What happened
Google Ads' built-in AI image generation picked up a new input: a reference image. Upload one — a product shot, an existing ad, any visual you want the output to resemble — and the tool uses it to guide the style of what it generates. Search Engine Roundtable flagged the addition on July 28.
Until now, the only steering you had was the prompt box. You described the look you wanted in words, the model interpreted those words however it pleased, and you either accepted the result or rewrote the prompt and rolled again. Style was a suggestion. For advertisers generating assets at any volume, that loop burned hours and still shipped images that looked nothing like the rest of the account.
A reference image changes the input, not the guarantee. The output is still generated imagery, and it still needs a human look before it lands anywhere a customer sees it. But steering with an actual picture of your brand beats describing your brand to a model and hoping.
Why this matters
Asset groups reward volume. Google keeps asking for more headlines, more images, more variations, and the path of least resistance is letting the machine fill the gaps. That is exactly where off-brand creative leaks in — the fourth lifestyle image nobody scrutinized, generated from a prompt someone wrote in thirty seconds. A style reference makes the lazy path less dangerous.
There is a harder line ecommerce stores should hold, though. Style guidance is fine for backgrounds, lifestyle scenes, and seasonal variations. It is not permission to generate the product itself. A generated image that misrepresents what arrives in the box is a returns problem and a trust problem, and no reference image fixes that. Use the feature to make filler match your look — keep real photography in every slot where the buyer is deciding what the product actually is.
Temper expectations, too. The pitch is that the reference guides the style, and generation tools have a habit of treating guidance loosely. The reference image is a knob, not a contract. Expect it to cut the number of bad generations rather than end them — which still matters, because every rejected asset costs review time, and review time is the real price of AI creative tooling.
What to do about it
Run a side-by-side test this week
Upload your strongest product photography as the reference and generate a batch of assets. Put them next to your real photos and judge honestly: same lighting, same color treatment, same feel? If the drift is still obvious, the feature is not ready for your account yet — and now you know before spend told you.
Write a pass/fail bar for generated assets
Decide what on-brand means in checkable terms — background style, color palette, composition — and reject anything that fails. One person should own that call. 'Looks fine' from whoever happened to be in the tool is how drift got in the first time.
Audit what is already live
Open your asset groups and review every generated image currently serving. Anything that predates a style reference was steered by prompt text alone. Replace the worst offenders first; they have been representing your brand without supervision.