Writing RSAs With AI That Google Actually Rewards

The hard part of responsive search ads was never writing fifteen headlines — any competent writer, human or model, can fill the slots. The hard parts are writing them against the queries your account actually pays for, and getting the finished ad into the account without a copy-paste chain that dies somewhere between the doc and the interface. A connected AI fixes both in one conversation: it reads your search terms, drafts in that language, and pushes the ad through an approval step.

Why RSA copy written blind underperforms

An RSA gives you up to fifteen headlines and four descriptions, from which Google assembles combinations per auction — a few headlines and descriptions shown at a time, mixed and matched against the query (Google's RSA documentation covers the mechanics). The design assumption is variety: different angles for different searchers.

Copy written from a brand brief fails that assumption in a predictable way. All fifteen headlines end up being the same headline in different clothes — the company's favourite claim, restated. Nothing matches the searcher who typed a problem rather than a product name, because nobody writing from the brief has seen what searchers type. The gap is measurable in the account itself: compare the vocabulary of your converting search terms with the vocabulary of your live headlines, and the overlap is usually thinner than anyone expects. Closing that gap is the whole job.

Your search terms report has seen it. It is a record of the exact language people used before spending your money — including the phrasings that converted. Copy grounded there starts with an unfair advantage: it echoes the query back, which is what both the searcher and the matching system reward.

Feeding the AI your actual queries, not your assumptions

The grounding prompt, before any drafting happens:

For ad group X in account 123-456-7890, read the search terms from
the last 90 days. Show me: the top converting terms, the phrasings
people use for the problem (not our product name), and any recurring
words that signal intent — pricing, comparison, urgency. Then wait.

The "then wait" matters. You want to see the language landscape before drafts exist, because you will spot things worth encoding — a use-case you had not considered a selling point, a qualifier ("for small business", "same day") that appears in converting queries but nowhere in your current ads. Then:

Draft 15 headlines (max 30 characters each) and 4 descriptions (max
90 characters) for this ad group, using that search-term language.
Vary the angle across headlines. Flag any character-count violations
and anything the landing page cannot substantiate.

The last clause is the quiet quality gate: an ad that promises what the page does not deliver costs you twice — once at Quality Score, once at conversion.

Headline architecture: pins, keyword coverage, proof points

Fifteen varied headlines is not fifteen random headlines. A slate that serves well covers roles:

Role	Slots	What it does
Keyword coverage	4-5	Echoes the query family — the relevance signal
Proof and specifics	3-4	Numbers, credentials, named outcomes the page supports
Offer and call to action	2-3	The reason to click now
Differentiators	2-3	What the competitor's ad cannot say
Compliance or brand line	1-2	The line legal requires — pinned if it must always show

On pinning: a pinned headline always serves in its position, which buys certainty and costs combinatorial freedom. Pin what compliance or brand rules genuinely require, and nothing else. Every additional pin shrinks the space of combinations Google can test, which is the mechanism you are paying for. Where pinning is unavoidable, pin in pairs when you can — two headlines pinned to the same position rotate against each other, which keeps a sliver of testing alive inside the constraint.

Ask the model to label its draft against this table — "mark each headline's role, and flag roles with no coverage". Gaps become visible in seconds.

Descriptions: the four slots everyone wastes

Descriptions get ninety characters each and half the attention. Give the four slots roles the way the headlines have them: one that expands the strongest proof point with specifics no headline can fit, one that handles the biggest objection — price, effort, risk — one that carries the offer and its mechanics, and one that speaks to the searcher earlier in their decision. The failure mode to draft against is four descriptions restating the headlines in longer clothes. The test is deletion: if removing a description loses no information, it was not doing a job, and the model should be sent back for one that does.

From draft to live: pushing RSAs through a connector with approval

This is where the connected workflow beats the doc-and-dashboard one outright. After your edit pass:

Create this RSA in ad group X, paused. Pin headline 1 to position 1.
Final URL: [url]. Show me the complete ad before creating it.

The client surfaces the exact tool call — every headline, every description, the pin, the URL — and waits for your click. The ad lands paused, attributed to your own Google sign-in in change history. You enable it from the interface after a final look, or ask the model to enable it once you are satisfied. Nothing in this chain involves retyping, and the approval step means nothing lands that you did not read.

Paused-first is optional but cheap: an ad that cannot serve cannot spend, so the cost of the habit is one extra click and the benefit is a calm final review. It also builds the habit that scales — every ad the model ever creates for you arrives visible, reviewable and reversible, which is what makes handing it the tedious half of copywriting feel unremarkable within a week.

Ad strength: what to optimise for and what to ignore

Ad strength is a pre-serving heuristic — it grades your inputs (headline variety, keyword inclusion, description length) on a scale from Poor to Excellent. It is worth a look and not worth obedience.

Optimise for what the rating genuinely proxies: real variety and keyword coverage, which the architecture table above already produces. Ignore it when it argues with your constraints — unpinning a legally required line to chase Excellent is trading a real obligation for a label. Ad strength is not a performance verdict; conversion data is. An ad rated Average that converts is a better ad than an Excellent one that does not, and no reporting column will ever say otherwise.

A 30-day RSA refresh cadence

RSAs decay quietly — queries drift, offers age, competitors adapt. A monthly loop keeps the slate honest without turning copywriting into a hobby:

1. Read. "Which of my RSAs have the weakest asset performance labels, and how have the converting search terms shifted since these ads were written?"
2. Replace, not pile. Swap the weakest third of headlines for new ones drafted from the current search-term language. Keep what performs.
3. Log. Note what you changed and why, one line — future you will want to know whether the December dip was seasonality or the refresh.
4. Check the pins. Compliance lines age too — offers expire, legal wording changes. A stale pinned line serves in every single impression, which makes it the most expensive stale text in the account.
5. Re-ground quarterly. Monthly refreshes work from the standing language map; once a quarter, redo the full search-term read, because markets drift in ways a month-to-month diff hides.

Thirty minutes a month, most of it judgement. For turning refreshes into genuine experiments with readable results, continue to the ad copy testing framework; for the campaign type where assets do even more of the work, see what you can control in Performance Max.