From Searching to Asking

The question a buyer asks an AI assistant today is not the same thing as a keyword. A keyword was a shorthand: "SEO agency New York," "brand refresh cost," "luxury marketing." It told you what someone wanted, stripped of everything else. A full question tells you something different. It tells you the situation, the concern, the stage, and the stakes.
That shift in how people phrase their searches changes what it means to be visible. And it surfaces a problem most companies have not yet named.
You cannot become the authoritative source for a question you do not know exists.
The rest of this article is about that problem: how large it is, why the standard explanation for it is wrong, and what it actually takes to measure the demand you are missing.
The question changed shape
A keyword was a fragment. "SEO audit." "Brand agency." "Market entry consultant." The user was compressing their need into something a search index could match. The search engine's job was to guess the rest.
An AI assistant does not need the compression. The buyer types the whole thought: "How long does an SEO audit take for a site with stalled traffic?" That sentence carries four distinct signals at once.
What a full question reveals that a keyword cannot
Signal | What it tells you |
What | The service or topic being researched (SEO audit) |
Situation | The context that triggered the search (stalled traffic) |
Concern | The underlying anxiety (time, cost, disruption, risk) |
Stage | Where in the buying process they are (evaluating, comparing, ready to act) |
A keyword gives you the first column. A full question gives you all four. That matters because the answer a buyer needs at the concern stage is different from the answer they need at the what stage, and a company that only knows the what is producing content for one signal while the other three go unaddressed.
This is not a small refinement. When buyers move from typing fragments to asking full questions, the information gap between what they reveal and what companies track widens dramatically. The keyword "SEO audit" tells you someone is interested. The question "How long does an SEO audit take for a site with stalled traffic?" tells you they are evaluating, they are worried about time, and they have already decided they need one. Those are three actionable facts the keyword never surfaced.
The companies that understand this shift are building content around the full question. The companies that do not are still optimizing for the fragment.
The shortlist is built before your site is visited
There is a sequence to how AI-assisted research works, and most companies have the order wrong.
The assumption is: buyer searches, lands on your site, reads, decides. In that model, your site is the first impression. It is where the evaluation begins.
The actual sequence is different. The buyer asks a question. The AI assistant draws on sources that are not yours, synthesizes an answer, and names the options worth considering. The buyer then visits those sites to confirm what the assistant already told them. Your site is checked third, to verify, not to inform.
Where the shortlist comes from
We track 1,757 source domains across AI responses in this topic area. The top two, by usage frequency, are google.com at rank 1 and youtube.com at rank 2, at 24.4% usage. Those are not company websites. They are platforms the AI assistant treats as authoritative aggregators of evidence before it ever reaches a brand's own domain.
This has a direct consequence for how you think about visibility. If the sources that shape the shortlist are platforms and publications rather than your own pages, then optimizing your site in isolation addresses only the confirmation step, not the selection step. To understand what SEO, AEO, and GEO actually mean for how that selection works, the mechanics are covered separately. The point here is narrower: the order matters, and the order puts your site later than most visibility strategies assume.
A brand that is absent from the sources an AI assistant draws on will not be on the shortlist the buyer receives. The site visit that would have changed that outcome never happens, because the shortlist was already closed.
The answer you get when you ask why you are invisible is a technical one
When a marketing team discovers their brand is absent from AI answers, the first move is usually to ask the model why. And the model answers. The answer is almost always the same.
It points to crawler access. It names robots.txt. It mentions GPTBot, OAI-SearchBot, PerplexityBot. It explains that if these crawlers are blocked, the AI system cannot read the site. The implication is clear: fix the technical access, fix the visibility.
We tested this directly through an analysis of the sub-queries generated while answering tracked prompts in this topic area. The sub-queries that surfaced were a documentation walkthrough: robots.txt, GPTBot, OAI-SearchBot, PerplexityBot, CDN and firewall configuration. That finding came from ChatGPT only, and the instrument limit is real. We have not verified that Perplexity, Gemini, or other assistants generate the same sub-queries behind the same prompt.
What we can observe more broadly is that the published advice ecosystem around AI visibility has converged on the same technical frame. Across the editorial sources and tool vendors that dominate this topic, the dominant answer to "why is my brand absent?" is a crawl-access checklist: robots.txt, bot permissions, rendering, WAF rules.
That is the wrong answer to the right question. Not because crawl access is irrelevant. It is not. But because the usual problem is not that the AI cannot reach your site. The usual problem is that nobody inside your company knows which questions are being asked about what you sell, which means there is no content addressing those questions for any crawler to find.
Fixing technical access on a site that does not answer the questions buyers are asking produces a perfectly accessible site that still does not appear. The constraint is upstream of the crawl.
google.com and openai.com together hold 3,533 citations in this topic area. The access is not the bottleneck. The questions are.
The questions you cannot see
We mapped the question space around what a professional services firm sells. Not the questions they were tracking. Not the questions they had published content for. The full space: every question a buyer might ask an AI assistant about their category, their services, and their competitors.
The question space around that business runs to roughly 64,000 questions.
Of those, 1,992 were available for measurement. The remaining approximately 62,045 were locked: questions being asked, questions shaping shortlists and buying decisions, questions the company had no visibility into and no content addressing.
That is about 3% accessible. Roughly 97% locked.
What the 3% looked like
The 1,992 visible questions were distributed across topics, buying stages, and branded versus generic intent. In the firm's largest single topic, there were 1,496 available questions out of a total estimated 20,287 in that space.
In one of its core service areas, we identified 107 questions. Not one of them was being tracked by the company's existing measurement program. The questions existed. Buyers were asking them. The company had no record of any of them.
This is not a content quality problem. It is a discovery problem. You cannot write the answer to a question you have never seen. You cannot measure your visibility on a prompt you have never tracked. And you cannot appear in an AI answer that is responding to a question your content does not address.
The 97% is not a failure of execution. It is a failure of instrumentation. Most companies measure their visibility on the questions they already know about, which is a small and self-selected sample of the demand that actually exists. The questions that would change their pipeline are in the 97% they have never mapped.
Not every question is worth the same
Knowing that roughly 64,000 questions exist is useful. Knowing which ones matter commercially is the work.
Questions are not equally valuable, and the distribution across buying stages tells you where the real opportunity sits. Of the 1,496 available questions in the firm's largest topic:
Stage | Questions | What it signals |
Top of funnel (TOFU) | 574 | Awareness and education: the buyer is learning |
Middle of funnel (MOFU) | 356 | Evaluation: the buyer is comparing options |
Bottom of funnel (BOFU) | 482 | Decision: the buyer is ready to act |
Post-purchase | 84 | Retention: the buyer is already a customer |
Nearly a third of the visible questions in this space come from people who are ready to buy. Those are not research questions. They are buying questions. And most companies, if they have any AI visibility program at all, are measuring their performance on informational prompts while the commercial prompts go untracked.
The post-purchase stage most companies have never considered
The fourth stage is the one that surprises most teams when they see it. Post-purchase questions are asked by existing customers: at renewal, when something goes wrong, when they are evaluating whether to expand the relationship or take it elsewhere.
AI assistants answer those questions too. If your content addresses the concerns of a buyer who is ready to leave, you can appear in that answer. If it does not, a competitor can. Most companies have never thought about AI visibility in terms of retention, because they have been thinking about AI visibility only in terms of acquisition. The question space does not make that distinction.
The funnel does not end at conversion. Neither does the demand.
Branded questions are the easy test and the least useful one
Of the 1,496 visible questions in the firm's largest topic, 721 were branded and 775 were generic.
That split matters because branded and generic questions test different things, and most companies, when they run any AI visibility check at all, run it on branded prompts. They ask the assistant what it knows about their company. They search for their own name. They measure whether the AI describes them accurately.
That is the easiest test to pass, and it is the least useful one commercially.
Why branded visibility is the wrong starting point
A branded question only exists after the shortlist has already named you. If a buyer is asking "What does [your company] specialize in?", they already know you exist. That question did not create the opportunity. Something else did, and that something else was an unbranded question earlier in the process.
New customers never type your name first. They type the problem. They type the category. They type the comparison: "[your brand] vs [competitor]." They arrive at your brand through the unbranded questions, and if your content does not appear on those prompts, the branded question never gets asked.
The 775 generic questions in this firm's space represent the demand that exists before a buyer knows which company to contact. That is where the shortlist is formed. That is where being cited determines whether the branded question ever follows.
But being absent from the unbranded questions means being quoted is not being chosen is a problem you never reach, because you were not in the conversation that preceded it.
Test your branded visibility if you want a baseline. Track your generic visibility if you want to understand your pipeline.
Your witnesses have to agree
An AI assistant does not read one source and form a view. It reads many sources simultaneously and looks for agreement. When sources conflict, the assistant has to choose, and a brand that sends conflicting signals across its own properties is harder to resolve with confidence.
For most companies, the sources an AI assistant reads include more than the company website. They include the store locator, the stockist listing, the retailer product page, and the local press roundup. For product companies, they include the spec sheet, the product page, the manual, and the retailer's description. For service companies, they include directory listings, review platforms, and any third-party coverage that describes what the company does.
The internal consistency problem nobody checks
Nobody inside most companies reads all of these side by side. The marketing team owns the website. The sales team owns the retailer relationship. The PR team owns the press. The operations team owns the store locator. Each is accurate in isolation. Together, they may describe a subtly different company.
Consider what happens when:
The website says you serve premium brands in fashion and hospitality
An outdated directory listing says you are a general marketing agency
The retailer product page uses a different brand name variant
The local press roundup from your last activation describes a service you no longer offer
Each source is a witness. The AI assistant is looking for corroboration. When the witnesses disagree, the brand becomes harder to resolve, and a brand that is hard to resolve is one the assistant will not put its name behind with confidence.
The consistency check is not a technical audit. It is a reading exercise. Pull every source an AI assistant might use to describe your company, read them in sequence, and ask whether a stranger would reach the same conclusion from each one. The gaps you find are the gaps the assistant finds too.
One run is an anecdote
AI answers are not stable. The response a model gives today is not the response it will give in two weeks. Sources change. Weights shift. New content enters the index. A brand that appeared in a strong share of answers to a given question last month may appear in a much smaller share this month, not because anything changed on their site, but because the landscape around them shifted.
This is not a theoretical concern. We measure it across client programs and publish the results honestly: across eight programs, 4 improved, 3 declined, and 1 was flat over thirty days. The median visibility rate across those programs is 3.5%, with a range from 1.4% to 14.0%. Those are not failure numbers. They are the current condition of the market. Most brands, at intake, are visible in a small share of the AI answers covering their category.
What the portfolio actually shows
The 3.5% median is a diagnosis, not a verdict. It reflects where brands are when they start measuring, not where they end up after sustained work.
One program in the portfolio has eighty days of structured work behind it. Over that period, visibility moved from 5.5 to 7.5, and brand mentions in AI answers went from 29 to 49, with a mention rate that rose from 14.4% to 24.4% across 201 responses per daily report. That is what the measurement looks like when the work is consistent and the question space is understood.
The portfolio shows the disease. That case shows the direction.
A sharp reader will notice the tension: we are describing a 3.5% median while also running the programs. The answer is that the median measures intake conditions, not outcomes. A brand that has never mapped the questions being asked about its category will appear in a small share of AI answers about that category. That is the starting point, not the ceiling.
Measure monthly or do not measure. A single check tells you nothing about direction. Sustained measurement is the only thing that tells you whether the work is working.
Start with the questions your buyers actually ask
We run AI visibility programs from our agency in Soho, New York. We measure every program monthly. The services we offer are set out on the questions your buyers actually ask. And the first thing we do with every client is not a technical audit. It is a question inventory.
The question inventory finds the questions buyers are asking about your category, labels them by topic, by branded or generic, and by buying stage. It shows you the 3% you can see and gives you a map of the 97% you cannot. It tells you which questions carry commercial intent, which ones are being asked at the moment of decision, and which ones your current content does not address at all.
That is the starting point. Not a ranking promise. Not a citation target. The questions.
Because you cannot become the authoritative source for a question you do not know exists. And right now, most of the questions being asked about what you sell are questions you have never seen.
If you want to know what buyers are asking about your category before a competitor finds out first, reach out. We will show you the map.
Frequently Asked Questions:
Which assistants and which markets should we track first?
Start with the assistant your buyers use most, which is usually the one that surfaces in your own sales conversations when prospects describe how they researched you. If that signal is absent, start with the assistant that dominates in your category's geography and buyer profile, then add a second once the first baseline is stable. On markets: track where your commercial intent questions concentrate, not where you have the broadest presence. A brand with strong U.S. demand and thin European traction should establish a U.S. baseline before splitting attention. Expanding across assistants and markets makes sense once you understand which prompt clusters are producing visibility and which are not. Starting everywhere at once produces noise before it produces insight.
What is the difference between a question a buyer types and one they ask an assistant?
A typed keyword compresses the need: "SEO agency," "brand refresh," "market entry." A question asked to an AI assistant carries the full context: the situation, the concern, the stage, and often the stakes. "How long does a brand refresh take for a company preparing for a U.S. launch?" is a different signal from "brand refresh." It tells you the buyer is in evaluation, they have a timeline, and they are already planning the launch. The content that answers the compressed keyword and the content that answers the full question are not the same, and most companies have only built the first.
Can you find these questions without a tool?
Partially. Sales call notes, customer support logs, RFP responses, and site search queries all surface real buyer language, and they are worth mining before any tool is involved. The limitation is coverage: those sources reflect the questions buyers ask after they have already found you. The questions asked before they find you, the ones shaping the shortlist, do not appear in your CRM. Surfacing those requires mapping the question space around your category, not just the questions your existing customers have already asked you.
How often do the answers change, and how often should we re-check?
AI answers move week to week. A brand that appears in a given share of answers this month may appear in a materially different share next month without any change on their site, because the sources the model draws on have shifted. Monthly measurement is the minimum cadence that produces a usable signal. A single check is an anecdote. Two checks are a comparison. Several months of checks begin to show whether the direction is improving, declining, or flat, which is the only information that tells you whether the work is working.
What does a post-purchase question look like, and why does it matter?
A post-purchase question is one an existing customer asks after the transaction: at renewal, when evaluating whether to expand the relationship, or when something has gone wrong. "Is [service category] worth renewing for another year?" is a post-purchase question. "What should I expect from my agency in month six?" is a post-purchase question. AI assistants answer these the same way they answer pre-purchase questions, which means a competitor can appear in the answer your existing customer receives when they are deciding whether to stay. Most AI visibility programs are built entirely around acquisition. The post-purchase stage is where retention is won or lost, and almost no company is measuring it.
Do different AI assistants ask different sub-questions behind the same buyer question?
Yes, and the decomposition varies by platform. When a buyer asks a complex question, each assistant breaks it into sub-questions internally before generating a response. The sub-questions it chooses determine which sources it retrieves. Two assistants responding to the same prompt may draw on different source sets because they decomposed the question differently. This is one reason visibility can differ across platforms even when the underlying content has not changed. It is also why tracking visibility on a single assistant and treating that as a complete picture understates the range of sub-questions your content needs to address.
Our competitors are named and we are not. Where do we start?
Start by mapping the questions where your competitors appear. The goal is not to understand why they were recommended instead of you. It is to understand which questions your content is not yet addressing that theirs is. A competitor appearing on a generic category question you have never tracked is a signal about the question, not about the competitor. Once you know which questions are producing visibility for others in your space, you know which questions your content needs to address. The discovery problem comes before the content problem.
How long before a change shows up in AI answers?
Content changes can begin to influence AI citations after reindexing, assuming the content directly addresses a question the model is retrieving sources for. Entity and consistency changes, updating listings, directories, and third-party descriptions, take longer because those sources update on their own schedules and the model's index reflects them with a lag. The honest answer is that there is no reliable single timeline. What is consistent across programs is that sustained monthly measurement is needed before directional conclusions are meaningful. Changes that look like progress early sometimes reverse before the pattern stabilizes.
What does a question inventory cost, and what do we get?
A question inventory maps the question space around what you sell: the full estimated universe of questions buyers ask about your category, labeled by topic, by branded or generic, and by buying stage. It shows you which questions are available for measurement, which are being asked at the moment of decision, and which your current content does not address. It is the starting point for an AI visibility program, not the program itself. The cost depends on the scope of the category and the number of service areas being mapped. The output is a structured view of the demand you are currently missing and a prioritized set of questions to address first.
We already rank well on Google. Does any of this apply to us?
Yes, for a specific reason. Ranking in Google and appearing in AI answers are different outcomes produced by different signals. A site can hold top positions in traditional search while being absent from AI answers on the same topic, because the questions buyers ask assistants are often phrased differently from the keywords the site ranks for, and the content that answers one is not always the content that answers the other. Strong Google rankings are a useful foundation. They do not substitute for understanding which questions buyers are asking AI assistants about your category, because those questions may not map to the keywords your current rankings were built around.


