Google put a conversation inside YouTube ads. Your product feed is now the script

Google's September Demand Gen Drop, published on 24 September 2026, contains one item that changes what an ad is. Business Agent puts a Gemini-powered conversation inside a YouTube in-stream ad, and it answers questions using your Merchant Center feed. Whatever is in that feed is now the script, read aloud to a stranger, without anyone at your company checking it first.

What was actually announced, and what it does

Two announcements sit behind this, a week apart, and conflating them is where most of the commentary has gone wrong.

On 16 September 2026, Google opened a US beta of Business Agent for YouTube Ads to selected advertisers. On 24 September 2026, the monthly Demand Gen Drop confirmed it publicly alongside two other items: one-click experiences on image ads in YouTube Shorts and Gmail, and affiliate location extensions surfacing promoted pins across Google Maps for retail networks and dealerships.

Business Agent itself is a conversational overlay attached to the ad unit. A viewer watching an in-stream ad can ask a question, does this fit a 2019 model, is it in stock in my size, what is the difference between these two variants, and get an answer inside the ad, without leaving the video. Reporting on the beta describes it pulling spec sheets, comparing SKUs and routing to a live agent where one is configured. The answers come from the product feed in Merchant Center.

That last sentence is the whole story. Everything else in this article follows from it.

Are you even eligible? The gate, in full

The beta is narrow, and it is worth checking before you spend a meeting on it. Every condition below has to be true at once.

Two implications for anyone still running standalone Display. First, the Display-to-Demand Gen migration you may have been deferring is now the gate to a product surface, not just a campaign-type change. Second, if you do not have a Merchant Center feed at all, this is not a feature you can turn on; it is a project.

Your feed stops being a catalogue and becomes a sales script

This is the reframe worth internalising, because it changes who in your business owns the feed.

A product feed has always been read by machines to decide whether to show your product, and by shoppers as a listing: title, image, price, availability. Errors were mostly quiet. A thin description cost you nothing visible; a missing attribute lowered your eligibility in ways nobody traced back to a spreadsheet.

An agent answering a live question works differently. It cannot decline to answer gracefully. Faced with "will this fit my door frame?" and a feed that has no dimensions, it has three options, all bad: say it does not know, answer from something adjacent and be wrong, or generalise. A listing with a gap looks incomplete. A conversation with a gap looks like an unhelpful salesperson, attached to your brand, at the exact moment of interest.

The table below is the version of feed quality that matters once the feed is being spoken. It maps the questions an in-ad agent actually gets to the field that answers them, and what the failure looks like when the field is empty.

Several of those fields are the conversational attributes Google added to the Merchant Center product data specification at Google Marketing Live 2026: structured question-and-answer pairs, related-product relationships typed as often_bought_with, required_part, accessory or substitute, and document links for manuals and sizing guides. They were optional and largely ignored when they shipped, because nothing visible depended on them. Something visible now does.

A feed-readiness score you can run in an afternoon

Do not audit the whole catalogue. Take your top twenty products by revenue and score each of the seven rows below: 0 if the field is absent, 1 if present but generic, 2 if specific and accurate. Maximum 14 per product.

1. Availability and lead time — accurate today, not last week.
2. Price including shipping reality — matches the landing page exactly.
3. Dimensions, size and material — as numbers and units, not adjectives.
4. Variant differentiation — each variant's title and description would let a stranger tell them apart.
5. Compatibility and fit — stated explicitly, including what it does not fit.
6. Related products typed correctly — particularly required_part, which prevents a wrong recommendation.
7. Documents linked — manual, spec sheet or sizing guide where the product needs one.

Read the result like this. 11 to 14: join the beta if you can get access; the agent has something to work with. 7 to 10: fix the critical rows first, then test on those products only. Below 7: do not put a conversation in front of this feed. You would be paying to have your worst data read aloud.

The useful part of this exercise is that it pays off whether or not you ever get into the beta. The same fields drive Shopping surfaces, AI Mode shopping answers and free listings. A feed audit is the rare piece of paid media work with a compounding organic return.

The measurement gap nobody has solved yet

Here is where scepticism is warranted, and it is not about the technology.

The conversation happens inside the ad unit. It does not happen on your website. Your analytics, your session recording, your on-site search logs and your heatmaps see none of it. A viewer can ask three questions, get three answers, decide against buying and leave, and the only trace in your account is an impression.

That produces two specific reporting distortions to expect:

1. Click-through rate may fall while conversions hold or rise. If the agent resolves a question that previously required a site visit to resolve, some clicks disappear and the remaining ones are better qualified. Judging this placement on CTR will make a working feature look broken.
2. Your landing page's job changes without notice. Visitors arriving after a conversation have already had the specification question answered. A landing page that leads with specifications is now repeating itself, and the page that converts is the one that gets straight to price, availability and checkout.

So instrument the things you can still see, and be explicit that they are proxies: conversion rate and revenue per impression for the placement rather than CTR; assisted conversions for feed-linked Demand Gen; and on the site side, the bounce and scroll behaviour of that traffic against your other Demand Gen traffic. If post-conversation visitors behave differently, your page needs a different opening. This is the same class of blind spot that makes on-site AI assistants hard to evaluate, with one extra twist: here the assistant is not yours, and you cannot read its logs.

What Google has and has not published

Worth separating, because a figure from this Drop is already being attached to the wrong feature.

Google's published number in the September Demand Gen Drop is that advertisers adding Gmail to their Demand Gen campaigns see, on average, a 40% increase in conversions from their image creatives at the same ROI. That figure is about the Gmail placement and the new one-click experience on image ads. It is not a Business Agent result, and it should not be quoted as one.

For Business Agent, Google has published no conversion lift, no CTR effect, no sample size and no rollout schedule beyond a US beta. That is normal for a beta and it is also a reason not to build a Q4 plan around it. The correct posture is: fix the feed now, because the feed pays either way, and treat the agent as a test you take if invited.

Should you test it, or wait?

The pattern across the last two years of Google ad releases holds here too: the feature is a wrapper, and the asset it wraps is your structured data. Advertisers who kept their feeds honest keep getting handed new surfaces that work on day one. Advertisers who treated the feed as an export job keep discovering, one launch at a time, that the export job was the strategy all along. It is the least glamorous line item in paid media work and reliably one of the highest-returning.

Frequently asked questions

What is Business Agent for YouTube Ads?

It is a conversational overlay attached to eligible YouTube in-stream Demand Gen ads, powered by Gemini, that lets a viewer ask product questions and get answers inside the ad unit without leaving the video. The answers are drawn from the advertiser's Merchant Center product feed. Google opened a US beta to selected advertisers on 16 September 2026 and confirmed it publicly in the September 2026 Demand Gen Drop on 24 September.

What do I need to be eligible for the Business Agent beta?

Reported requirements are all of the following at once: United States targeting, English-language ads, an active Demand Gen campaign linked to a Merchant Center product feed, mobile in-stream placement, a budget of around $10 per day minimum, and selection or invitation by Google. Shorts placements are excluded. Advertisers without a Merchant Center feed cannot use it at all.

How does Business Agent change what my product feed needs to contain?

The feed stops being a listing and becomes the source of spoken answers. Availability, price, dimensions, size, material, compatibility, variant differentiation, related-product relationships and linked documents all become things a viewer can ask about directly. Google's conversational attributes in the Merchant Center product data specification, added at Google Marketing Live 2026, cover structured question-and-answer pairs, typed related products such as required_part and accessory, and document links for manuals and sizing guides.

Does the 40% conversion figure apply to Business Agent?

No. The 40% figure Google published in the September 2026 Demand Gen Drop refers to advertisers adding Gmail to their Demand Gen campaigns, who on average see a 40% increase in conversions from image creatives at the same ROI. It describes the Gmail placement and the new one-click experience, not Business Agent. Google has published no conversion lift, CTR effect or sample size for Business Agent itself.

Will Business Agent reduce my click-through rate?

It plausibly will, and that is not necessarily a problem. If the agent answers a question that previously required a site visit, some clicks disappear while the remaining ones arrive better qualified. Judge the placement on conversion rate and revenue per impression rather than CTR, and expect post-conversation visitors to need a landing page that leads with price, availability and checkout rather than with specifications.

Should a small business test this in Q4 2026?

Only if you are a US retailer with a strong feed and existing Demand Gen spend, and even then not during your peak weeks. A new ad surface with no published performance data is a poor risk in your highest-revenue period. The work that pays regardless is the feed audit: the same fields drive Shopping surfaces, free listings and AI Mode shopping answers, so improving them returns value whether or not you ever enable the agent.

The takeaway

Strip away the novelty and this release does something simple: it takes a spreadsheet most companies treat as plumbing and makes it speak to customers. That is a governance question as much as a media one, because nobody currently signs off on feed copy the way they sign off on ad copy. Start where the return is certain. Take your twenty best-selling products, score the seven feed fields above, and fix the critical rows. If the beta reaches you, you will be ready for it. If it never does, you will still have improved the data behind your Shopping listings, your free listings and every AI shopping answer that reads your catalogue.