Responsive Search Ads: Feed the Combination Machine

Responsive search ads are not one ad format among several anymore. Since June 30, 2022, expanded text ads can no longer be created or edited, which makes the RSA the only standard search ad you can actually build. That changes what writing a search ad means. You are not composing a fixed message, you are supplying parts to an assembly system - up to 15 headlines and 4 descriptions that Google recombines per query - and the craft has moved from writing the ad to designing the parts so that every assembly is one you would have approved.

Most RSA advice I see treats this as a compliance exercise: fill the slots, chase the Ad Strength rating, move on. I think that gets the format backwards. This guide covers how the assembly actually works by Google’s own documentation, what Ad Strength does and does not tell you, when pinning is worth its cost, and the part of the job the machine still cannot do.

How the assembly works

The mechanics, straight from Google’s RSA documentation: you provide up to 15 headlines of 30 characters and up to 4 descriptions of 90 characters. At serving time, Google Ads “assembles the text into multiple ad combinations in a way that avoids redundancy,” shows up to 3 headlines at once, and then “will test the most promising ad combinations, and learn which combinations are the most relevant for different queries.”

The sentence I want you to sit with is the last one. Different queries get different assemblies. The same RSA can lead with your price headline for a bargain-hunting query and your credibility headline for a comparison query, which is the entire argument for the format. It is also the constraint that shapes all the writing advice below: any headline can appear next to any other headline, in any position, so the set has to work as a set.

My mental model is a bench of players, not a finished team sheet. I am picking who is available. Google picks the lineup per match. Complaining that the wrong headline showed for a query is usually a confession that the wrong headline was on the bench.

Writing the fifteen, in practice

The mistake I see most is fifteen rephrasings of the same sentence. Google’s assembly avoids literal redundancy, but it cannot manufacture variety you did not supply, and its own best-practice guidance is to add as many genuinely unique headlines and descriptions as you can so the system has real combinations to test.

My working recipe for a fifteen-headline set, which I treat as slots to fill rather than a word count to hit:

• Three to four keyword headlines. The query mirrored back, close variants included. These earn relevance and tend to show in position one.
• Three to four benefit headlines. What the customer gets, each one a different benefit, not the same benefit reworded.
• Two to three proof headlines. The number, the guarantee, the named credential. Anything here must survive the same substantiation bar as any ad claim.
• Two to three offer and action headlines. The price, the trial, the deadline that is real, the verb.
• One or two odd ones. A question, an objection answered. These are the lottery tickets the format exists to test.

Every headline gets one further check: read it next to each of the others and confirm no pairing is nonsense or an accidental overclaim. “Free trial” next to “From $99/month” is a combination the machine will happily serve. Drafting variants at this volume is exactly the work I hand to a model with my ad copy prompts, with the character limits in the prompt, but the pairing check stays a human job because the model does not know which of your claims can legally sit together.

Descriptions follow the same logic at four slots: I write them as two pairs, one pair carrying the offer and one carrying proof, each able to stand alone.

What Ad Strength is worth

Ad Strength rates your asset mix from Incomplete through Poor, Average and Good to Excellent, and Google’s documentation attaches a number to caring about it: advertisers who improve Ad Strength for their responsive search ads and sitelinks from Poor to Excellent see 15% more conversions on average.

I use the rating, and I also keep it in its place. Two things are worth being clear-eyed about. First, that figure is the platform grading its own homework - an average across advertisers who did many things differently, not a controlled promise to your account. I treat it the way I treat any platform’s claim about its own product: plausible direction, unverified magnitude. Second, Ad Strength rates the input mix, not the outcome. It estimates whether you gave the system enough relevant, diverse material to work with. It does not know your margins, your compliance constraints or your brand voice.

So my rule: fix Incomplete and Poor immediately, because those ratings mean the machine is genuinely starved. Get to Good. Treat the last step to Excellent as optional, and never buy it by adding headlines you do not actually want shown. A conversion report beats a grade. Google’s own account-level recommendation is at least 2 RSAs with Good or Excellent Ad Strength per ad group, which doubles as a testing structure: two sets of parts, and the ad-level results tell you which set of raw material wins.

Pinning: the control tax

Pinning locks a headline or description to a position, and it is the one lever that pushes back against the whole design of the format. Google’s documentation is unusually blunt about the cost: pinning reduces the number of combinations that can be matched to a search, is not recommended for most advertisers, and can affect Ad Strength.

I pin anyway in exactly two situations. Regulated claims, where a disclaimer or a required phrase must appear every time the ad shows - there the combination machine is a compliance risk, and the tax is just the price of running ads at all. And brand mandates, where the client requires the trademark line in position one. In both cases I soften the cost the way the documentation suggests: pin two or three interchangeable variants to the position rather than a single asset, so the position is constrained but not frozen.

What I do not do is pin to restore the feeling of authorship. Pinning everything turns an RSA back into an expanded text ad, except now it is an expanded text ad you assembled through a worse interface, with a Poor rating attached. If the machine keeps surfacing a headline you hate, the honest fix is to remove that headline, not to pin around it.

Reading the reports without fooling yourself

Two reports matter, and each answers a narrower question than it appears to. Asset reporting shows per-asset performance statistics - impressions and the rest - which mostly reflect how often Google chose to serve an asset. An asset with few impressions is not necessarily bad copy. It lost auditions, possibly for queries you do not even want. The combinations report shows which assemblies actually served, and I read it for one thing only: whether the top combinations are messages I would have written on purpose. If the most-served combination is incoherent, that is a parts problem, and I go back to the fifteen.

What I refuse to do is micro-manage removals off a week of asset data. Small samples, uneven serving and query mixes make per-asset judgments slower than they look - the same patience discipline as judging a Smart Bidding change, where the evaluation window gets agreed before the change is made. I review asset reports monthly, replace the clear losers with new candidates, and otherwise let the system run.

Where the RSA fits in the automation stack

An RSA is one third of a machine. The bid strategy decides how much to pay, the ad decides what to say, and the two interact: auction-time bidding can only exploit a query-message match your asset set makes possible. That is why I stopped thinking of ad copy as the creative garnish on a bidding setup - thin, samey headlines starve the bidder of the variety it prices. The same input discipline applies one level up in Performance Max, where asset groups play the RSA role across every Google surface at once, and the guardrails have to exist before spend starts.

The through-line of everything I write about paid media: the platforms automated the assembly and kept the receipts, so the marketer’s leverage now lives entirely in the inputs and the audit. For RSAs that means the claims you allow into the asset set, the pairings you rule out, and the monthly look at what the machine actually built with your words.

FAQ

How many headlines should I actually write - is fewer ever better? I fill all fifteen when I have fifteen genuinely different things to say, and stop earlier when I do not. A padded set is worse than a short one, because padding is where accidental duplicates and weak claims hide. Google’s guidance is “as many unique headlines as you can,” and the operative word is unique.

Should I still run more than one RSA per ad group? Yes - Google’s stated best practice is at least two with Good or Excellent Ad Strength. I use the second slot to test a different message strategy, not a shuffled copy of the first. Two ads built from the same fifteen headlines test nothing.

Does Ad Strength affect how my ads rank in the auction? Ad Strength is a pre-serving input check, a rating of your asset mix. Treat it as a linting tool: worth clearing the warnings, not worth optimizing as a score in itself.

Can I use ChatGPT to write the headline set? For drafting the candidates, yes, and it is one of the best-fitting tasks in the format - fifteen distinct 30-character headlines is exactly the variant-expansion work models do well. My prompts for it bake in the character limits and a compliance fence. The selection, the claim-pairing check and anything you would pin stay human.

This post is part of the paid media guides, alongside choosing a Smart Bidding strategy - the other half of the same machine - and the Performance Max guardrails. The wider practice starts at the complete guide to AI marketing.