RSA Google Ads + IA : génération et test rotation

On the 14-21 day A/B tests we continuously run in 2026 on accompanied accounts, a well-prompted AI-generated RSA delivers a CTR 5 to 8% higher than a pure-human RSA — but with a conversion rate 0 to 3% lower on niche B2B and premium brand ad groups. CPA equivalent to +/- 5%. The real gain isn't on pure performance but on production time: 45 min well-prompted AI vs 2-3h pure human per complete RSA. ChatGPT isn't magic on Google Ads RSAs — it's a production accelerator on standardized ad groups, and a trap if deployed without process on strategic ad groups.

This article walks through the complete 2026 workflow: prompt template per intent (4 variants), matrix quality scoring, clean ad rotation deployment despite Google's enforced 2024+ optimize constraint, 14-21 day holdout A/B test in isolated ad groups, AI vs human incrementality measurement. No hype — a structured process that delivers 75-88% usable output instead of 40-55% from a naive prompt. For pure RSA mechanics (7-theme matrix, pinning, Ad Strength), see our RSA writing method. For the AI Google Ads pillar, our article on 30 JSON Google Ads prompts. Our free CTR calculator compares your click-through rate to US 2026 medians by vertical.

RSA + AI: why it remains a human process in 2026

RSA generation by AI in 2026 is technically trivial — a frontier model produces 15 headlines and 4 descriptions in under 10 seconds — but the performance differential plays out on the structured brief, quality scoring, and in-account testing, not on the generation engine. On the aggregated 2025-2026 Google Ads data, a well-prompted AI-augmented workflow delivers 75 to 88% usable output vs 40 to 55% for a naive prompt, with production time halved at equivalent RSA. The diagram below summarizes the four steps of the AI-augmented RSA → A/B test workflow.

RSA generation via AI in 2026 is technically trivial — any frontier model (GPT-5, Claude Opus 4.7, Gemini 2.5 Pro) produces 15 headlines and 4 descriptions in under 10 seconds. The difficulty isn't generation but quality, scoring, and in-account testing. The 2026 ad copywriter doesn't disappear — their role changes: feed the AI engine with a structured brief, validate outputs, calibrate message-market matching, measure incrementality.

Three persistent illusions about AI RSAs:

• "AI writes better than a human" — false on average. On serious A/B tests, AI = +5-8% CTR but -0-3% conversion rate. Net business often neutral or slightly positive.
• "The more variants generated, the better" — false. Beyond 30 outputs, diversity caps and human review becomes the bottleneck.
• "GPT-5 is strictly superior" — false in 2026. Claude Opus 4.7 often surpasses GPT-5 on B2B RSAs (prose coherence, stakeholder-aware tone), GPT-5 is more creative on mass-market consumer angles, Gemini 2.5 Pro excels in contexts needing real-time web grounding.

What AI does well (validated in-account 2025-2026):

• Fast production of 30+ variants for matrix (time gain -65%).
• Strict character-count constraint adherence (94-99% in structured JSON).
• Multi-account brand voice consistency (agency industrialization).
• Multi-language generation from a canonical EN (local consistency).
• Differentiation angle suggestions humans haven't seen.

What AI does poorly (and requires humans):

• Calibrate the specific message-market match on niche B2B.
• Detect risky angles (legal, brand safety, off-brand tone).
• Evaluate emotional resonance vs simple formal compliance.
• Understand implicit vertical codes (luxury, health, finance, religion).
• Anticipate mobile vs desktop reading ambiguities.

Production ratio observed on mature workflows:

Official Google reference on RSAs: the RSA best-practices documentation on support.google.com and the ad rotation policy article on ad rotation. Google's recommendations converge with our tactical method: 15 headlines, 7 themes, 1 pin max.

The prompt template (4 versions per intent)

The same RSA prompt doesn't work for every ad group. The 4 most frequent intents — long-tail, brand defense, comparative, lead gen — call for 4 distinct templates. The JSON structure stays similar; the constraint content varies by intent.

Template 1 — Long-tail RSA (specific query volume):

{
  "role": "You are a Google Ads RSA copywriter, native English, long-tail expert.",
  "intent": "long_tail",
  "context": {
    "vertical": "[To fill]",
    "icp": "[Precise persona]",
    "long_tail_keywords_top_10": "[Paste top 10 SQR queries]",
    "differentiators": ["[List 3-5 differentiators]"]
  },
  "task": "Generate 30 headlines (2x the 15 finals) and 8 descriptions (2x the 4 finals).",
  "constraints": {
    "headline_max_chars": 30,
    "description_max_chars": 90,
    "theme_distribution_target": {
      "main_keyword": 6,
      "long_tail_variation": 6,
      "quantified_benefit": 4,
      "proof_point": 4,
      "direct_cta": 4,
      "differentiation": 4,
      "brand_only": 2
    },
    "include_long_tail_modifier_in_8_headlines": true,
    "no_repetition_keyword_exact": true,
    "no_external_benchmarks": true,
    "no_emojis": true,
    "no_caps_lock": true
  },
  "output_format": "JSON array: headline, theme, char_count, long_tail_modifier_used"
}

Template 2 — Brand defense RSA (competitor bidding on your brand):

{
  "role": "Brand defense PPC copywriter.",
  "intent": "brand_defense",
  "context": {
    "brand_name": "[Your brand]",
    "competitor_attacking": "[Competitor name]",
    "differentiators_vs_competitor": ["[3-5 specific strengths vs this competitor]"],
    "brand_proof_points": ["[2-3 proof points like rating, years, clients]"]
  },
  "task": "Generate defense RSA for brand exact-match ad group.",
  "constraints": {
    "include_brand_in_minimum_5_headlines": true,
    "tone": "confident without aggressive, no direct bashing",
    "implicit_comparison": true,
    "no_competitor_name_mention": true,
    "headline_max_chars": 30,
    "include_proof_points_credibility": "minimum 3 headlines"
  },
  "output_format": "JSON array: headline, theme, brand_present, char_count"
}

Template 3 — Comparative RSA (vs direct competitor):

{
  "role": "Comparative PPC copywriter (legal-aware).",
  "intent": "comparative",
  "context": {
    "your_solution": "[Your product]",
    "competitor_to_compare": "[Compared competitor]",
    "comparison_axes": ["price", "features", "support", "integrations"],
    "concrete_advantages": ["[Real quantified advantages]"]
  },
  "task": "Generate comparative RSA for 'vs Competitor' ad group capturing queries like [your brand vs competitor].",
  "constraints": {
    "headline_max_chars": 30,
    "tone": "factual evidence-based, no gratuitous superlatives",
    "no_misleading_claims": true,
    "include_minimum_3_concrete_numbers": true,
    "comparative_advantage_per_axis": "1 headline minimum per axis"
  },
  "output_format": "JSON array: headline, comparison_axis, evidence_level, char_count"
}

Template 4 — Lead gen RSA (qualification + objection handling):

{
  "role": "Lead gen copywriter, qualification focus.",
  "intent": "lead_gen",
  "context": {
    "service_offered": "[Your service]",
    "icp_target": "[Precise persona]",
    "icp_anti_target": "[Who you do NOT want to attract]",
    "common_objections": ["[3-5 typical objections]"],
    "qualification_criteria": ["[Lead qualification criteria]"]
  },
  "task": "Generate RSA for lead gen ad group with qualification objective, not volume.",
  "constraints": {
    "headline_max_chars": 30,
    "include_qualifying_signals_minimum_3_headlines": true,
    "include_objection_handling_minimum_2_descriptions": true,
    "tone": "professional, no artificial urgency",
    "no_clickbait": true,
    "exclude_terms_attracting_unqualified": "[Terms to exclude like 'free', 'no commitment' if you want paid intent]"
  },
  "output_format": "JSON array: headline, qualifying_signal, objection_handled, char_count"
}

These 4 templates cover ~80% of typical account RSA use cases. For special cases (seasonal, multi-language, regulated sector), create derived templates by adding specific constraints without changing the global JSON structure.

Quality scoring: criteria and thresholds

Quality scoring is the step that separates mature AI workflows from amateur ones. Without scoring, you take the first 15 generated headlines — outputs often technically valid but qualitatively mediocre. Matrix scoring filters outputs before human review, reducing review time by 60-70%.

6 scoring criteria (3 algorithmic + 3 human):

{
  "scoring_rubric": {
    "char_count_compliance": {
      "type": "algorithmic",
      "rule": "headline <= 30 chars AND description <= 90 chars",
      "weight": 1,
      "binary": true
    },
    "theme_tag_valid": {
      "type": "algorithmic",
      "rule": "theme tag in [keyword, benefit, proof, cta, offer, differentiation, brand]",
      "weight": 1,
      "binary": true
    },
    "no_excluded_terms": {
      "type": "algorithmic",
      "rule": "no term from excluded_terms list present",
      "weight": 1,
      "binary": true
    },
    "no_keyword_repetition_exact": {
      "type": "algorithmic",
      "rule": "keyword exact appears max 3 times across 30 headlines",
      "weight": 1,
      "binary": true
    },
    "brand_voice_match": {
      "type": "human",
      "rule": "tone aligns with brand guidelines",
      "weight": 1,
      "binary": false,
      "scale": "0-3"
    },
    "proof_credibility": {
      "type": "human",
      "rule": "proof points are credible and verifiable",
      "weight": 1,
      "binary": false,
      "scale": "0-3"
    }
  },
  "filter_threshold": "score_total >= 5/6 (algorithmic) + brand_voice >= 2 + proof_credibility >= 2"
}

Scoring workflow in practice:

# Pseudo-code AI RSA scoring pipeline
def score_rsa_outputs(outputs, scoring_rubric, brand_voice_guidelines):
    scored = []
    for output in outputs:
        score = {
            "char_count_compliance": check_chars(output),
            "theme_tag_valid": check_theme(output),
            "no_excluded_terms": check_excluded(output, excluded_list),
            "no_keyword_repetition": check_repetition(outputs, output),
        }
        # Algorithmic score 0-4
        algo_score = sum(score.values())
        if algo_score < 4:
            scored.append({"output": output, "passed": False, "reason": "algorithmic"})
            continue

        # Human review queue
        scored.append({
            "output": output,
            "passed": "pending_human_review",
            "algorithmic_score": algo_score,
            "human_criteria_to_review": ["brand_voice", "proof_credibility"]
        })
    return scored

Thresholds observed on 200 scored RSAs (aggregated Google Ads benchmarks):

• 30 AI outputs generated (2x targets), algorithmic filtering: ~25 pass (83%).
• 25 outputs in human review: ~18 pass brand_voice + proof_credibility (72% of remaining).
• 18 validated outputs, final selection of 15 per 7-theme matrix: 15 retained, 3 rejected for thematic redundancy.
• Final output ratio: 15/30 = 50% of AI generations end up in production RSAs. Normal, healthy.

Clean ad rotation: optimize vs rotate evenly

Ad rotation is the parameter that dictates how Google serves an ad group's RSAs. Before late 2024, two options were fully available: optimize (Google preferentially serves the best-performing RSAs) and rotate evenly (Google serves RSAs in balanced alternation over 90 days). Since late 2024, Google has phased out rotate evenly on the majority of accounts — only some legacy or edge-case old campaigns retain it.

What this changes for AI vs human RSA A/B tests:

• Before 2024 — you could place 2 RSAs in the same ad group, rotate evenly, and compare apples-to-apples over 90 days.
• Since late 2024 — Google forces optimize, so it's impossible to test 2 RSAs serving 50/50 in the same ad group. The RSA that "wins" the first week receives 80%+ of serving thereafter.
• Clean 2026 method — create 2 isolated ad groups: AI_only and Human_only, same keywords, same budget, same landing, same match types. The ad group becomes the A/B test unit.

Setup of isolated A/B ad groups (procedure):

# Pseudo-code Google Ads API A/B ad groups setup
def create_ab_test_ad_groups(campaign_id, keywords, landing_url, budget_per_ad_group):
    # Ad group A: AI-only
    ad_group_a = create_ad_group(
        name="RSA_AI_test_a",
        campaign_id=campaign_id,
        max_cpc_default=None  # Inherit from Smart Bidding
    )
    add_keywords(ad_group_a.id, keywords)
    add_rsa(ad_group_a.id, headlines=ai_generated_15, descriptions=ai_generated_4)

    # Ad group B: Human-only
    ad_group_b = create_ad_group(
        name="RSA_human_test_b",
        campaign_id=campaign_id,
        max_cpc_default=None
    )
    add_keywords(ad_group_b.id, keywords)
    add_rsa(ad_group_b.id, headlines=human_written_15, descriptions=human_written_4)

    # Optionally adjust ad rotation (limited 2026)
    set_ad_rotation_optimize(ad_group_a.id)
    set_ad_rotation_optimize(ad_group_b.id)

    return {"ai_group": ad_group_a, "human_group": ad_group_b}

Critical precautions for test rigor:

• Same exact-match-type keywords. No broad-match variation on one side and phrase-match on the other — immediate bias.
• Same shared budget or identical per-ad-group budgets. No asymmetric Smart Bidding learning phase.
• Same landing page URL across all RSAs. Testing a different page = another confounding variable.
• No modification during the 14-21 day test. No headlines added, no keywords adjusted, no budget changed.
• Identical geo-targeting between the 2 ad groups. Otherwise market bias.
• No different audience signal between the 2 ad groups.

Confounding variables that ruin the test:

• Different device bid adjustments (mobile vs desktop) between ad groups.
• Different scheduling (active days / hours).
• Different network targeting (Search Partners on/off).
• Different extensions (different sitelinks, callouts).

All these variables must be strictly identical between the 2 ad groups. Otherwise you're testing "AI RSA + bid +20% mobile" vs "human RSA + bid 0% mobile", which says nothing about RSA quality itself.

14-day A/B test: holdout split methodology

The clean A/B test lasts 14 days minimum, ideally 21 days, with 5,000 impressions minimum per ad group. Below that, day-to-day variance exceeds the AI vs human RSA gap, and you cut on noise. The holdout-split methodology applies the same principles as incrementality holdout tests (cf. our Discovery Ads incremental guide) — applied at the ad-group RSA scale.

Stopping criteria and result reading:

{
  "test_completion_criteria": {
    "min_duration_days": 14,
    "min_impressions_per_ad_group": 5000,
    "min_clicks_per_ad_group": 200,
    "min_conversions_per_ad_group": 10
  },
  "decision_rules": {
    "ctr_significant_improvement": "+8% relative AND p_value < 0.05",
    "conv_rate_no_significant_loss": "loss < 5% relative",
    "cpa_no_significant_loss": "loss < 8% relative"
  },
  "winner_definition": {
    "ai_wins_if": "ctr_significant_improvement AND no_significant_loss",
    "human_wins_if": "ai_does_not_meet_criteria OR conv_rate_loss > 8%",
    "tie_if": "no clear winner — choose by production time"
  }
}

Example reading on a fashion e-com ad group (aggregated Google Ads benchmarks Q1 2026):

Cases where humans win (Google Ads data, niche B2B ad groups):

• AI conv rate often -8 to -15% in niche B2B (complex message-market matching).
• Off-tone brand voice detected in human review (AI tends to flatten distinctive angles).
• AI CTR comparable to or below human on ad groups where specificity beats hook.
• Practical conclusion: on niche B2B, premium brand, top revenue ad groups, prioritize humans.

Industrialization decision matrix:

• If AI wins on CTR AND no conv rate loss AND production time -50%+: industrialize AI on similar ad groups (same vertical, same intent).
• If AI equivalent to human AND production time -50%+: industrialize AI for productivity gain.
• If AI loses on conv rate (above 5%): keep human on these ad groups.
• If tie: pick AI on standardized ad groups, human on strategic ad groups.

Measuring AI vs human incrementality

AI vs human incrementality is different from campaign vs holdout incrementality. Here, we measure not whether the ad exists or not, but whether the AI version delivers a net gain over the human version — across 3 dimensions: pure performance (CTR / conv rate), production time, brand voice quality.

The measurement happens at 3 levels:

• Pure performance — 14-21 day isolated ad group A/B test (cf. section 5). It's the most visible measure but often the least discriminating.
• Production time — strict timing of steps: brief, generation, scoring, selection, calibration. Compared on 10 RSAs produced per method.
• Brand voice quality — qualitative blind review by 3 human reviewers who don't know who wrote (AI or human). Score 0-5 on brand consistency.

Typical results from aggregated 2025-2026 Google Ads data (n=78 blind-tested RSAs):

Business reading of the results:

Well-prompted AI is neither strictly superior nor strictly inferior to humans — it shifts the production frontier. At equivalent performance (-/+5% per metric), it frees 50-60% of production time. This time gained can be reallocated to strategy (which strategic ad groups deserve pure human), tracking (Enhanced Conversions, offline), or scaling (more thematic ad groups).

The real 2026 question isn't "AI vs human" but "where to allocate human time budget":

• Standardized ad groups (mass-market e-com, volume lead gen) → well-prompted AI by default.
• Strategic ad groups (premium brand, niche B2B, top revenue) → pure human.
• Multi-language ad groups (cross-country industrialization) → well-prompted AI + local human review.
• Fast seasonal ad groups (weekly refresh) → well-prompted AI for speed.
• New product launch ad groups → pure human, AI as support.

Naive ChatGPT RSAs (without structured prompt, without scoring, without A/B) are never a recommended option. They average -5 to -12% conversion rate vs human baseline, with brand voice quality 2.1/5 and high stat hallucination risk. Apparent time gain is offset by performance losses and reputational risk.

Common mistakes (over-fitting to the prompt)

On AI RSA workflows referenced in 2025-2026, here are the 6 recurring mistakes — each one reduces real AI ROI and explains why many advertisers wrongly conclude that "AI doesn't work on Google Ads." Often, it's not AI that doesn't work — it's the workflow.

Mistake 1 — Naive prompts without structured constraints. Asking "write me 15 RSA headlines for my company" without context, without theme distribution, without character_max, without excluded_terms produces 40-55% usable output. With a structured JSON prompt, you climb to 75-88%. The gain isn't in the model but in the precision of constraints.

Mistake 2 — No quality scoring before human review. Taking the first 15 generated headlines without algorithmic filtering wastes 60-70% of human review time on outputs that don't even meet character count or theme distribution constraints. Always filter algorithmically before human review.

Mistake 3 — Over-fitting to the initial prompt. Iterating the prompt 15 times to "perfect" the output on a specific ad group produces a non-reusable prompt. The right workflow: 80% reusable generic prompt template + 20% context customization. If you iterate more than 3 times on the prompt for 1 ad group, the prompt template needs enrichment, not over-optimization for one particular case.

Mistake 4 — Testing 2 RSAs in the same ad group under Google's enforced 2024+ optimize. Since late 2024, Google forces ad rotation optimize, so 2 RSAs in the same ad group don't serve 50/50 — the first that performs in the first 7 days captures 80%+ of serving. Any intra-ad-group A/B conclusion is biased. Clean method = 2 isolated ad groups, same keywords.

Mistake 5 — Cutting the test under 14 days and 5,000 impressions. Day-to-day variance often exceeds the AI vs human RSA gap. Cutting too early = decision on noise. Strict rule: 14 days minimum, 5,000 impressions minimum per ad group, ideally 21 days and 10,000 impressions to absorb 3 full weekly cycles.

Mistake 6 — Industrializing AI on every ad group without discernment. AI is a production accelerator on standardized ad groups (mass-market e-com, volume lead gen), but it degrades strategic ad groups (premium brand, niche B2B, top revenue) where complex message-market matching prevails. Industrializing everything to AI is as naive as industrializing everything to human — 2026 sophistication is in human time allocation by ad group criticality.

To automate production pipeline deployment without building the prompt + scoring + A/B infrastructure yourself, our SteerAds audit integrates the workflow above and proposes an AI industrialization plan segmented by ad group criticality, with a pilot A/B test on 2-3 ad groups before global rollout. To go further on the AI Google Ads pillar, see our article on 30 JSON Google Ads prompts and its visual extension AI images Veo3 Flux Midjourney. AI RSA is neither magical nor useless — it's what your surrounding workflow makes of it. Without scoring, without isolated A/B, without human review, it's a trap of apparent productivity. With methodological discipline, it's the cleanest 2026 productivity lever for acquisition teams — see also official Google Ads documentation for more details.

To go further, see also our guides on AI negative keywords discovery clustering, Python API automation, Zapier Make Google Ads.

Sources

Official sources consulted for this guide:

• schema.org
• blog.google
• anthropic.com
• platform.openai.com

FAQ