How to Use Gemini Inside Google Ads Without Shipping Generic Creative

Before you click Generate, write a one-page rule: Gemini may draft extra headlines and descriptions; Gemini may suggest crops and asset mix; Gemini may not be the only video in a Demand Gen group; auto-apply image enhancements stay off until you have reviewed a sample. If you do not draw that line, the UI will draw it for you in the direction of more automation and less distinctiveness.

Google's generative tools now sit in several places at once: asset generation for Performance Max and Demand Gen, Asset Studio for images and video, Gemini-powered recommendations while you pick Demand Gen assets, and text customization inside AI Max for Search. Treat them as one system with one policy. The policy should name what must stay human: the primary vertical video, the offer line, and any claim that could get you disapproved. Suggestions around those assets are useful. Replacing those assets with a generated smile and a 'Shop the look' headline is how accounts fill with creative that cannot lose a test because nothing specific was ever tested.

Turn on a checklist in the campaign brief, not in your head: which asset types are generated, which are uploaded, which are pinned. Pinning is the actual control. Unpinned generated headlines will cycle with whatever else Gemini adds later, including lines you would never type. If your team cannot state the pin rule in one sentence, you do not have a Gemini workflow yet. You have a default. Defaults optimize for coverage and speed, not for a brand that needs to look like a person used the product on camera.

When you generate assets, do not stop at the URL Google wants to crawl. Add a three-line brief: who the buyer is, the one offer, and the one proof you can show. Edit the product summary until it matches the SKU you are actually advertising. A homepage crawl produces homepage ads. A SKU brief produces something you can defend in a screenshot review.

Paste customer language into the prompt the same way you would for a UGC script: the complaint, the result, the object they hold. Then generate headlines and descriptions, not the whole identity of the brand. Ask for twenty, keep five. The keepers should contain a concrete noun or a true constraint (time, quantity, what is in the box). Kill anything that would still work if you swapped your domain for a competitor's. Gemini is good at fluent generics because your category page is fluent generics. Your job is to starve it of that page and feed it the packet instead.

For ecommerce, generate from the product URL or Merchant Center item, not the collection page. Collection pages teach the model your catalog voice, which is usually feature soup. If you use generated images, start from a real pack shot and edit the scene rather than inventing a lifestyle that does not match the PDP. Shoppers who click through from YouTube or Discover will bounce if the person, the kitchen, and the lighting cannot exist on the landing page. Continuity is a conversion problem, not an aesthetic one.

Export or copy the Gemini headline set into a doc next to your approved claims. Mark each line as usable, hedged, or dead. Dead lines get deleted, not 'saved for later.' Hedged lines get a human rewrite that names the real proof. This pass is faster than it feels and is the difference between AI assistance and AI ad copy you would not put your name on.

Demand Gen headlines max out at 40 characters; descriptions at 90. Those limits make Gemini extra likely to reach for empty intensifiers because there is no room for a clause. Force a formula: object + true constraint, or result + time you can defend. 'Glass bottle, ships tomorrow' beats 'Elevate hydration today.' If a generated line includes 'best,' 'guaranteed,' or a medical-adjacent verb, it is dead even if the rest is clever. Clever is not a policy exemption. Put the surviving lines back into the asset group as pinned text so they cannot be diluted by a later auto-generated batch.

Do the same for long headlines and RSA assets if you are also using Gemini on Search. Synonyms can change legal meaning: 'tested' is not 'proven,' 'helps' is not 'stops,' 'up to' is not optional. If you sell across countries, run this pass per market rather than translating a US winner into a market where the claim is tighter. Gemini will translate fluently and still be wrong. Fluency is the trap. The claims list is the filter. No line ships without a human initial on that doc, even when the UI called it high-performing predicted text.

When Gemini suggests how to optimize assets for YouTube, read the suggestion as a coverage checklist: you may be missing vertical video, a square crop, or a first frame that reads at a glance. Do not let it replace a working UGC cut with a generated or auto-resized clip just to tick the box. Coverage without a native master is how Shorts inventory gets a letterboxed catalog video.

June 2026 Demand Gen updates pushed Gemini recommendations at asset selection and broader auto-resizing (vertical to square, vertical to landscape, and the reverse). Resizing is useful for filling placements after you have a real 9:16 master. It is a poor substitute for shooting or generating that master. If the recommendation is 'add video,' add a talking-head-plus-demo cut, not a Ken Burns pan across a pack shot. If the recommendation is 'add more images,' add images that match the video's first frame rather than a new lifestyle set that splits the brand in two.

Look at the asset-group report after learning, not at the recommendation badge at setup. Gemini can be right that you lack a ratio and wrong about which file should fill it. Rank assets on conversions and view-through where you have them, then pin the top two humans and the one generated extra that actually moved. Unpin the rest. An asset group that contains twelve Gemini images and one weak video will spend like it has variety while teaching the auction a generic look. Variety is distinct files that can win. It is not a folder of near-duplicates.

Produce one 9:16 talking-head-plus-product cut from a locked script, then let Gemini and Google's video enhancements make ratio variants from that master. Do not start from a generated video and hope it feels like UGC on Shorts. The face, the hands, and the first two seconds still do most of the work in feed inventory. Gemini should follow that file, not invent a different spokesperson.

Klip Kanvas can generate that vertical UGC master from your product URL — hook, demo, captions — so the file you upload to Demand Gen already looks like social, not like a resized TV ad. Once it is in the asset group, use Google's resizing to reach 1:1 and 16:9 rather than rendering three unrelated videos. Keep the same hook line in the first two seconds across ratios. If Gemini later offers a generated video, compare it against the UGC master on a phone, sound off. If you cannot tell what is being sold in the first frame, it does not ship, no matter how new the model badge in the UI looks.

For catalogs, pair the UGC master with product-feed units rather than replacing the person with a grid of SKUs. Feed units carry price and availability; the UGC cut carries trust. Gemini image generation is at its best as background cleanup and extra angles of the real product, not as a fake creator. If you would be embarrassed to put the generated person next to your real reviews, do not put them next to your real bid. YouTube viewers are not confused about which one is an ad. They are deciding whether the ad looks like everyone else's Gemini output.

After seven days of stable delivery, open asset-level reporting. Pin the headlines and videos that hold CPA or conversion value. Remove generated assets that spent without converting. Generate a small replacement batch only for the slots you emptied, using the winning lines as the new brief. A living asset group beats a one-time Gemini dump you never touch.

Do not wait for a statistically perfect read on every headline. You need directional evidence and a bias toward specific language that matches the landing page. If a generated image spends because it is bright and empty, that is not a win. That is a tax. Replace it with a closer crop of the product in use from the UGC master. If a Gemini headline wins, promote its formula into the next brief instead of asking for 'more like this' with no example. Models copy examples. They hallucinate vibes when you only give adjectives.

Keep a simple log: date, assets added, assets pinned, assets removed, and whether Gemini auto-apply was on. When performance breaks, that log tells you whether the auction changed or whether the asset group quietly filled with new generated text. Quiet fill is the failure mode of Gemini inside Google Ads. The UI is designed to keep adding. Your workflow is designed to keep subtracting until what remains could only be your product, spoken in your claims language, on a frame you would show a founder.

Watch the winning Demand Gen ads on a phone with sound off, then open the landing page in another tab. Confirm the first frame, the offer, and the product match. Gemini is good at drifting those three things independently, especially after a resize or a new generated headline. Scaling a mismatch trains the campaign to buy clicks that bounce and then look like a bidding problem.

Check safe zones for Shorts: no essential text in the bottom fifth, no claims in captions that you already killed in headlines. Confirm auto-resized landscape versions do not crop the product out of the talking-head shot. If the PDP price or bundle changed since generation, rebuild the text assets rather than hoping Gemini's old summary is still true. Then — and only then — raise budget. Gemini speed is a gift at draft time. At scale time, slowness is the point: you are choosing which of its sentences are allowed to represent the brand on YouTube.