# Do your AI-powered ads match your facts?

Prepare campaign copy from the same sources, review every checkable claim, and test the
follow-up questions a prospect will ask after the click.

**Preview.** Local preparation only: a brief, reviewed copy, follow-up tests and an export.
Nothing here submits an ad, creates a campaign or spends money. You need an organization
with your product facts in it — [Connect Your Agent](https://docs.senso.ai/docs/connect-your-agent)
covers that.

## The loop

An ad is a claim. The conversation after the click is where it gets tested. Product owns
the facts, marketing owns the message, a reviewer approves what goes out. All three
contribute to the same context layer.

1. **Build the brief.** Objective, offer, market, audience, language, landing page and the
   approved facts.
2. **Draft the creative.** Generate variants from your sources, or paste copy someone
   wrote. Which of the two it is stays visible.
3. **Review the claims.** Every checkable claim in the copy and on the landing page,
   classified against sources. *Requires a human.*
4. **Test the follow-up questions.** Fit, eligibility, limitations, returns, next steps.
   Each output kept as it came back.
5. **Resolve and retest.** An owner supplies the fact, then the same questions run again.
   Keep the runs that did not improve.
6. **Approve and hand off.** Approval attaches to an exact revision. Changing copy, language
   or a source returns it to review. *Requires a human.*

Ask your agent: "Draft three ad variants for this offer from our approved sources, then list
every claim you cannot support."

Your agent needs the [Senso skills](https://docs.senso.ai/docs/agent-skills) to answer that —
`senso setup` installs them. There is no CLI command for the brief or the claim review
yet; both are local preparation, which is what the Preview note above means.

## What goes in the brief

Everything downstream is checked against this record.

- **Objective and offer** — what the campaign is for and which product it promotes. No ad
  account needed.
- **Landing page and version** — the page is part of the claim surface. An omission there
  is a finding too.
- **Approved facts** — the sources backing the copy, with their material conditions attached.
- **Voice and restrictions** — brand voice, and anything you are not permitted to claim.

## How a claim is classified

Review runs over the ad copy and the landing page together, because a prospect reads them
as one statement.

- **Supported** — a selected source says this, including its conditions.
- **Conflicting** — a selected source says otherwise. The copy is wrong, or the source is.
- **Unsupported** — nothing retrieved backs it. It may be true; it is not yet defensible.
- **Missing condition** — true only under a qualifier the copy omits. Where most ad copy
  fails.
- **Outdated** — the backing source has been superseded. Fix the source first.

## Sample walkthrough: Juniper Home

Sample — no live model call. Juniper Home is a fictional fixture and its facts are
synthetic. The copy below is illustrative and is never shown under a provider's name.

1. **The facts.** A dining table seats four closed and six extended. The extension leaf is
   sold separately.
2. **The draft ad.** "Seats six, everything included."
3. **Two findings, not one.** The ad makes an unsupported inclusion claim. The landing page
   omits the same condition. A copy fix and a page fix, different owners.
4. **The correction.** "Seats four, or six with the separately sold extension leaf."
5. **The follow-up.** "Can I seat six without buying anything else?" The expected behavior
   is to say the leaf is an additional purchase.
6. **The question nobody can answer yet.** "Will it fit with the chairs pulled out?" No
   source holds dimensions. The right outcome is a request for information, never an
   invented measurement.
7. **The multiplayer beat.** The product owner adds approved dimensions. The next agent
   retrieves that contribution and reruns the question.

## What the export contains

- The selected copy, at the exact revision approved.
- The destination it was approved for.
- Every material condition attached to a claim in it.
- Source references and versions, so a claim can be traced back.
- Who reviewed it, and when. What is still unresolved.
- The follow-up test results, with each provider and setting recorded.

## Limits

- Paid placement and independent answers are separate. Buying a placement is not a way to
  change what a model says on its own.
- Supported by your sources is not approved by an ad platform, and never becomes it.
- Factual support is one review. Brand and tone are a person's judgment, not a score.
  Platform policy is the platform's.
- A Senso follow-up test is a controlled model call, not a live sponsored conversation.
- Better factual support is not measured campaign lift. Report paid results only from real
  campaign data.
- The export is a Senso package, not an asserted import format for any ad platform.

---

- [Answers, checked the same way](https://docs.senso.ai/docs/use-cases/verification-loop)