose prompts around customer decisions
Build a small prompt set around the decisions your audience makes. For a project-management product, a useful set might include a category question, a comparison and a task-specific question. For example: “What should a small agency consider when choosing project-management software?”
Keep explicitly branded questions in their own group. Asking “What is ExampleBrand?” makes a mention much more likely by construction than asking an unbranded category question. Combining the two groups can make your visibility appear stronger without reflecting a change in independent discovery.
Give each prompt a stable identifier. Save the exact wording, its intent group and the date you added it. When wording changes, create a new version so a report can distinguish prompt changes from changes in the observed answers.
Preserve the context of each observation
For an authorized observation, retain the prompt, collection time, answer text and available source URLs. Record the observable mode, market and conversation context where those details are known. Treat a follow-up in an existing conversation as a different experiment from the same question in a fresh conversation.
Do not fill missing metadata with assumptions. If the exact model or mode is not exposed, record it as unknown. An analysis is easier to trust when its limits are visible in the dataset.
Score mentions and citations separately
Use three fields in your analysis:
- Brand mentioned: the answer names the brand or a reviewed alias.
- Owned domain cited: at least one returned source points to a domain you own.
- Relevant description: a reviewer checks whether the surrounding text describes the product accurately.
A mention may be negative, incidental or confused with another company. A citation may support a general statement without recommending your product. Keep a short evidence excerpt and the original answer so a reviewer can verify the classification.
Do not turn citation order into a search position. The first source link in an answer is not evidence that the page ranks first in an organic search result list.
Report coverage alongside visibility
Suppose an illustrative run schedules ten observations, completes eight and finds your brand in three complete answers. Its completion rate is 80%, and its mention rate among completed observations is 37.5%. Show both figures and the counts behind them. Treating the two failed observations as confirmed absences would mix collection reliability with brand visibility.
Compare the same prompt cohort over time and retain individual observations. A small sample supports a description of what was observed; it does not establish what every ChatGPT user sees or prove that a content change caused a later mention.
Where ProtoScrape fits today
ProtoScrape provides a job API, account workspace and API keys. Submit a ChatGPT observation with engine set to chatgpt, retain the returned job ID, then retrieve its result after the job succeeds. Each validated successful ChatGPT result uses eight credits. Failed jobs release their credit reservation.
curl https://api.protoscrape.com/v1/jobs \
-H "Authorization: Bearer $PROTOSCRAPE_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: chatgpt-visibility-001" \
-d '{"engine":"chatgpt","query":"What should a small agency consider when choosing project-management software?","country":"US"}'
Poll GET /v1/jobs/{job_id} and, after success, read GET /v1/jobs/{job_id}/result. Keep answer_text, citations and the collection timestamp with your prompt cohort. Citation coverage varies; ProtoScrape does not guarantee that every answer invokes search or contains sources.
Check the live engine availability endpoint before scheduling collection. Use a fresh idempotency key for a new observation, and reuse the original key when retrying the same submission.
Read the API documentation, then use the AI visibility measurement guide to define your report. ProtoScrape is independent and is not affiliated with OpenAI.