A Capture Standard for AI Answer Screenshots
Capture the whole answer, not the flattering sentence. Record the prompt verbatim, the assistant and surface, whether you were signed in and personalised, the model shown if any, web search on or off, date, time and timezone, and your location. Save the full answer text next to the image, because text can be searched and re-counted and pixels can't, and name both files so they sort by date and prompt. Then never present a screenshot alone: pair it with the rate from repeated runs of the same prompt. A screenshot shows that an answer happened once. The capture standard makes it checkable, and the rate makes it representative.
Screenshots are the currency of AI visibility conversations. A client sends one showing a competitor recommended; a salesperson sends one showing you first. Both are true, and both are single draws from a system that answers differently from run to run.
What to capture
| Field | Example (illustrative) | Why |
|---|---|---|
| Prompt, verbatim | "best invoicing tool for a 5-person agency" | A paraphrased prompt can't be re-run |
| Assistant and surface | Assistant A, iOS app | Apps, web and API are different products |
| Signed in, personalisation | Signed out | A personalised answer isn't what others see |
| Model or mode shown | As displayed, or 'not shown' | Ties the answer to a version |
| Web search | On, sources displayed | Changes what can be cited |
| Date, time, timezone | 2026-09-28 10:14 IST | Places it against platform changes |
| Location and VPN | Bengaluru, no VPN | Search results can be localised |
| Full answer text | Saved as .txt beside the image | Searchable, countable, quotable |
| Whole answer in frame | Scrolling capture if needed | A crop can hide the caveat two lines down |
A file name that sorts
{date}_{time}{tz}_{assistant}-{surface}_{promptId}_run{n}.png
{date}_{time}{tz}_{assistant}-{surface}_{promptId}_run{n}.txt
2026-09-28_1014IST_assistantA-ios_P07_run2.png (illustrative)
2026-09-28_1014IST_assistantA-ios_P07_run2.txtThe prompt id and run number do the real work. They let anyone find every other run of the same question, which is exactly what a sceptical reader will ask for.
Pair every screenshot with a rate
In a report, a screenshot illustrates a measured rate; it isn't the measurement. Caption it with the context: 'Run 2 of 5 on 28 September; named in 3 of 5.' If it isn't typical, say so: 'the only run of 5 where we appeared'. Choosing the best of five without saying so is the screenshot version of dropping failed runs from a denominator, and it's found out the same way, when someone asks for the other four.
What a screenshot can't show
Which sources were retrieved but not cited. What another user saw at the same moment. Whether it'll happen again tomorrow. Those questions need repeated, stored runs, and the capture standard exists so that each screenshot can sit inside that record rather than stand in for it.
FAQ
Should I capture signed in or signed out? — Match the state your buyer is likely in, and record which you used. Settings differ by product: ChatGPT's August 27, 2026 release notes, for example, say its temporary chats are not personalised by default and don't use memory, custom instructions or plugins.
Is a share link better than a screenshot? — Keep one where the assistant offers it, but don't rely on it alone. You don't control whether the link keeps working. The local image and text are the record you own.
How many screenshots belong in a report? — One or two, as examples of a rate the report states. A report made mostly of screenshots is showing its best draws.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.