At Google Marketing Live in May, among a pile of announcements about AI shopping, there was a quiet one that I think matters more than the loud ones.
Google added six new optional fields to Merchant Center, which is where online shops send their product data. They are called conversational attributes, and they exist for one purpose: to help AI understand products well enough to answer specific questions about them.
The most interesting of the six is simply question and answer. Structured pairs of “somebody asks this” and “the answer is this,” submitted as product data.
Google’s own example is a phone: “Does it have a headphone jack?” “This version does not have a headphone jack.”
That is an FAQ, submitted as data.
Why that is a bigger deal than it looks
Standard product feeds were never built to carry that sort of information. They hold titles, prices, images, identifiers. Facts about what a thing is.
What they could not carry was answers to the things people actually ask before buying, and that is precisely what an AI needs when somebody says “I want one of these but does it work with that.”
So Google has effectively said: give us your FAQs in a form we can read, because we are now answering questions rather than listing results.
They are optional and they do not affect whether your products get approved. But the direction of travel could not be clearer. Fill them in and you are answering the question. Leave them blank and Google either guesses or picks the competitor who filled theirs in.
If you sell products online, go and do this. It is a supplemental feed, it does not touch your main one, and the question and answer field alone is worth the afternoon.
Now the part that applies to everybody else
Here is why I am writing about this when most of my readers do not have a Merchant Center feed.
The principle underneath it applies to every business.
AI is answering specific questions rather than presenting a list. To do that, it needs specific answers. Google has built a formal pipe for shops to hand those over. Service businesses have no such pipe, which means the only place that information can live is on your own site, in a form something can read.
The problem is that you cannot put all of it on your main service page without ruining the page. Nobody wants to scroll through forty questions to find your phone number.
So the answer is supplementary pages. Your main page stays clean and readable for humans. Underneath it sit deeper pages holding the detail: the awkward questions, the exact terms, the comparisons, the things that only matter to somebody who is seriously considering it.
A human will rarely click through to those. An AI absolutely will, because it is not tired and it is not scanning.
What that looks like for my own business
Let me use myself, since I have three plans and exactly this problem.
My pricing page has to stay simple. Three options, clear prices, enough for somebody to work out whether it is for them. If I loaded it with every detail, nobody would read any of it.
But there is a great deal a genuinely interested person, or an AI acting for them, would want to know:
- What is the difference between a one page site at £150 a year and the £99 monthly plan, and which suits a business that only really needs to be findable?
- What happens after the minimum term?
- What exactly is included in “changes” and where does that stop?
- Who owns the site if I leave, and what do I keep?
- What is genuinely not included?
- How long does it take, and what holds it up?
- Does it suit a business with no photographs and no logo?
Those are real questions I get asked. Every one of them is a decision point, and every one of them is something an AI would need in order to recommend me confidently rather than vaguely.
Some of those already have posts behind them. The rest are a to-do list, and writing this out has been quite a useful exercise in seeing which.
That is the practical version of this for a service business: work out the questions somebody asks before they commit, and make sure each one is answered somewhere findable, in plain language, whether or not it fits on the main page.
And ChatGPT has built the same thing
Google is not doing this alone, which is the part that turns a Merchant Center update into a genuine shift.
OpenAI has its own equivalent, the Agentic Commerce Protocol, developed with Stripe and open sourced. Merchants sign up at chatgpt.com/merchants and supply a structured product feed containing titles, descriptions, prices, availability, media and fulfilment details.
There is one detail in there worth pausing on, because it changes the model completely.
ChatGPT does not crawl your product pages. You push it a feed.
That is a reversal of how the web has worked for twenty five years. Search engines came and looked, and your job was to make your site easy to look at. Here the merchant sends a file to an endpoint, and can refresh it as often as every fifteen minutes so prices and stock stay accurate.
Two things follow from that.
Accuracy becomes your responsibility rather than a crawling problem. No more waiting for a price change to be noticed. If the feed is wrong, the answer is wrong, and it is your fault.
And if you are not in the feed, you are not in the conversation. There is no crawling to fall back on. A shop that has not set this up is not being ranked poorly, it is absent.
Worth knowing too: OpenAI states that these results are not ads and are not influenced by paid placement, but depend on feed quality. Which is unusually good news for a small business, because it means the lever is data quality rather than budget.
Building with the protocol is open to anybody. The in-chat Instant Checkout part is currently limited to approved partners, starting with Etsy sellers and Shopify merchants, but the feed itself is not gated behind that.
The bit I would be slightly wary about
One honest observation about the direction of all this.
Google would like that information inside Google. Merchant Center is Google’s product, and data you put there is not data you own or control. Rivals cannot see it, and you become a little more dependent on one company. Exactly the same is true of OpenAI’s feed.
And now you can see the shape of the problem. Two companies, two separate specifications, two sets of field names, describing the same products. Next year there will be a third.
I would not be at all surprised to see something similar arrive for Google Business Profile, letting service businesses submit structured questions and answers about what they do. It is the obvious next step, and it would be genuinely useful.
My advice would be to do all of it. Give each platform whatever it asks for, because the visibility is real and there is no prize for being principled and invisible. But do not let that be the only place the information exists.
Optimising for one company’s field names is a game of whack-a-mole. The durable version is to have the underlying information written down properly, on something you own, in plain language, so that feeding it to whichever platform asks next is a formatting job rather than a research project.
Get it right once. Reformat it forever.
What to actually do
Whatever you sell, the job is the same and it starts offline.
Write down every question you get asked before somebody buys. Not the ones you wish they asked. The real ones, including the awkward ones about price and the ones people are slightly embarrassed to ask.
Answer each one plainly. Short, factual, no hedging. If the answer is “it depends,” say what it depends on.
Then find somewhere for each answer to live. On the page if it fits, on a supplementary page if it does not, and in Google’s fields if you sell products and they have given you somewhere to put it.
And keep it current, because a confidently wrong answer is worse than no answer at all.
A prompt to get you started
If you do sell products and the idea of writing question and answer pairs for four hundred items is making you tired, this is exactly the sort of job AI is good at, provided you keep it honest.
Copy this into whichever assistant you use. It is written to look up Google’s current specification itself rather than trusting my summary of it, because these fields will change and the post you are reading will not.
I sell products online and I want to fill in the structured product data that AI shopping surfaces now expect.
First, look up the current documentation for two things and tell me the exact field names, formats and limits that apply right now: Google Merchant Center’s conversational attributes, and OpenAI’s Agentic Commerce Protocol product feed specification. Do not rely on what you already think you know, because both were introduced recently and are still changing. Tell me where the two overlap, so I write each answer once rather than twice.
Then work with me on the content, one product at a time to start with.
Critical rules, and I would rather you refused than guessed:
- Every answer must come from information I give you or from my own website. Do not invent specifications, materials, measurements, compatibility, timescales or policies.
- If you do not know something, write “NEEDS INPUT” and tell me exactly what to go and find out. A blank field is fine. A wrong one is not.
- Write in plain language, the way a customer would actually ask and the way a person would actually answer. No marketing language.
- Answer the awkward questions honestly, including what the product does not do or is not suitable for. Those are the ones that build trust and stop returns.
For the first product I give you, produce:
- Eight to twelve question and answer pairs covering what a real buyer asks before purchasing, including at least two about limitations or unsuitability
- Anything else the current spec supports that I can honestly fill in from the information provided
- A list of gaps where I need to supply more information
Ask me for the first product and its details now. Once we have one right, we will agree a pattern and scale it up.
Two things to hold onto when you use it.
Make it refuse. The instruction to write “NEEDS INPUT” rather than guess is the important line in that whole prompt. Left alone, an AI will happily invent a plausible-sounding warranty length, and you will have published a promise you did not make.
Do one properly first. Get a single product right, decide what good looks like, then scale. Generating four hundred at once and reviewing them afterwards is how mistakes get published, because nobody reviews four hundred of anything properly.
The service business version
Same idea, no feed involved.
I run a [type of business] in [area]. I want to work out every question a customer asks before they commit to buying from me, so I can make sure each one is answered somewhere on my website.
Interview me rather than guessing. Ask me about my services, my prices, my process and my terms, one question at a time.
Then produce:
- A list of the questions a serious buyer would want answered before committing, phrased in their words rather than mine
- The awkward ones people are reluctant to ask out loud, particularly about cost, what happens if it goes wrong, and what is not included
- For each question, whether the answer belongs on the main service page or on a separate deeper page, and why
- Anything I have told you that contradicts itself, because that is what confuses both customers and AI
Do not write the answers for me. I will write those, because they need to be true.
That last line matters. The questions can be generated. The answers have to be yours.
None of that requires you to understand feeds, schemas or protocols. It requires you to write down what you already know and stop keeping it in your head.
Which, I am increasingly convinced, is most of what this whole shift actually asks of a small business.
Want to work out which questions your website is not answering? Get your free audit, or email hello@inkon.co.uk. I read every message myself.