Search Ad & Asset Testing

An ad is a hypothesis until the evidence says otherwise.

Responsive search ads can compare combinations of headlines and descriptions, but the comparison needs multiple RSAs and enough unpinned assets to create meaningful variation. We define the reason for each variant, launch a controlled comparison, and wait for enough evidence before interpreting the result. For ad groups with enough traffic to reach a real read, and reviewers available to check copy before launch.

An ad-testing rhythm where each live variant answers a specific question and the answer remains available for the next test.

A specialist comparing two responsive search ad variants on a testing board with a hypothesis written above each

Some of the 500+ brands we've worked with

See all references
  • MediaMarkt
  • LC Waikiki
  • Bayer
  • PWC Türkiye
  • Sigortam.net
  • Apsiyon
  • Yatsan

How the work runs

Every ad test leaves a traceable learning record

AI can draft variant copy and track the sample size. A person reviews brand-sensitive claims and signs off before a live ad launches.

How we hold ourselves to it

  • Every new asset starts as a stated hypothesis, not a guess
  • Pinning is a tradeoff you choose on purpose, every time
  • A test needs a real sample before anyone calls it
  • Every result gets written down, win or not

  1. Form the hypothesis

    Name the specific signal driving the test, whether that is low Ad Strength, weak click-through rate, or an unproven claim, and state exactly what the variant changes.

    Written hypothesis naming the driving signal and exactly what the variant changes.

    AI assist
    Drafts the hypothesis from the flagged signal and past learnings.
    Human gate
    Paid-search lead confirms the hypothesis is worth testing.
    Owners
    Paid-search lead + copy lead

  2. Draft the variant assets

    Write headlines and descriptions against the hypothesis, deciding deliberately what, if anything, gets pinned and accepting the combination trade-off that comes with it.

    Variant headline and description set with the pinning decision and its combination trade-off recorded.

    AI assist
    Drafts variant copy options against the stated hypothesis.
    Human gate
    Copy lead selects and edits the assets that actually launch.
    Owners
    Copy lead + paid-search lead

  3. Run it through policy and brand review

    Check claims, restricted terms, and brand voice before anything goes live, so a complaint or a disapproval never gets the first look.

    Cleared policy and brand review covering claims, restricted terms and voice.

    AI assist
    Flags likely policy or brand issues against the review checklist.
    Human gate
    Brand or compliance reviewer signs off before launch.
    Owners
    Brand reviewer + copy lead

  4. Launch a controlled comparison

    Run at least two RSAs per ad group with matched targeting and unique final URLs, so the comparison actually isolates the variable being tested.

    Controlled comparison running at least two responsive search ads per ad group with matched targeting and unique final URLs.

    AI assist
    Confirms the launched set matches the planned comparison structure.
    Human gate
    Paid-search lead approves the live comparison before it starts.
    Owners
    Paid-search lead + media specialist

  5. Hold for a real sample

    Let the test run until it reaches a defensible sample size and duration for the ad group's actual traffic.

    Run record showing the test reached a defensible sample size and duration for that ad group's traffic.

    AI assist
    Tracks accumulated volume against the planned sample size.
    Human gate
    Paid-search lead approves any early stop.
    Owners
    Paid-search lead + analytics lead

  6. Read the result and record it

    Call it a win, a loss, or inconclusive against the original hypothesis, then write the learning into the record for the next ad group.

    Learning record calling the test a win, a loss or inconclusive against the original hypothesis.

    AI assist
    Drafts the readout from the accumulated data.
    Human gate
    Paid-search lead signs the learning before it gets applied elsewhere.
    Owners
    Paid-search lead + copy lead

What lands with your team

A test history you can actually search

Each test keeps its hypothesis, comparison, and result together, giving the next ad group evidence it can consult before repeating the same question.

  • Test matrix

    Every hypothesis, its variant, the ad group it ran in, and its current status, in one running view.

    Accepted when

    Every hypothesis, its variant, its ad group and its current status appear in one running view rather than in separate spreadsheets.

    Cadence: On launch or conclusion

  • Asset library with pinning rules

    What's pinned, why, and the combination trade-off accepted for each ad group's asset set.

    Accepted when

    Each pin states why it exists and which combination trade-off the ad group accepted for it.

    Cadence: On add or retire

  • Policy and brand pre-flight checklist

    The specific claims, terms, and brand checks cleared before each variant went live.

    Accepted when

    The specific claims, restricted terms and brand checks are cleared and recorded before the variant serves.

    Cadence: Before every launch

  • Learning record

    Win, loss, or inconclusive, with the reasoning, kept searchable so a past result can inform the next ad group.

    Accepted when

    Every closed test is filed with its verdict and reasoning in a searchable form, including the inconclusive ones.

    Cadence: Close of every test

Before the work starts

When this work is the right next step

An unchanged ad stops producing useful learning. Google recommends running at least two responsive search ads with a Good or Excellent Ad Strength rating per ad group, each with its own final URL, precisely so there's something to compare. Pin every headline out of caution and you shrink the very combination pool the format is built to test. Call a result a win after two days and a handful of clicks, and all you've actually measured is noise.

A good fit when

  • Your ad group draws enough traffic for two responsive search ads, so the comparison can move well beyond a handful of clicks.
  • A brand or policy reviewer can check claims and restricted terms before launch, so every cleared variant enters the learning record with its approval attached.
  • The planned sample and hold period can govern the read, even when an early click-through swing makes one headline look convincing.
  • An ad group has run one RSA for months with no second variant to compare it against.
  • Every headline and description is pinned, so the combinations the format is built to test barely happen.
  • A new variant gets called a "win" after a couple of days and a small number of clicks.
  • Nobody flagged a policy-sensitive claim before the ad went live.

Better handled as other work when

  • Traffic is too thin for two responsive search ads to reach a meaningful comparison, so the test matrix would remain inconclusive.
  • A specific headline has to launch today, before a hypothesis, controlled comparison, and planned sample can be completed.
  • No brand or policy reviewer can clear sensitive claims before launch, so the variant needs copy approval outside this testing method first.

People who own a single channel

Paid Search, Paid Social, CRO, and Programmatic each run under a named owner at Zeo. The consultants below are matched to the channel this page is about, so you can see who you'd actually work with.

Tools we use

Tools behind this work

  • Adalysis

    runs the RSA variant comparison itself, from cloning assets to reading which one actually won

Put your ad copy through a real test

Run ad tests with enough evidence to make a call

Bring us the current ad groups. We'll set the hypotheses and comparisons, then keep each result in a learning record the next test can use.
Start testing your ad copy

The questions we hear most

At least two, matching Google's own recommendation, each with a unique final URL so the comparison is real, though more can run if the ad group's traffic supports reading them separately.