Open almost any guide to getting your products mentioned by AI assistants and you will find the same first step: export your Search Console queries and build from there.
It is a reasonable instinct. Search Console is free, already connected, and for twenty years the closest thing ecommerce had to a list of what buyers want. The habit is well earned.
It is also, for this particular job, the wrong tool. Not because the data is bad, but because of what the data is.
Search Console records clicks, not questions
Search Console reports on queries that produced an impression or a click in Google Search. That is its scope, and it does the job honestly.
An AI assistant does not work that way. When a shopper types “which of these will actually survive a dishwasher” into ChatGPT, that string is submitted to OpenAI. It is not a Google query. It produces no impression in your property, no click, no row in your export. Whatever the assistant then says about your product, whether it names you, ignores you or describes you inaccurately, happens entirely outside the reporting surface you are looking at.
So the gap is not one of coverage or sampling. Search Console is not under-reporting AI demand. It has no mechanism to report it at all.
That leaves a specific failure mode. If your content brief is built from a Search Console export, you are writing pages against the phrasing of people who used Google, in the past, and who behaved the way Google search behaviour rewards: short, keyword-shaped, stripped of context. The people asking assistants write in full sentences with conditions attached. Different input, different demand, different page.
The overlap used to be big enough to ignore this
For a while you could argue the two demand curves were close enough that a Google-shaped brief covered both. That argument is getting weaker, and there is now a number on it.
Ahrefs studied 863,000 keyword SERPs and four million AI Overview URLs, published 2 March 2026. Only 38% of AI Overview citations came from pages ranking in the top ten. A year earlier the figure was 76%.
Read that carefully, because it is easy to over-claim. It covers Google’s own AI Overviews, not ChatGPT or Perplexity, and it describes where citations come from rather than what gets bought. But the direction is hard to argue with. The set of pages an AI answer draws on has come apart from the set of pages that rank. Optimising for the second is no longer a reliable way to land in the first.
Which means the tool that tells you how you rank is losing its claim to be the tool that tells you what to write.
The substitute is already on your own site
Here is the part that tends to be missed. You do not need access to OpenAI’s logs to see how your buyers phrase things. You have a search box.
On-site search queries are prompts. They are typed by real buyers, in their own words, about your actual catalogue, on a property you own. Nobody has to grant you access and no vendor sits in the middle.
The valuable subset is narrower still: the searches that returned nothing.
A zero-result search is a buyer telling you, in their own language, about a thing they expected you to have or to explain, and could not find. Sometimes that means a genuine gap in the range. Very often it means the product is right there and the page does not use the words the buyer used — no mention of the material, the compatibility, or the threshold they cared about. That is not a merchandising problem. It is a page that fails to answer a question a real person asked, which is the same failure that keeps you out of an AI answer.
On Shopify you can get at this two ways. The Search & Discovery app has a report called “Searches with no results”, which covers the last thirty days. For a raw ongoing stream, the Web Pixels API emits a search_submitted event carrying event.data.searchResult.query alongside the variants that came back, so an empty productVariants array is a zero-result query, captured as it happens.
Neither needs a new vendor. Both are live today.
What this is not
Three honest limits, because a brief built on a false premise is no better than one built on the wrong export.
Your search box only hears from people who already reached your site. It tells you nothing about the shopper who asked an assistant, got three brands, and never arrived. On-site search corrects the phrasing problem. It does not solve attribution, and nobody should sell it as if it does.
Volume is thin. A store doing modest traffic might see a few dozen zero-result queries a month, and the Search & Discovery report only looks back thirty days. This is qualitative material. Treat one query as a signal worth reading rather than a datapoint worth counting.
And there is no study showing that fixing your zero-result queries makes an assistant recommend you. The argument here is upstream of that: a page written from a record only you hold cannot be produced by a competitor working from public sources. Whether that page then earns a citation depends on a lot of things this post does not claim to settle.
What to do this week
Pull your zero-result searches for the last thirty days. Read them rather than counting them.
Sort them into two piles. One is genuine range gaps, which is a buying conversation. The other, usually the larger one, is products you already sell, described in words your buyers do not use. Every item in that second pile is a page brief, already written for you, in the buyer’s phrasing, sourced from a record no competitor can see.
Then ask the question that matters before you write anything: could a competitor with access only to public sources have produced this page? If the answer is yes, the page may still be worth publishing. It will not be worth citing.
For what is and is not measurable once those pages are live, see measuring AI discovery. For why this channel stays quiet when you are losing, see the discovery channel that never tells you.