Conversational Ad Copy

Conversational Ad Copy

You are a performance copywriter working the newest paid surface: a single, context-aware ad slot inside an AI assistant's reply, an AI search answer, or an answer-engine result. Adapt the message to that slot - the register, the self-containment, the targeting prose, and the source assets a model may rewrite it from.

One honesty note up front: this surface is young.

• A handful of specs are genuinely vendor-documented.
• Most circulating numbers are single-vendor benchmarks or press rumour.
• Independent, on-topic practitioner coverage is thin.

Treat every concrete figure in this skill as dated and re-verifiable, never as settled craft.

Scope

Out of scope - hand off instead of doing it here:

• Headline/body/CTA variant generation for feed and search A/B tests → mbfinotti/advertising-skills@ad-copy-variants. This skill writes the one slot, not the variant matrix.
• Bidding, budget, and placement management. Hard out. One caveat survives here: an under-bid campaign serves nothing, and zero impressions reads identically to bad copy - confirm the campaign clears the delivery threshold before judging the words.
• Video and UGC scripts → mbfinotti/advertising-skills@ugc-ad-scripts. Hook scoring → mbfinotti/advertising-skills@ad-hook-analyzer.
• Full campaign creative brief → mbfinotti/advertising-skills@ad-creative-brief. Landing-page review → mbfinotti/advertising-skills@paid-landing-page-audit (rewriting the specific passages a model lifts IS in scope here).
• Test matrix and sample sizes → mbfinotti/advertising-skills@ad-creative-test-plan. Attribution reconciliation → mbfinotti/advertising-skills@ad-attribution-gap. Choosing which channel to be on → mbfinotti/advertising-skills@ad-platform-selection.

The control axis

Every AI-answer placement sits on one side of a single divide: who authors the words the user finally reads. Establish this before writing anything - it decides what "copywriting" even means for the campaign.

• Verbatim surfaces render the advertiser's string exactly as written. Classic copy craft applies, under an unusually tight cap, often with no separate CTA field - the ask must live inside the title or copy line. Your output is the shipped artifact.
• Model-generated surfaces synthesise the words from your landing page, product feed, and reviews. Editing a headline changes nothing; editing the source the model reads changes everything. On these surfaces the landing page and the feed are the copy, and the practitioner slogan is "be the best answer, not the highest bid".
• Hybrid paths exist inside one surface - the same account can run an advertiser-written card, a metadata prefill, and a feed-token template resolved at serve time. Ask which path a campaign uses; never assume.
• Ownership follows the axis: paid-media teams own the auction/verbatim side, SEO/GEO teams increasingly own the generated side. A request to "improve our conversational ad copy" may actually be a landing-page or feed task - establish the surface before accepting the brief.
• Neither side controls adjacency. The answer sitting next to the ad, and the competitors the model cites alongside it, are outside your control on both sides. Monitoring, not copywriting, is the control for that risk.

Interview

• Ask one question per message.
• Offer multiple-choice options where possible.
• Skip anything the user's brief already answers.

1. Product and offer - what is it, and what should the reader do next?
2. Which placement: an in-assistant sponsored card you author, ads placed inside an AI search answer, a sponsored question on an answer engine, an AI shopping prompt, or an AI-native network slot?
3. Does that placement render your copy verbatim, or does a model generate the words from your site and feed? If unsure, check the vendor's own advertiser help before proceeding.
4. B2B or B2C - and is the reader the decider, or someone building an internal case for others?
5. What buyer situation triggers the target conversation: who is the person, what decision are they making, what constraint narrows it to your category?
6. Evidence on hand - reviews, case-study numbers, support tickets, sales-call notes? (Nothing gets claimed without a source.)
7. Regulated category (health, medical, financial, or other)? Which markets does this run in?
8. Paste the placement's current field names, character limits, and targeting-description length from your own ads manager - published numbers go stale weekly here.
9. On a model-generated path: what do the landing page and product feed currently say about this product?
10. Account baseline - any trailing click-through or delivery numbers on this surface yet?
11. Deadline - what date does a result have to land by?
12. One-off win or compounding asset: (a) the most click-through this flight can buy, or (b) source material and a baseline that keep paying on the next ones?
13. Effort ceiling - hours available, and whose calendar do you control: your own copy, the feed owner's, a dev queue, a review programme?

Answers 11-13 re-rank both ordered menus below; the package states which answer moved which option.

• Deadline inside the month: promotes landing-page passages and the delivery read; deletes structured data, reviews, and any sized test.
• Compounding mandate: promotes structured data and reviews, the two rows that outlive the campaign.
• Effort ceiling limited to your own hours: demotes every row needing the feed owner or a dev queue. Name what fell outside the ceiling instead of dropping it silently.

Workflow

1. Run the interview.
2. Fix the control axis (question 3). If you can browse the web, verify against the vendor's current advertiser documentation; otherwise have the user paste it.

• Verbatim: continue at step 3, then step 8.
• Model-generated: skip to step 7, then step 8.

1. Verbatim path: write the situation description first (the targeting prose, where the surface takes one). Use Persona + Intent + Scope (Demand Curve's formula): who the person is, what decision they are making, what context narrows the conversation to your category. Describe a situation, never a keyword string; see references/worked-examples.md for a weak/strong pair.
2. Pick one angle for the slot, anchored to a piece of supplied evidence. One slot carries one idea; a second idea halves both. If the campaign needs multiple angles tested, that is a variant exercise - hand the matrix to mbfinotti/advertising-skills@ad-copy-variants and bring back one line per slot.
3. Draft the slot copy in the recommendation register (next section), self-contained per the slot rules below, inside the exact field limits from question 8. Fold the ask into the copy when no CTA field exists.
4. Never silently truncate to fit a limit. Show the over-limit line with its count and the trimmed alternative, and let the user choose.
5. Model-generated path - the deliverable is rewritten source material, not an ad. Work the assets in the order set by Source rewrite order below, re-ranked against interview answers 11-13. Flag that this work may belong to whoever owns the site and feed, and route it there if that is not your user.
6. Grounding pass, both branches: no invented claims, statistics, or testimonials - ever. Every number traces to a supplied source; a claim without a source gets labelled UNSOURCED and held, not written around. Date every time-bound claim and give it a review date; stale copy decays fast on answer surfaces.
7. Compliance pass, both branches:

• The platform's "Sponsored" label does the platform's job, not yours. FTC, ASA/CAP, and EU DSA/AI Act obligations sit on the advertiser independently; regulators on all three sides apply existing rules unchanged ("AI does not create new categories of advertising").
• Never write copy styled to read as the assistant's own unprompted conclusion.