Google started writing the description on your Shopping ads
On July 30, descriptions no advertiser had written began appearing next to Shopping and Product ads. Brodie Clark spotted them and published the sighting on his SERP Alert page: AI-generated text, sitting inside the ad block, on a listing somebody is paying for. It is not a new product or a format you opt into. It is a layer that appeared on top of campaigns already running.
Google confirmed it in one sentence: "a small experiment to see if adding AI-generated context to Search ads helps people make more informed decisions." The same test had been running since early July on Search ads. What happened on the 30th is its extension into Shopping, which is a rather different place.
What Google confirmed, and what it left out
The confirmation covers the existence of the test and its intent, and nothing else. Google published no geographic scope, no share of auctions affected, no end date, and no word on whether the experiment expands. It also documented no opt-out: as of this writing there is no checkbox, no campaign setting and no policy that lets an advertiser decline.
That absence matters more than the format itself. A test the advertiser cannot switch off is not a test for the advertiser, it is a product change in an observation phase. The difference is who absorbs the risk while it runs.
Shopping is the worst possible place for someone else's text
In Search, advertisers write headlines and descriptions and have lived for years with the idea that the system recombines them. In Shopping the relationship with text is different: the product title, its attributes and its description are the asset teams work on for months, because auction relevance and click decisions both come out of that chain. Entire ecommerce teams spend their week there.
On top of that work now sits text the advertiser did not draft, did not approve and cannot edit. If it describes an attribute poorly, flattens a difference between variants, or implies something the product does not do, the cost lands in the account paying for the click, not with whoever generated the sentence.
The measurement problem shows up immediately
A change like this would normally be measured with a test, and there is no clean way to build one. The advertiser does not control exposure: there is no way to segment who sees the generated description and who does not, and no dimension separating impressions with it from impressions without. Any CTR movement arrives mixed with seasonality, competition and auction shifts.
What is available today is a poor but honest reading: note the date the text appears in each account and compare against the prior period, knowing it is a correlation and not an experiment. It is worth recording that note anyway, because in two months nobody will remember when it started. It is the same principle we worked through writing about the measurement gate: without a timestamp and a question asked in advance, a change cannot be judged, only narrated.
If the model generates from the feed, the feed is copy
Here is the actionable part, and it is a single idea. The generated description does not come from nowhere: it leans on the product's structured information, which is exactly what the advertiser still fully controls. Feed title, attributes, category and description stop being a catalogue task and become the raw material somebody else writes from.
That changes the standard for reviewing them. A tidy feed for the search engine prioritises completeness and consistency. A feed that feeds a text generator also needs precision in whatever distinguishes one variant from another, and no ambiguity in attributes that could be read wrong. If a product description says "ideal for everyday use" without stating the material, the model will fill that gap with something.
We would look at it in that order: first the products concentrating the spend, then the ones with variants that resemble each other, and only then the rest of the catalogue.
The pattern, which is no longer an anecdote
This does not arrive alone. When we wrote about the official MCP servers across ads platforms, the point was that the account became operable by agents, on permissions each team decides. When we wrote about Manus AI inside Ads Manager, we flagged that the conflict of interest is structural: the platform charging for the space also proposes what to do with it. And looking at search turning into a checkout, what moved was the conversion point.
The generated description is the same move applied to creative. Each step is defensible on its own and sounds like an experience improvement. Added up, they describe a direction: the surface the advertiser controls keeps shrinking, and what remains under control becomes more valuable, not less.
Who this matters to right now
Anyone running Shopping with a large catalogue, especially where variants resemble each other. That is where the risk of a generated description flattening real differences runs highest, and the feed is the first thing worth putting in order.
Ecommerce teams that report CTR as a channel health metric, because it is worth explaining now, before the number moves, that there is a new variable outside their control.
And it does not matter yet, except as a read on direction, to anyone not running Shopping. It is a test with no published scope that may come to nothing. What is unlikely to come to nothing is the trend containing it.