AI Ad Copy Quality Report 2026: The Shorter the Ad, the More Generic It Gets

We scored 590 AI-generated ad copy assets across 10 landing pages, 5 languages and 3 ad formats against the pages they were written from. 100% fitted their platform's character limit and only 1 asserted a figure absent from the source page. 6.6% were generic — copy that shares almost no vocabulary with the page it promotes.

Key findings

6.6% of assets were generic (39 of 590) — they could be moved to a competitor's page unchanged.
Genericness scales inversely with slot length. 13.1% of 30-character Google Search headlines were generic, against 0.8% of Meta primary text, which has 500 characters to work with.
100% fitted the character limit (0 over). Character compliance is solved when the limit is stated as a counted rule.
1 of 590 assets cited an unsupported figure. Instructing a model to copy digits verbatim closes this gap far better than telling it not to invent things.
50.3% carried a concrete figure taken from the page.

Results by ad format

Format	Assets	Within limits	Grounded	Unsupported figure
Google Responsive Search Ads	260	100%	86.9%	0%
Meta (Facebook & Instagram) Ads	130	100%	99.2%	0%
Google Responsive Display Ads	200	100%	98%	0.5%

Most common defects

Google headline not grounded in the page — 33 occurrence(s)
Google headline near-duplicate — 6 occurrence(s)
Meta primary text hook runs past 125 chars — 5 occurrence(s)
Display short headline not grounded in the page — 4 occurrence(s)
Meta headline near-duplicate — 2 occurrence(s)
Display description under-uses the limit (52/90) — 2 occurrence(s)
Display description under-uses the limit (41/90) — 2 occurrence(s)
Meta headline not grounded in the page — 1 occurrence(s)
Google description not grounded in the page — 1 occurrence(s)
Display description under-uses the limit (49/90) — 1 occurrence(s)

Method

On 2026-07-25 we generated ad copy for 10 landing pages spanning B2B SaaS, e-commerce, local services, non-profit, regulated finance and a deliberately thin page, in Chinese, English, German, Spanish, Vietnamese. Each page was run through 3 ad formats, producing 30 ad sets and 590 individual assets.

Every asset was scored deterministically against its own source page — no human raters and no model judging its own output. "Grounded" means at least a third of an asset's content words appear on the source page. "Unsupported figure" means the asset states a number whose digits appear nowhere on that page; abbreviating a page figure (45 seconds to 45s) does not count. Scores are computed on raw model output, before validation, because that measures the generator rather than the safety net around it.

Scope. This measures one engine on a fixed corpus at prompt revision 2026-07-25.2. It is not a comparison of AI ad tools and the figures are not industry averages. Generation is stochastic, so a repeat run will differ slightly.