Guide
How to keep forecasts and sources apart when AI reads a research report
Keep forecasts and sources apart by labeling every estimate as a forecast right next to the number, attributing each data point to its source in the same sentence or table row, and separating your own analysis from cited data. Then publish a version where those labels and references are part of the data, so AI summaries cannot quietly drop them.
How AI blurs forecasts, facts and sources
A research report mixes three kinds of statements: measured data from third parties, your own estimates and projections, and interpretation. In the layout these are clear: a shaded forecast column, a source line under a chart, a methodology footnote.
When an assistant summarizes the report, layout is the first thing lost. A 2030 market estimate becomes "the market is", a footnoted source vanishes, and a scenario turns into a prediction. The summary looks confident and is wrong in ways that are hard to spot.
Write so the label travels with the number
- Put the status in the sentence: "forecast", "estimate", "base case", "survey of n respondents".
- Mark forecast columns in the header ("2027F") rather than only with shading.
- Attribute in-line: "according to [source], [year]" next to the figure, not only in endnotes.
- Give each chart a source line and a small data table.
- Separate sections for data, methodology, and the authors’ view.
- Keep ranges as ranges; do not let "up to" or "between" collapse into a single point.
A labeling scheme that AI can follow
| Statement type | Label | Example |
|---|---|---|
| Measured fact | Source and date | Shipments of 1.2M units in 2025 (source: agency report, 2026) |
| Your estimate | Estimate + method | Estimated 2025 share of 14% (bottom-up model) |
| Forecast | Forecast + scenario + horizon | Base-case forecast: 9% CAGR 2026–2030 |
| Opinion | Authors’ view | We expect consolidation among smaller vendors |
Test the report before you publish
- Ask an assistant: "List every forecast in this report with its horizon and scenario."
- Ask: "Which figures come from third-party sources, and which?"
- Ask for the source page of three key numbers.
- Check whether any forecast is described as a fact, or any range as a point.
How The Papers keeps them apart
The Papers verifies a report before publishing its AI Version: every number, date and period must match the source, no figure may appear that is not there, and qualifiers that change meaning (forecast, unaudited, pro forma, "up to") stay attached to the value. Each value keeps a reference to its source page or slide, and native charts in PowerPoint keep their series data.
The JSON version stores metrics with period, currency and classification, plus qualifiers and claims, so an agent can tell a projection from a measurement. Where the report itself is ambiguous, you confirm and the answer is recorded as owner-confirmed. Published at a permanent link, the report can be updated without breaking citations, and analytics show which AI agents read it.
FAQ
- Does this work for scanned reports?
- The Papers uses OCR as a fallback for scanned pages, but a text-based export gives better results and keeps tables intact.
- Can AI cite the page a figure came from?
- It can when the version it reads contains that reference. In The Papers each value keeps its source page, slide or cell.
- Should paid research be public?
- Not necessarily. Use unlisted (link only, never indexed) or private (Data Room members) visibility for client deliverables.
- Does fidelity check check third-party data?
- No. It checks that the AI version matches your report. The accuracy of cited sources remains with the authors.