Skip to main content
Pretty has two routes to the web and one to your own material, and it treats each differently.

​
The two web routes

Researching

Live web research across many sources. Returns dense factual prose plus a structured list of sources, each with a title, a URL and a last-updated date.Use it when you do not know which pages hold the answer: market sizing, competitive landscape, trends, recent financials.

Reading page

One URL that you name. Pretty fetches it, converts it to text, and pulls out the parts relevant to your question.Use it when the answer is on a page you already have: a pricing page, an agenda, a press release, a spec sheet.
You do not pick between them by name. Paste a URL and you get Reading page. Ask an open question about the world and you get Researching.

​
Asking for research

​
Asking for one page

Ask a page a sharp question, not a topic. “The full session-by-session agenda with times and speakers” gets you the agenda. “Info about the conference” gets you vague text.
Content a page renders with JavaScript after loading is not visible to the fetch. If a page comes back thin or empty, that is usually why. Try a different URL or attach a PDF export.

​
Sources end up on the slide

When Pretty drafts slide content from web research, it is required to add a small Sources: footer label listing the domains it drew on. Attribution on research-derived slides is a rule it follows, not a case-by-case decision. If you do not want the footer on a particular slide, say so, and move the attribution somewhere you do want it.

​
Grounding in your own documents instead

Often the web is the wrong source and your own material is the right one. Attach it and say so. Drag files onto the pane (it shows Drop to reference in chat) or use Upload files. Several at once is fine, and so is a .zip.
Without that last paragraph, Pretty may reach for the web to fill gaps. Say explicitly when you want a closed-world answer.
Typing or pasting a file path does not attach anything. Pretty cannot reach your filesystem. The upload is the only route in.

​
Asking for citations, and checking them

Pretty will attribute what it drew from research, but it will not footnote every line unless you ask. When precision matters, ask for it:
Then check two things before the deck leaves your hands:
  • Open the sources. The research result carries real URLs and last-updated dates. A source from four years ago behind a claim about “current” market share is a problem the model will not flag for you.
  • Match every number to a source. If a figure has no source next to it and did not come from your own attached file, treat it as unverified until you find where it came from.
Ask Pretty to be explicit about gaps: “If the research does not support a claim, write ‘not found’ on the slide rather than filling it in.” An obvious hole is easier to fix than a confident guess.

​
Research turns take longer

A Researching row runs many searches and reads many pages before it returns anything, so a turn that includes research is noticeably slower than one that just builds a slide. Nothing is stuck. If you want to move faster, do the research in one turn, check what came back, then ask for the slides in the next.