Last verified: August 12, 2026
What You Can Now Do
Context Memo now lets you choose which AI model generates your ad recommendations, and attaches source links to every recommendation the system returns. Before this, a generation run gave you three or four angles, hooks, or targeting ideas with no view into which model produced them or what informed them. Now each recommendation carries its provenance: whether it traces back to a competitor's live ad, a public benchmark, a platform documentation page, or the model's own pattern matching with no external grounding. The task that changes is triage. You stop weighing every suggestion equally and start acting on the ones with evidence behind them.
Where It Is in Context Memo
Both controls live on the ad recommendations screen. The model picker sits at the point of generation, next to the campaign context fields, so you select before you run rather than after. Source links appear on the returned recommendations themselves, one set per recommendation, expandable inline so you don't lose the run while checking a citation.
How to Use It
- Open the ad recommendations screen and enter your campaign context: product, audience, objective, and platform. This is the brief the model sees, so specificity here changes what comes back.
- Pick a model in the picker before you generate. A reasoning-heavy model tends to return more structured targeting logic. A model trained on fresher web data is more likely to reference current campaign patterns. A faster, cheaper model is the right call when you're iterating at volume and don't need depth.
- Generate. Each recommendation returns with its source links attached.
- Expand the sources on any recommendation you intend to act on. Check whether it points to external material or to the model's own pattern matching. That single check tells you how much weight the idea deserves and whether it needs validation before it reaches a media plan.
- Re-run the same brief against a second model and compare outputs side by side. Same context, different reasoning, visibly different recommendations. Keep the run you can defend.
- When you escalate a recommendation to a client or an executive, send the source links with it. The audit trail is already attached, so you're not reconstructing the reasoning after the fact.
Why We Built It
Paid media teams kept telling us the same thing: "I can't take this to a CMO when the only answer to why is the AI said so." That gap cost them twice. First in rework, when a recommendation got challenged and someone had to go rebuild the rationale by hand. Second in adoption, because teams quietly stopped using AI output for anything that required sign-off, which meant the fast part of the workflow only ever applied to low-stakes decisions. Exposing the model and the sources closes both gaps at generation time instead of at review time.
What It Does Not Do Yet
- We don't rank the models for you. The picker surfaces the tradeoff; it doesn't resolve it. Comparing outputs on your own brief is still the only reliable way to learn which model suits your category, and that comparison is a manual re-run.
- Model selection is a per-run choice. You pick at generation, not once per campaign, so each run is a deliberate decision rather than an inherited default.
- Some recommendations will return no external source. That isn't a failure state. It means the output came from the model's own pattern matching, and the absence of a link is the signal. Treat those as hypotheses to test, not as findings.
- This screen recommends; it doesn't publish. Turning a validated angle into citation-grade content that AI models can cite back to you remains a separate step in memo generation.
If you're running high-volume iteration, start with a faster model and reserve the reasoning-heavy runs for the briefs that will get escalated. That split tends to be where the source links earn their keep.