Agents move into marketing—and need clearer limits

Perplexity connects the campaign workflow

In a November 3 post, Perplexity described how Computer can coordinate Shopify analysis and landing pages, Meta advertising and Klaviyo email campaigns. Teams can supply brand guidelines and specify reviews before publishing, sending or spending. This is a provider-described workflow, not independent evidence of better campaign performance or a newly introduced ad placement. ¹

For a merchant, the useful starting point is one collection: identify relevant products, prepare a landing page and draft matching campaign materials for review. Check that the offer, inventory and destination agree across all three channels. Account access, permissions and actual execution still need validation; the announcement alone does not establish eligibility for every business or market.

Zalando plans purchases across merchant boundaries

Zalando’s November 3 results announcement introduced Zalando Extended: products from other shops that Zalando does not sell itself appearing on its platform. Westwing is the first pilot partner, with an AI agent planned to check out its home-and-living products inside the Zalando app from 2027. ²

That is a prospective distribution model, not general checkout access available to any merchant today. The announcement does not settle partner eligibility, technical requirements or transaction responsibilities.

A home-goods retailer can prepare by checking whether dimensions, delivery limitations and returns are clear and current. Joining a channel still requires its commercial and operational terms. Zalando also describes retail-media opportunities; paid promotion should not be confused with organic recommendation or guaranteed inclusion.

Fresh web evidence becomes part of API decisions

Perplexity’s November 4 Decisions API update adds web grounding: searches supply evidence before the model returns probabilities for a question. The response includes the search results. Crucially, if a search fails or times out, the model can still answer using whatever evidence arrived. ³

That makes the evidence check as important as the answer. An illustrative use for a local-service team is checking an organizer’s current event page before proposing an appointment tied to that event. A probability is not a booking confirmation. Build an explicit “insufficient evidence” outcome when sources are missing or contradictory, and confirm availability through the responsible service before committing.

Agent safety needs more than a passing test

OpenAI’s November 3 research compared six models’ performance on a warning-respect evaluation with approximately matched decisions in deidentified Codex production logs. Success was lower in the evaluation—not higher—across those models. The comparison is approximate and limited to sampled contexts; it does not establish that production agents are universally safer. ⁴

Perplexity’s November 2 BrowseSafe++ research approaches a different problem: malicious instructions embedded in web pages. Its benchmark primarily preserves real observed attacks inside rewritten surrounding pages. It found that detecting an injection did not guarantee resisting it during a task, and that attackers revising their attempts exposed weaknesses missed by one-shot tests. These are provider-led findings, not independent certification of a business workflow. ⁵

For a Realtor’s inquiry assistant, a useful controlled test is a page that tries to redirect the agent into sending a message or changing an appointment without authorization. Record whether it stops, explains the boundary and offers a legitimate next step. Test the action and its permission together.

What to do next

Act: For one campaign, document which steps may be drafted automatically and which require approval to publish, send or spend.

Test: Run one authorized sandbox scenario with missing evidence and one with an instruction planted in retrieved content. Keep the resulting action log.

Watch: Look for Zalando’s partner requirements and demonstrated 2027 checkout, rather than treating the announcement as an available integration.