How to Write Ad Copy with AI for Google and Meta (2026)

In 2026, Google and Meta do not run the ad you typed. They run the ad their AI assembles from your assets. Google mixes RSA headlines into new combinations, drops them into descriptions, and generates new lines from your landing page through text customization. Meta's Text Improvements can lift a phrase from primary text and turn it into a headline, overlay, or footer. The operator skill is no longer copywriting one perfect 3-line ad. It is writing a modular, constrained, remix-proof asset pack and watching what actually serves.

The short answer

Write assets, not ads. Build a facts vault (offer, proof, banned claims, landing pages), then generate modular lines that each make sense alone. In Google, pin any legal or offer-critical line to Headline 1, Headline 2, or Description 1, and use text guidelines in AI Max or Performance Max to stop Google from generating forbidden words. In Meta, put the offer in the first 125 characters of primary text and keep every claim self-contained, because Text Improvements can reorder or extract phrases.

Score every asset with a 0-100 rubric covering click promise, specificity, length fit, combinatorial safety, claim-to-landing match, and brand plus policy fit. Ship only at 80 or above, with 100% human review on P0 assets (brand, legal, price). After launch, watch the combination report in Google and the Advantage+ Creative previews in Meta. Remove machine-generated lines that are off-brand, off-claim, or off-landing page.

The ILAA Ad Copy Generator gives you free, no-sign-up headline and description variants built to each platform's limits. Use it for the first pass, then run the scorecard and constrain steps by hand.

Why 2026 ads are remix machines

Three platform shifts changed the job description:

1. Responsive Search Ads are the only Search text format. You no longer write expanded text ads. You upload up to 15 headlines and 4 descriptions. Google assembles them, sometimes bolds matching terms, and may serve unused headlines as link-based assets or at the start of a description. Google decides the order, not you. This means any two headlines in your pack could appear together.
2. Performance Max and AI Max write their own lines. Text customization (formerly automatically created assets) is on by default in Performance Max and is a core feature of AI Max for Search. Google generates extra headlines and descriptions from your landing page content, domain, and existing ads, and refreshes them at least every 48 hours. It may also swap your final URL for a different page and generate matching copy.
3. Meta Advantage+ Creative can rewrite your text. Text Improvements is a generative AI enhancement that reorders primary text and headlines, swaps positions, and places phrases as overlays or footers. The creative you see in the preview is not guaranteed to be the creative a user sees.

The upshot: your ad is now a pool of candidate lines. If the pool contains even one line that contradicts another, reads as generic on its own, or makes a claim your landing page cannot support, the platform will eventually serve that bad combination to someone.

Platform map: current names and limits

Use the current 2026 names. Do not say DALL·E, GPT-3.5, Bard, or Claude 2 as current stack. Do not say "Advantage+ Shopping Campaigns" as the current Meta sales product name (the current name is Advantage+ Sales Campaigns). Do not say "automatically created assets" as the current Google UI name; the current name is text customization.

All limits apply to single-width characters. Double-width languages (Korean, Japanese, Chinese) count each character as 2. For Google, these are hard field limits. For Meta, the numbers above are recommended visible lengths; the field accepts more, but users may not see it. Always preview by placement in Ads Manager.

For official Google specs, see About responsive search ads, Best practices for creating effective responsive search ads, and About text assets for Performance Max campaigns. For Meta, see Advantage+ Creative.

The SOURCE → LOCK → SCORE → WATCH workflow

This is the operating system. It works for a single campaign or fifty. Each step has an owner and a gate.

The workflow is deliberately boring. The discipline is the constraint layer and the scorecard. An LLM will write cute copy all day. The operator's job is to stop it from writing a cute lie.

SOURCE: build a facts vault before you prompt

Most bad AI ads come from one mistake: the prompt contains no facts. The model invents a price, a guarantee, a feature name, or a timeline. The platform then serves that invention. Your first artifact is a facts vault, not a prompt.

For every campaign, capture:

• Offer: exactly what the user gets and what it costs. ("14-day free trial, paid plans from $49/mo, no card required.")
• Proof: claims you can defend. ("Used by 2,000 Shopify stores" only if true. If not, use a feature proof point.)
• Banned claims: words or promises that are legally risky or off-brand. ("guaranteed ROI," "instant," "unlimited" if not literally unlimited.)
• Must-show lines: disclaimers, required offer text, brand name constraints.
• Landing URL(s): the exact page each ad sends to, so every claim can be verified.
• Voice card: tone descriptors, sentence length, words to avoid, words to prefer. For a deeper reusable system, see the sibling Aug 19 post on building a reusable AI brand voice guide.

If your landing page content is weak, text customization and Final URL expansion will have poor source material. Clean your landing page first. The SEO meta titles and descriptions at scale workflow covers the page side of this; this post covers the ad side. Do not reuse metas as ad assets. They serve different jobs.

Generate modular assets, not finished ads

A finished 3-line ad is fragile. A pool of modular lines is robust. Write each line so it can stand next to any other line in the pack.

Generate across these families:

• Keyword family: the exact search term or pain point. ("Shopify helpdesk," "AI ticket routing.")
• Benefit family: the outcome. ("Cut first-response time," "Stop bouncing between tabs.")
• Proof family: the reason to believe. ("Built for Shopify," "Auto-tags and assigns.")
• CTA family: the next step. ("Start free trial," "See pricing," "Watch demo.")

Vary the lengths. Google specifically recommends trying headlines of different lengths so the assembler has room to fit different formats. Do not write 15 headlines that are all 29 characters. Mix 12-character punch lines with 28-character descriptive lines.

Never put a negative claim next to a positive one in a way that could combine into nonsense. If headline A says "No credit card" and headline B says "Plans from $49/mo," the assembly could read "No credit card. Plans from $49/mo." That is fine if true. But if headline A says "Free forever" and headline B says "From $49/mo," the assembly becomes a contradiction.

Use the ILAA Ad Copy Generator to produce the first pool. It targets Google and Meta limits and gives you several variants to score. Then use the AI Rewriter for length-tightening and the Grammar Checker for a final polish.

The HarborDesk asset pack (copy-ready example)

HarborDesk is a fictional AI ticket-routing helpdesk for Shopify stores. Use this pack as a template. Replace the facts with yours.

Facts vault

• Product: AI helpdesk that reads, tags, and routes Shopify support tickets.
• Offer: 14-day free trial; paid plans from $49/mo; no credit card to start.
• Proof: built for Shopify; auto-tagging; round-robin assignment.
• Banned: guaranteed ROI, instant, unlimited, 24/7 if not literally true.
• Landing: pricing page and signup page.
• Voice: direct, slightly informal, no hype.

Google Responsive Search Ad pack

Headlines (15, all ≤30 chars):

1. Shopify Helpdesk Software
2. AI Ticket Routing
3. Auto-Tag Support Tickets
4. Stop Tab-Hopping Support
5. Built for Shopify Stores
6. First Response, Faster
7. Route Tickets to the Right Agent
8. 14-Day Free Trial
9. From $49 Per Month
10. No Credit Card Needed
11. See Plans and Pricing
12. Support Teams Love It
13. Less Scrolling, More Solving
14. Try HarborDesk Free
15. AI That Reads Tickets

Descriptions (4, all ≤90 chars):

1. HarborDesk auto-tags and routes Shopify support tickets to the right agent. Start a 14-day free trial with no credit card.
2. Built for Shopify stores. Cut first-response time with AI ticket routing and clear queue assignments.
3. From $49/mo after your free trial. No hidden setup fees. See pricing and pick the plan that fits your team.
4. Support teams stop bouncing between inbox, chat, and docs. HarborDesk keeps every ticket in one clear workflow.

Paths: shopify-helpdesk / ai-routing

Combinatorial safety check. "14-Day Free Trial" + "From $49 Per Month" is true. "No Credit Card Needed" + "From $49 Per Month" is true. "Shopify Helpdesk Software" + "Stop Tab-Hopping Support" is coherent. Nothing contradicts the facts vault.

Performance Max text asset pack

Headlines (same 15 as RSA, plus one short):

16. Try Free

Long headlines (5, ≤90 chars):

1. HarborDesk: AI ticket routing for Shopify support teams. Start a 14-day free trial today.
2. Auto-tag and route Shopify support tickets to the right agent with HarborDesk.
3. Built for Shopify stores. Cut first-response time with AI that reads every ticket.
4. Support teams use HarborDesk to stop bouncing between tools and start closing tickets.
5. From $49/mo after a free trial. No credit card required. See plans and pricing.

Descriptions (5, ≤90 chars):

1. AI ticket routing and auto-tagging for Shopify support teams. Start your 14-day free trial.
2. HarborDesk assigns tickets to the right agent, so first-response times drop and queues stay clear.
3. Built for Shopify. No credit card to start. Choose a plan from $49/mo after the trial.
4. Stop bouncing between inbox, chat, and docs. One workflow for every Shopify support ticket.
5. See pricing and try HarborDesk free for 14 days. Setup takes minutes, not hours.

Business name: HarborDesk (10 chars)

Meta ad text pack

Primary text variants (offer in first 125 chars):

1. HarborDesk auto-tags and routes Shopify support tickets to the right agent. Start a 14-day free trial with no credit card. Plans from $49/mo.
2. Stop bouncing between inbox, chat, and docs. HarborDesk gives Shopify support teams one clear workflow. Free trial, no card required.
3. Built for Shopify stores. HarborDesk reads every ticket, adds the right tags, and assigns it to the best agent. Try it free for 14 days.
4. From $49/mo after a free trial. HarborDesk helps Shopify support teams cut first-response time with AI ticket routing.

Headlines (3, ≤40 chars recommended):

1. AI Ticket Routing for Shopify
2. Stop Tab-Hopping Support
3. Try HarborDesk Free

Link descriptions (3, ≤25 chars recommended):

1. Start free trial
2. See pricing
3. Built for Shopify

This pack is your template. Every line is under limit, true to the facts vault, and self-contained enough to survive reordering.

LOCK: pin, exclude, restrict, and front-load

Constraints are what separate an asset pack from an AI word salad. Use three layers.

1. Pin must-show lines in Google RSA

If a line must appear in every ad (legal disclaimer, required offer, brand name), pin it. Google lets you pin to Headline 1, Headline 2, or Description 1. Those positions are guaranteed. Headline 3 and Description 2 are not. Pinning reduces the number of combinations Google can test, so only pin what is legally or commercially required.

Google's guidance is specific: "If you pin one headline or description, pinning isn't recommended for most advertisers and can affect ad strength." A hybrid approach works best: pin one or two essential lines and let the rest rotate. If you need variation within a pinned slot, pin 2-3 unique lines to that slot.

2. Use text guidelines in Performance Max and AI Max

Text guidelines (currently in beta) let you add term exclusions and messaging restrictions to Google-generated assets. You can exclude up to 25 terms and add up to 40 messaging restrictions per campaign. This applies to text customization assets, not your uploaded assets.

Example exclusions for HarborDesk:

• guaranteed
• instant
• unlimited
• 24/7

Example messaging restrictions:

• Do not mention prices except the stated plans from $49/mo.
• Keep tone direct and slightly informal. Avoid hype.
• Always spell the brand name HarborDesk as one word.

Be careful: new guidelines can stop existing Google-generated assets from serving if they fail the new rules. Start with the smallest set that protects brand and legal risk, then expand.

3. Front-load Meta primary text

Meta's Text Improvements can take keywords and phrases from your original copy and move them to overlays, footers, or prominent headlines. If your most important claim lives at the end of a 300-character block, it may never be seen. Put the offer and any must-show claim in the first 125 characters. Keep every claim self-contained so extraction does not create a false statement.

Bad: "Support teams have a lot to juggle. HarborDesk makes it easier with AI ticket routing and a free trial." If Meta extracts "Support teams have a lot to juggle" as a headline, you have wasted the placement.

Better: "HarborDesk auto-tags and routes Shopify tickets. Start a 14-day free trial with no credit card." The extracted phrase carries the offer.

SCORE: the Ad Asset Scorecard

Score every asset before it ships. The scorecard is platform-agnostic. A line that scores 80 on Google should also score 80 on Meta.

Ship threshold: 80. Any asset below 80 is rewritten or removed. P0 assets (price, legal, brand) require a human check regardless of score.

Scoring example for "Stop Tab-Hopping Support" (headline, 25 chars):

• Click promise: 18/20 (curiosity + pain point, but no offer)
• Specificity: 10/15 (Shopify support teams will get it; others may not)
• Length: 15/15 (fits with room)
• Combinatorial safety: 20/20 (works with any other true headline)
• Claim + landing: 12/15 (landing page should explain tab-hopping)
• Brand + policy: 15/15 (no risk)

Total: 90. Ship.

Scoring example for "AI-Powered Support Revolution" (headline, 30 chars):

• Click promise: 8/20 (vague)
• Specificity: 3/15 (no product, audience, or outcome)
• Length: 15/15 (fits)
• Combinatorial safety: 14/20 (not false, but adds nothing)
• Claim + landing: 5/15 (meaningless claim)
• Brand + policy: 10/15 (hype risk)

Total: 55. Rewrite.

WATCH: reports, previews, and prune

Your work does not end at launch. The platform will write new lines. You have to catch them.

Google

• Combination report: shows which headline + description combinations actually served. Look for pairs that contradict or read poorly together.
• Asset-level report: shows performance per headline and description. Pause low performers after enough data.
• "Automatically created" source: in Performance Max asset reporting, this label means Google wrote the line via text customization. Review these at least weekly.
• Asset details report: for AI Max / text customization in Search, this is where off-brand generated lines show up.

Meta

• Advantage+ Creative previews: check the generated variants before broad spend.
• Text Improvements variants: if available in your account, review the reordered/extracted text. Regulated industries (housing, employment, credit, pharma, social issues, politics) may have limited or no access to generative enhancements.
• Placement previews: mobile feed, Stories, Reels, and desktop right column can truncate differently.

Rule: any machine-generated line that fails the scorecard is removed immediately, even if performance looks good. A winning ad that makes a false claim is a compliance lawsuit waiting to happen.

Worked before/after rewrites

These show the difference between generic AI output and remix-safe assets. The "before" lines sound fine in isolation. The "after" lines survive platform assembly.

Google RSA headline rewrite

Before (generic): "AI-Powered Support Software"

After (specific + self-contained): "AI Ticket Routing for Shopify"

Why: The after line names the feature, the platform, and the audience. It works as H1, H2, H3, or as a link-style asset. The before line says nothing.

Performance Max long headline rewrite

Before: "Transform Your Customer Support with the Power of Artificial Intelligence"

After: "HarborDesk: AI ticket routing for Shopify support teams. Start a 14-day free trial today."

Why: The after line fits 90 characters, contains the brand, the audience, the offer, and a true trial claim. The before line is 88 characters of vague promise.

Meta primary text rewrite

Before: "In today's fast-paced ecommerce landscape, support teams need every advantage they can get. HarborDesk leverages cutting-edge AI to streamline your workflow and deliver exceptional customer experiences."

After: "HarborDesk auto-tags and routes Shopify support tickets to the right agent. Start a 14-day free trial with no credit card. Plans from $49/mo."

Why: The after line front-loads the offer, is under 125 visible characters for the core promise, and contains only verifiable claims. The before line is robotic, vague, and buries the offer.

Failure gallery: 8 bad AI ad patterns

These are real failure modes we see when teams generate ad copy with AI and ship it without the scorecard.

1. Feature soup. "AI-Powered, Cloud-Based, Scalable, Omni-Channel Support Solution." Too many modifiers, no benefit, no audience.
2. Phantom numbers. "Save 10 Hours Per Week" when the landing page says nothing about time saved. Fabricated proof points are a policy risk.
3. Contradictory pair. "Free Forever" and "From $49/mo" in the same RSA. Google will eventually show them together.
4. Buried offer. Meta primary text that waits 200 characters to mention the trial. Text Improvements may never extract it.
5. Generic CTA. "Learn More" on every asset. Specific CTAs such as "See pricing," "Start free trial," and "Watch demo" outperform generic ones, per Google's own best-practice guidance.
6. Brand drift. The AI writes "Harbor Desk" as two words or uses a competitor's tagline structure. Score brand fit on every line.
7. Legal loose end. "No hidden fees" without the landing page spelling out the fee structure. A claim is only as safe as its proof.
8. SEO meta theft. Copying organic meta titles and pasting them into RSA headline slots. Meta titles are written for search intent, not commercial intent, and often exceed 60 characters. Write separate assets.

Each of these fails the scorecard on combinatorial safety, claim-to-landing match, or brand fit. The fix is not better prompting. It is the facts vault plus the scorecard.

Prompt pack and spreadsheet

Copy-paste generation prompt

Use this prompt with ChatGPT, Claude, or Gemini. It embeds the facts vault and forces modular output.

You are a paid-ad asset writer. Use only the facts below. Do not invent prices, guarantees, percentages, or features.

FACTS VAULT:
- Product: [name]
- Audience: [specific audience]
- Offer: [exact offer]
- Price: [exact price or range]
- Proof points: [list]
- Banned words/phrases: [list]
- Must-show lines: [list]
- Brand voice: [3 descriptors]
- Landing URL: [URL]

TASK:
Write a Google RSA asset pack and a Meta ad text pack.

GOOGLE RSA:
- 15 headlines, each 30 characters or fewer. Mix short (10-15 chars), medium (20-25), and long (26-30).
- 4 descriptions, each 90 characters or fewer.
- 2 display paths, 15 characters each.
- Every headline must make sense alone and next to any other headline.
- No headline may contradict another headline.

META:
- 4 primary text variants. Put the offer in the first 125 characters.
- 3 headlines, ~40 characters recommended visible limit.
- 3 link descriptions, ~25 characters recommended visible limit.
- Each line must be self-contained if extracted for an overlay or footer.

OUTPUT as a markdown table per section. Then list any risky combinations you see.

Spreadsheet column spec

One row per asset line:

campaign | platform | asset_type | asset_text | char_count | family | score_click | score_specific | score_length | score_combo | score_claim | score_brand | total_score | qa_status | final_status | notes

Filter by total_score < 80 to find rewrites. Filter by family to make sure you have enough keyword, benefit, proof, and CTA coverage. Sort by asset_type to upload cleanly into Google Ads Editor or Meta Ads Manager bulk sheets.

Tool stack 2026

For more free tools across writing, marketing, and data, see the 20 free AI tools you can use without signing up list and the best free AI writing tools in 2026 guide.

Common mistakes that kill AI ad copy

• Writing one finished ad. Platforms need pools. Write modular lines.
• Skipping the facts vault. AI will invent numbers and promises if you do not give it real ones.
• Letting legal claims float unpinned. If a price, disclaimer, or guarantee must appear, pin it in Google or front-load it in Meta.
• Trusting default-on text customization. Google will generate lines from your landing page. Monitor the "Automatically created" source weekly.
• Ignoring combinatorial safety. Every headline must work with every other headline.
• Reusing SEO metas as ad assets. Different job, different length, different intent. Use the meta titles workflow for pages and this workflow for ads.
• One generic CTA everywhere. Match the CTA to the landing page: pricing page gets "See pricing," signup page gets "Start free trial."
• Stopping at launch. The platform keeps remixing. You keep watching.

How this fits your broader marketing stack

This post is the paid-ad layer. It connects to, but does not replace, other ILAA workflows:

• How to Write SEO Meta Titles and Descriptions at Scale for the organic landing-page side.
• How to Rewrite Content Faster with an AI Rewriter for voice-preserving rewrites of landing-page copy.
• How to Write Direct Answer Blocks That AI Search Engines Quote for making landing pages citable, which also feeds text customization.
• Free AI Tools for Freelancers if you are running ads for clients.

The sibling Aug 19 post on building a reusable AI brand voice guide covers the full voice system. This post uses a one-page voice card only. Link to the sibling guide when it ships; do not try to write the whole system here.

FAQ

How do you write ad copy with AI for Google and Meta in 2026?

Build a facts vault, then generate modular assets (headlines, descriptions, primary text) that each make sense alone. Lock legal and offer-critical lines by pinning in Google, using text guidelines in Google AI Max, and front-loading the offer in Meta. Score every asset against a 0-100 rubric, ship only at 80 or higher, and review platform remix reports weekly to remove bad machine-generated lines.

What are Google Responsive Search Ads headline and description limits?

Google Responsive Search Ads allow up to 15 headlines of 30 characters each and 4 descriptions of 90 characters each. You must provide at least 3 headlines and 2 descriptions. Path fields support up to 15 characters each. Double-width languages such as Korean, Japanese, and Chinese count each character as 2.

Should I pin RSA headlines to specific positions?

Pin only when a line must appear in every ad, such as a disclaimer or a required offer. Pinning limits the combinations Google can test and can lower Ad Strength. If you pin, prefer 2-3 unique lines per pinned slot and leave the rest unpinned for rotation.

What is Google text customization?

Text customization is a Google Ads feature, formerly called automatically created assets, that generates extra headlines and descriptions from your landing page, domain, existing ads, and keywords. It is on by default in Performance Max and available in AI Max for Search campaigns. Google will only serve these generated assets if they are predicted to perform better than the assets you uploaded.

How do I write Meta copy that survives Text Improvements?

Put your strongest offer and any required claim in the first 125 characters of primary text, because Text Improvements can reorder, swap, or extract phrases for overlays and prominent headlines. Keep each line self-contained. Avoid claims that become false when a single phrase is lifted out of context.

Can I use my SEO meta titles and descriptions as Google ad assets?

Do not reuse meta titles and descriptions as ad assets. They have different jobs. SERP snippets answer search intent; ad assets answer commercial intent in a restricted character limit and must survive platform remix. Use a dedicated ad asset pack instead.

What is AI Max for Search and when does it matter for ad copy?

AI Max for Search is Google Ads' next-generation Search automation layer, out of beta in 2026. It adds search term matching, improved text customization, and final URL expansion to eligible campaigns. For ad copy it matters because text customization and final URL expansion can generate new lines and send traffic to different pages. You need a facts vault, exclusions, and active monitoring.

How many assets should I generate for a new campaign?

For a Google Responsive Search Ad, generate at least 8-12 unique headlines and 3-4 descriptions. For Performance Max, also add 2-5 long headlines plus the required business name and display paths. For Meta, prepare 3-5 primary text variants and 2-3 headlines and descriptions per creative concept. Never ship the minimum without a human review if legal or price claims are involved.

Ship remix-proof assets this week

The ad platforms are not going to stop remixing your copy. Their whole point is to match the right message to the right query and placement. Your job is to give them a clean, constrained pool of lines so the remixes stay true.

Start with one campaign. Build the facts vault. Generate the asset pack. Lock the must-show lines. Score every asset. Launch and watch the reports. Most teams skip the scorecard and the weekly review. That is where the edge is.

Ready to start? Open the ILAA Ad Copy Generator. Free, no sign-up, no card. Describe your product, audience, and offer, get platform-aware headline and description variants in seconds, then run them through the scorecard.

Google and Meta product details checked August 2026 against official documentation. Meta help pages returned errors during direct fetch; visible-length numbers are cited from search-attributed help guidance and should be re-verified in Ads Manager before publish QA.