AI vs Human Ad Copy: The Studies Measure the Writer, Google Serves the Query

A simple question: are AI generated headlines better than their human counterparts? If we look at the published studies on AI generated ad copy vs human written ad copy, the larger samples find that AI ads can match or even better human written ads. One study of RSA copy on Google search, the only head to head of its kind, did test human written ad copy against AI written ad copy, with the human writer coming out ahead in a test that rests on just 91 clicks in total. But there is one mechanical reason for the studies coming to different conclusions: each study scores the writer, while Google serves a Responsive Search Ad (RSA), the search format that takes up to 15 headlines and four descriptions, as combinations it tests one query at a time. For an RSA the writer of the ad is not the unit of measure you are looking to score.

This piece contains the evidence in the form of the Google Help pages and six studies from the past few years, each described with sample size and with caveats.

What do the published tests of AI ad copy show?

Four sources investigate the ad copy written by AI against human copywriters and none of them even score in the same way.

Source and date	Sample	What it scored	Result
Hop Skip Media, Search Engine Journal, April 2, 2024	One RSA per side, 8 weeks, $500	Click through rate, Copy.ai against human writers	Human 4.98% (65 clicks, 1,306 impressions). AI 3.65% (26 clicks, 713 impressions)
Taboola with researchers from Columbia, Harvard, TUM and Carnegie Mellon, January 28, 2026	Hundreds of thousands of ads on Taboola's Realize platform, 500 million plus impressions	Sibling ads made with and without Realize's GenAI Ad Maker, a test of AI generated visuals	Raw CTR 0.76% AI, 0.65% human. Comparable under the tightest controls (vendor reported)
University of Queensland and University of Zurich, arXiv preprint, December 3, 2025	400 and 800 participants	Stated preference between full ads, with AI text and AI images on one side	Parity at 51.1% to 48.9%. AI preferred 59.1% to 40.9% on persuasion appeals
Brainlabs, Search Engine Land, March 12, 2026	23 tests, 16 advertisers, nine months	Google generated assets inside AI Max	Better ROAS on generated assets. Weighted Quality Score 6.8 to 7.3

The one published comparison of a writing tool to human writers on Google search RSAs is a pilot study written up by Ameet Khabra, founder of Hop Skip Media, who named the budget and the timeline as its limits. He writes up a two way comparison of a writing tool vs human writers for search ads, and on 91 total clicks a two proportion test of the published click counts gives a z-score of about 1.4 and p-value of about 0.17, by our own arithmetic rather than his. In other words, such a CTR difference is likely to occur by chance about once in six such tests of human vs AI writing. The human written and AI written ads were served by Google 1,306 times and 713 times, respectively. In other words, the serving system gave the two ads different shares of the auction, quite possibly on different queries. Thus the CTR difference reported in the study carries that query mix inside it.

The studies with large sample sizes lean the other way, finding that AI generated ads perform as well or even better than the human written ads, though each one needs its label read. (Note that Taboola is a company that owns the platform it studied, its ads ran outside of search, and by Taboola's own account the AI side of each pair was imagery from its GenAI Ad Maker, which makes this vendor study a finding about generated visuals rather than written copy. A study by researchers in Queensland and Zurich, a preprint measuring stated preference, also gave its AI generated ads generated images, so it too is not a pure test of written copy. A study by Brainlabs, an advertising agency, reports on its own tests run for its advertisers, but the agency does not name the advertisers in question.)

Why do the studies disagree?

All these studies look at different units (from click through rate on one RSA per side to survey preference) so they disagree, and Google's RSA help page describes a serving system with no fixed unit at all, one that tests combinations over time and learns "which combinations are the most relevant for different queries."

Each study, including the write up, has its own set of limitations. Because Google serves up RSA ads as combinations of different assets for different search queries, the studies on ad level click through rate average out the queries that the writer couldn't have chosen. Thus none of the studies can effectively score writer vs. query level optimization.

Of course the closest we get to a query level reading is Brainlabs, whose write up notes that the Google Ads interface reports Quality Score only when the search query exactly matches the keyword, a narrow window of exact match queries. Brainlabs, for example, was able to report a rise in weighted Quality Score for exactly matched queries in its AI Max before and after tests (up from 6.8 to 7.3).

What does Google say its generated headlines do?

This feature, Text Customization, is currently set at the campaign level for Search campaigns within the Google AI feature bundle, AI Max. Text Customization generates additional ad headlines and descriptions in your campaigns from your domain, landing page, existing ads and keywords. Google help documentation for this feature describes the way in which extractive snippets are mixed with generative AI that Google says is grounded in your landing page, and in Google's words the output "is guaranteed to be relevant to your products and offerings." It gives the example of a model that would not write a makeup related headline for a shampoo only page.

Text customization (formerly called automatically created assets) in AI Max for Search campaigns names the landing page as its main grounding source, which means the page is the strongest prompt you control and a generic page produces generic headlines. In the text customization page, Google adds that generated assets "will only serve if they are predicted to perform better than the headlines and description lines you uploaded." A specific FAQ answer is also given, i.e. generated assets do not count against your 15 headlines and four descriptions and can be removed one at a time. The same grounding applies in Performance Max (as described for PMax automated assets in our piece here).

Google's new steering controls for text (in Beta) for 'text guidelines' allow up to 25 terms to exclude and up to 40 Messaging Restriction messages per campaign. Note that Google lists "Don't target purely informational queries" as an ineffective guideline, because query targeting is managed in a totally separate system (keyword or audience targeting) and therefore cannot be controlled through the words.

On April 15, 2026 Google announced that campaigns already using automatically created assets will be automatically upgraded to AI Max starting in September, with Search term matching and Text Customization on by default. Google expects the transition to be complete by end of September 2026.

Does a high Ad Strength rating mean AI generated RSA headlines are good?

Ad Strength is Google's rating of how well your assets match your keywords and how many combinations you give the system. However, Optmyzr's study published September 9, 2024 found that the rating did not track performance. The best CPA and ROAS were for RSAs with an Average Ad Strength rating, and in general click through rate barely moved across the labels.

That 2024 study was a study of 22,000 accounts across more than one million ads, looking at how Ad Strength lined up with CPA and ROAS. Note the study only looked at accounts that Optmyzr has access to, a sample Optmyzr says skews toward advanced marketers, so read it as a strong signal from a selected sample (Optmyzr sells optimization software after all). A follow up study across roughly 20,000 accounts was released April 6, 2026. That study also found Average ads to be best at CPA ($12.43) and Excellent ads to be worst ($28.68), from the same kind of selected sample.

Google's Ad Strength page says the rating is not used to calculate Ad Rank or Quality Score. Furthermore, Google publishes a conversion figure of its own: Google reports that advertisers who improve Ad Strength from Poor to Excellent see 15% more conversions on average, based on internal data dated August 15 through August 20, 2025, six days with no method attached.

It is also worth noting that in Google Ads, there is a hard ceiling on the number of headlines that can be served. Once you have added 15 headlines to an RSA, you can add up to 2 more enabled RSAs with 15 headlines each, since Google's account limits allow three per ad group (for a total of 45 advertiser created headlines). According to Google's internal data, a second RSA rated Good or Excellent averages 6.6% more conversions at a similar cost per conversion, while a third adds a further 3.7%, so the gains flatten fast, and nothing in Google's description of Ad Strength compares a headline with the search that triggered it.

What is the strongest case for letting AI write the headlines?

Brainlabs' write up on Search Engine Land found that in their test of Google created Assets for Search, using the Added by segment in the Assets report, generated assets improved ROAS on both headlines and descriptions, and that in their agency reported tests all 3 AI Max features running together yielded a 40% higher uplift in test success rates. Of all the test cases, only half used Text Customization. (A graphical representation of the increased ROAS for Headlines is also included in the write up.)

In terms of evidence for the various persuasive elements, Queensland and Zurich present some very interesting data in their preprint on persuasive elements of generated creative, with AI generated ads winning authority appeals at 63.0% average and consensus appeals at 62.5% average in their tests. Taboola's vendor reported sibling pairs (i.e. drawn from the same advertiser's campaign on the same day) found parity for generated creative, which, as noted above, bears on headlines for ads only by extension.

While writing 15 truly different headlines is a lot of work, a model can produce many concise alternative versions at almost no cost. Optmyzr's 2024 data says headlines "appear to benefit from concision," and the record so far shows that at scale AI works as a drafter that matches human written ad copy. There is no indication from the evidence of how the mix of search queries might impact this in any way.

How do you check generated headlines against your queries?

Open the Search Terms Report for the last 90 days and sort by cost. Write down the desired answer for each of the top 20 search terms for which you are spending the most money. This will take about 20 minutes.

Now switch over to the Assets report for the biggest ad group for those search terms. Sort by Added by to see yours on top and the Google generated ones below. Look at the report for combinations of these and sort by % of impressions for each ad. Google Ads reports queries and assets in separate places, with no view that joins a search term to the headlines shown on it, so look at each of your search terms one at a time and see if any of the headlines in the ad group's set answer what the searcher wanted. (A headline that names your product is very different from a headline that answers a question for a search term that posed a question.) If a high cost search term has no such headlines, write a headline that answers the question posed by that search term for that ad group. Remove any of the generated asset headlines that promise things that your search terms did not ask for. If you run text customization, add a term exclusion or a messaging restriction with an example, since Google says the model "can only take hard requirements."

Pinning should be used sparingly, since Google says it reduces the number of headlines that can be matched to a search, and Optmyzr's 2024 study found some pinning and no pinning close on CPA. Harrison Jack Hepp, a Google Ads consultant, noted that most changes tested "saw no more than a 0.5-1% change across the CTR." This means that judging the effectiveness of your headline tests is best done by measuring the conversion per click and cost per conversion for the tests.

Query drift refers to the way that the search queries that your ads are showing up for can change over time, often unexpectedly. Broad match keywords and AI Max will serve ads against a wider and wider set of search queries, and every new query lands on a headline pool written for the old ones. Our article on what broad match does to account structure covers this topic in more depth. In a large enterprise paid search account with hundreds of ad groups, this exercise will likely focus on the highest cost ad groups.

Where AI ad copy still needs an expert

Using a model or even the text customization within AI Max to create a first draft is something that is strongly backed up by the evidence reviewed above. Although none of the various sources analyzed within the above piece scored which promise of benefits belongs on which search query, Google's documentation outlines a process where one query at a time is used to select the best combination of headlines and descriptions. Thus the judgment of which headline fits which query stays with the expert who reads the search terms report against the asset set, whether the human or the model wrote the initial headlines. Text customization earns its place particularly when the landing page is specific, as the generator is using the landing page as the primary grounding source for the written headlines. Get a list of your 20 highest cost search queries by cost this week and then read them against the current written headlines in the specified ad groups.