conversational-ads

name: conversational-adsdescription: >  Plan, write and measure ads inside AI assistants and AI search (ChatGPT Ads,  Google AI Overviews and AI Mode, Microsoft Copilot). Use when testing ChatGPT  ads, answer-adjacent copy, or incrementality for AI placements.license: MIT + Commons Clausemetadata:   version: 1.0.0  author: borghei  category: marketing  domain: paid-media  updated: 2026-09-21  tags: [chatgpt-ads, ai-overviews-ads, ai-mode, copilot-ads, conversational-ads, incrementality, paid-media, ad-copy]

Conversational Ads

Paid placements next to AI answers: ChatGPT Ads (test began February 2026, self-serve Ads Manager since May 2026), Google ads in AI Overviews and AI Mode, and ads in Microsoft Copilot. This skill covers what is different from search, which platforms your category and market can use, how to write copy that sits next to an answer without borrowing its authority, how to get feeds and landing pages ready, and how to prove incrementality when two of the three platforms give no placement-level reporting.

Platform facts are as of September 2026 and come from each platform's own help pages — see references/platform-specs.md. These products change monthly; re-check the linked pages before committing budget.

---

When to use this skill

---

Clarify First

Before planning, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.

• Category and markets/language: sensitive categories are excluded on ChatGPT and AI Overviews, and AI Overviews ads are English-only in 12 countries.
• Monthly budget and target CPA: decides how many platforms can be tested with a readable holdout (≥50 conversions per arm).
• Real customer prompts: 20-50 questions people ask assistants, tagged research / compare / purchase / support.
• Measurement readiness: server-side conversions and the ability to hold out regions; without these the test cannot be read.

Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.

---

Quick start

# 1. Plan: eligibility, mix, holdout test, KPI targets, readiness (exit 2 = blocked)
python3 scripts/conversational_ad_planner.py assets/sample_plan_input.json
python3 scripts/conversational_ad_planner.py my_brief.json --format json --min-conversions 100

# 2. Lint copy before upload (exit 2 = errors; --strict fails on warnings too)
python3 scripts/answer_adjacent_copy_linter.py assets/sample_ad_copy.json

# 3. Fill assets/test-plan-template.md and freeze it before launch

The shipped samples fail on purpose: the planner exits 2 because the Copilot arm is underpowered at USD 12,000/month and a USD 60 CPA; the linter exits 2 with 10 errors across three broken ads (assistant-endorsement copy, unsubstantiated "clinically proven", over-length Google headline, unverified testimonial, http landing URL). Two ads pass clean.

---

How conversational placements differ from search [RECOMMENDED]

Position: optimise inputs (feed, assets, landing pages, negatives) and read outcomes with holdout experiments. Do not try to optimise a placement you cannot see.

---

Platform snapshot (as of September 2026)

Perplexity: 2024 ads experiment announced; current availability unverified — excluded from the planner.

---

Workflows

Workflow: first test plan

1. Collect prompts and tag each with a stage and a weight (relative volume or value).
2. Write the brief (assets/sample_plan_input.json shows every field).
3. Run the planner. Resolve BLOCKERS first: drop a platform, raise budget, extend weeks, or install tracking.
4. Resolve READINESS items marked MISSING — landing pages that answer the prompt and a clean feed matter more than bids. [RECOMMENDED]
5. Copy the plan into assets/test-plan-template.md, pick treated and holdout regions, and set the freeze window.

Workflow: copy for answer-adjacent placements

1. One prompt cluster → 5-10 variations, each with a different angle (spec, price, delivery, fit, compatibility). OpenAI recommends many diverse variations.
2. Lead with a checkable fact; end with a specific soft CTA ("Compare widths").
3. Run the linter. Fix every ERROR; review WARNINGs — ChatGPT length ranges are recommendations, not limits.
4. Record evidence for any claim you keep in substantiated_claims.

Workflow: measurement

1. Server-side conversions on each platform, deduplicated.
2. UTMs with utm_medium=cpc so paid clicks stay out of GA4's organic AI Assistant channel.
3. Geo holdout (10-20% of matched regions) or time-based on/off if regions are too few. [RECOMMENDED]
4. Two-week learning period, then freeze. Read at week 6: incremental conversions, incremental CPA, brand-search lift.
5. Decide: scale if incremental CPA ≤ target; iterate inputs if within 30%; stop otherwise.

For Google and Microsoft, the holdout measures the campaign change that made you eligible (e.g. moving to broad match / AI Max / PMax), not the AI placement alone. State that in the readout.

---

Exit code contract [PROVEN]

---

Anti-Patterns

Borrowing The Assistant's Voice

Mistake: "ChatGPT's top pick", "Recommended by Copilot", copy styled to look like the answer. Why it happens: The answer carries trust and teams want some of it. Instead: Ads are labelled Sponsored and separated from answers by design, and OpenAI's ad policies prohibit false endorsements. Lead with a fact the answer lacks. The linter blocks this pattern.

Reading Platform ROAS As Incrementality

Mistake: Scaling on platform-reported conversions in week 3. Why it happens: It is the only number available, and Google/Microsoft do not isolate AI placements. Instead: Build the holdout before launch and decide on incremental CPA.

Budgeting A Placement You Cannot Buy

Mistake: A media-plan line called "AI Overviews" with its own budget. Why it happens: Plans expect one line per placement. Instead: On Google and Microsoft, fund the underlying Search broad / AI Max / Shopping / PMax campaigns and test the change that made you eligible.

Spreading A Small Budget Across Every Assistant

Mistake: A few thousand dollars split three ways for a four-week test. Why it happens: Fear of missing the next channel. Instead: Concentrate until each arm can reach ~50 conversions. The planner exits 2 on underpowered arms.

Feed Neglect

Mistake: Polishing creative while product titles lack the attributes people ask about and prices lag the site. Why it happens: Feeds belong to another team. Instead: Treat the feed as creative. Audit the top SKUs against real prompts and use delta feeds for price and availability.

More in references/anti-patterns.md.

---

Troubleshooting

---

Scripts

Both: Python 3.8+ standard library only, --format text|json, deterministic.

Scope & Limitations

In Scope:

• Paid placements next to AI answers: ChatGPT Ads, Google ads in AI Overviews and AI Mode, and ads in Microsoft Copilot
• Test planning with scripts/conversational_ad_planner.py: platform eligibility, placement mix, holdout design, KPI targets, readiness checks, and blockers for underpowered arms
• Copy linting with scripts/answer_adjacent_copy_linter.py: length, unsubstantiated claims, assistant-endorsement wording, restricted categories, CTA, URL and UTM, price, testimonial and AI-media checks
• Measurement with server-side conversions, UTMs and geo or time-based holdouts, plus a test-plan template

Out of Scope:

• Earning organic citations in AI answers (see marketing/aeo/ and marketing/ai-seo/)
• Labelling AI-generated creative beyond a lint warning (see marketing/ai-content-disclosure/)
• Classic search, social and display campaigns (see marketing/paid-ads/)
• Perplexity ads, whose current availability is unverified and which the planner excludes
• Buying media, uploading ads or reading platform reports; the tools have no platform API access

Important Caveats:

• Platform facts are as of September 2026 and these products change monthly. Eligibility, excluded categories, markets and copy lengths need re-checking on the linked platform pages before budget is committed.
• Google and Microsoft give no placement-level reporting for AI surfaces, so a holdout measures the campaign change that made you eligible, not the AI placement alone.
• The planner's power check uses your budget and target CPA with a default of 50 conversions per arm. It estimates readability; it does not guarantee a significant result.
• The linter's category and claim checks use deliberately conservative patterns. A clean lint is not approval under platform ad policies or advertising law.

References

• platform-specs.md — ChatGPT Ads, Google AI Overviews / AI Mode, Microsoft Copilot, Perplexity status, GA4 AI Assistant channel; official links
• conversational-ads-playbook.md — differences from search, planning sequence, channel roles, creative rules, test designs, KPIs, risks
• anti-patterns.md — extended anti-pattern catalogue

Assets

• assets/sample_plan_input.json — planning brief (exits 2: underpowered arm)
• assets/sample_ad_copy.json — five ads, two clean and three broken (exits 2)
• assets/test-plan-template.md — hypothesis, prompts, eligibility, design, power, KPIs, readiness, readout