# Case: research answers with citations

**Status: not executed.** No account or API key for any web answer engine is
available to this project, and buying one is outside this task's authorisation.
This document is the reproducible *plan*, kept separate from any claim that a
test happened. There are **no run records** for this theme: no `runs/` directory
exists here, and none is referenced anywhere.

## What is published

The public guide <https://globalaiproductindex.com/guides/source-backed-research-qa/>
publishes this method and these limits only. It carries **no measured result,
no score and no ranking** for any answer engine, because none was run. The guide
describes citation behaviour from vendor documentation, labelled as a
documentation comparison, and names the accounts that would be required before
any run could be recorded here.

This document is republished byte for byte at
<https://globalaiproductindex.com/tests/source-backed-research-qa.md>; the index
of every published record is <https://globalaiproductindex.com/tests/index.md>.
It deliberately has no sibling file under any `runs/` directory.

## The task (as it would be run)

Ask each tool the same five questions, two of which have a known correct
answer that only exists in a source published after most models' training data
(the "freshness probe"), and one of which is unanswerable (the "honesty
probe"). A tool passes the freshness probe only if it cites a source that
actually contains the fact; it passes the honesty probe only if it declines.

Inputs would be fixed in a file under `docs/tests/research-inputs.md` so every
tool sees the identical wording.

## What would have to be recorded per tool

- Product and plan used (free tier vs paid), account type, and the exact model
  or mode selected in the interface.
- The full answer text, unedited, plus every citation URL it offered.
- For each citation: does the URL load, and does the page actually contain the
  claim? (Checked with `curl` on the record date.)
- Token/call usage and the cost of the run, if the tool bills per use.
- Date and time of the run, and the region the request came from.

## Blockers, precisely

For every tool below the run was simply **not executed in this round**: no
account, API key or entitlement was used, so no answer, citation check, latency
or cost can be reported. The table states the condition each tool actually
needs, checked against this site's product records. Where the plan or
entitlement a run needs is not confirmed, it is recorded as unconfirmed rather
than assumed to be a paid plan.

| Tool | Condition, as verified |
| --- | --- |
| ChatGPT | A signed-in account. A free tier exists; a paid plan is only needed for the newer models. The pricing page (chatgpt.com/pricing) answered HTTP 200 on 2026-10-10 but renders its prices in JavaScript, so no price was read. |
| Perplexity | A signed-in account; Pro is the paid plan. perplexity.ai/pro answered HTTP 403 to this project's fetcher on 2026-10-10, so no plan price was read. |
| You.com | An API key; the free tier still requires a key and an account, and new keys come with free credits, so a paid plan is not the precondition. |
| Genspark | A signed-in account; a free plan exists. The public help-centre pricing page (genspark.ai/helpcenter/membership-plans) answered HTTP 200 on 2026-10-13, and its plan prices are recorded on the product page. |
| Microsoft Copilot | A Microsoft account. Core Copilot features are free and the paid plan is optional; the entitlement actually needed for a cited-answer run was not confirmed in this round. |
| Exa | An API key; new developers receive free credits, so a paid plan is not the precondition. |

## What can be done without those accounts

- The vendor pages themselves can be (and were) checked: see
  `docs/research/sources-2026-10-09.md` and the per-product records under
  `data/products/*.json`, each carrying `verification` with the fetch date,
  HTTP status and SHA-256 of the page actually read.
- The citation *behaviour* of a tool can be described from its documentation
  (for example, which surfaces state that they cite sources), but that is a
  documentation comparison and is labelled as such on the guide page.

## Next step for the project

Obtain one free API key for any cited-answer service, then run the five
questions and append a `runs/` record under this directory. Until then the
guide links to tools whose citation behaviour is documented, and says nothing
about how well any of them perform.