Know Your Customer When They Research Through AI

Atamyrat Hangeldiyev

Systems Architect

February 11, 2026

The transcript was four exchanges long and the owner of the regional HVAC company had never seen it: a facilities manager had asked an assistant "is it risky to hire a smaller HVAC contractor for a multi-site commercial account, or do I need one of the national ones", got an answer about insurance limits, bonding, and how to check a contractor's commercial track record, asked a follow-up about what to look for in references, and only then searched, landed on three sites, and filled in one form. By the time that form arrived the decision was substantially made, and it had been made against a checklist the owner did not know existed, on a question the owner was never asked, in a conversation the owner could not see. The form said "interested in a quote". The transcript said the buyer had already decided this company was either credible or not before the site ever loaded.

Knowing your customer is the documented understanding of who the buyer is, the job they are hiring you for, and the questions they actually ask before they decide, including the research they now do inside an AI assistant where you never see it, in the context of small and mid-sized businesses with no marketing team. It is not demographics, not a persona slide with a stock photo, and not a survey. It is a written, testable picture of a specific buyer and a specific job, built from what buyers do and ask rather than what an owner assumes, and the part of it that used to be invisible is now most of it.

What knowing your customer actually is, and what it is not

Customer knowledge has three parts, and all three have to be written down to count. The first is who the buyer actually is: not their age and company size, but their situation, what triggers them to look, and what they are afraid of getting wrong. The second is the job they are hiring you for, stated as the outcome they want and the thing that would make them regret the choice. The third is the set of questions they actually ask before they decide, in their words, including the ones they ask an assistant in private and would never ask a salesperson. Hold those three and you can predict what a buyer will respond to. Hold a persona slide instead and you can only predict what a persona slide says.

The reason "written down" is load-bearing is that an owner who talks to customers every day genuinely believes they know the customer, and that belief is mostly memory of the easy conversations. The buyer who said yes is vivid. The buyer who quietly went elsewhere after asking an assistant one question never became a memory at all, because that buyer never made contact. Customer knowledge is the discipline of replacing the remembered customer with the documented one, including the buyers who never reached you.

Customer knowledge stops at the buyer picture: the ICP and the buyer research, done honestly, with the AI-mediated research step reconstructed. The words you say back to the buyer are covered in the value proposition guide. How that picture sharpens your position is covered in the positioning guide. How you get surfaced inside the assistant in the first place is covered in the AI-visibility guide. Each is linked further down at the point it connects. What follows builds the buyer picture itself, because the message, the position, and the visibility all depend on it being true.

The documented buyer, their job-to-be-done, and what they actually ask

The job-to-be-done is the part owners get wrong most often, because the obvious answer is the product. A two-location dental group does not think of itself as hiring a marketing firm; the office manager is hiring something that stops the schedule from having empty chairs on Tuesdays without making the practice look discount. A B2B parts distributor's customer is not hiring "parts"; a maintenance lead is hiring the certainty that the line will not be down waiting on a part nobody can confirm is in stock. The product is what you sell. The job is what they are trying to get done, and the job is where the buying decision is actually made. Two buyers with identical demographics and different jobs will make opposite decisions, and a picture built on demographics will not see it coming.

What they actually ask is the third part and the one that changed. Before assistants, the questions a buyer asked were mostly the ones they were willing to ask a person: scoped, slightly performed, aware they were talking to someone selling. The questions a buyer asks an assistant are the unscoped ones, the ones they would be embarrassed to ask a salesperson, the ones that decide whether you are even a candidate: "is this kind of company trustworthy for X", "what goes wrong when a small firm does this", "how do I tell a good one from a bad one without knowing the field". Those questions are the real buyer research, and they happen before contact. Customer knowledge that does not include them is a picture of the buyer after they have already filtered themselves.

An example: the assumed customer and the real one side by side

Take a niche industrial-supply shop that sells specialized fasteners and fittings to fabrication shops. Here is the customer the owner described, and the customer the work surfaced, changed in nothing but accuracy.

The gap between those two columns is not a detail. The assumed customer responds to a price-and-stock message. The real customer responds to proof of certification and a way to verify it, and ignores the price message entirely until that fear is handled. An owner working from the left column would spend a year competing on price against a buyer who was never deciding on price, and would never know why the close rate stayed flat.

The customer you think you have is not the one deciding

The assumed customer and the real one diverge by default, not by accident, and the divergence is structural. An owner's mental model of the customer is built from a biased sample: the buyers who made contact, the deals that closed, the complaints loud enough to reach the owner. It systematically excludes the buyers who looked, got an answer somewhere the owner could not see, and silently chose someone else. That excluded group is not a rounding error. For most SMBs it is the larger group, and it is the group whose decision logic the owner most needs and least has.

The cost of the divergence is specific and compounding. Every message, every page, every sales conversation, and every channel choice is aimed at the assumed customer. If the assumed customer is wrong about the job, the entire output is precisely aimed at the wrong target, and effort does not fix aim. A business can run a disciplined content program, a clean site, and a tight sales process, all of it pointed at a buyer who is deciding on something the business never addresses, and the result is consistent mediocre performance that looks like a marketing problem and is actually a customer-knowledge problem. You cannot out-execute a wrong picture of who is deciding.

Where the assumed and real buyer diverge, and what it costs

There are four reliable divergence points, and naming them tells you where to look. The first is the job: the owner names the product, the buyer is hiring an outcome, and the outcome is usually about avoiding a specific regret rather than gaining a specific upside. The second is the decider: the owner pictures the title on the org chart, the actual decision is shaped by the person who carries the blame if it goes wrong, and that is often not the same person. The third is the trigger: the owner thinks buyers shop when they need the product, but buyers usually start when something failed, someone got criticized, or a deadline appeared, and the emotional state at the trigger changes what they respond to. The fourth is the question set: the owner answers the questions buyers ask out loud, and the decision turns on the questions they ask in private.

Each divergence has a price tag. A wrong job means the message is true and irrelevant. A wrong decider means the message reaches someone with no skin in the decision. A wrong trigger means the marketing arrives when nobody is looking and is absent when they are. A wrong question set means the buyer's actual deciding question is answered by a competitor, or by the assistant, and not by you. None of these show up as an obvious failure. They show up as a conversion rate that will not move and a sales team that says the leads are weak.

The research step you never see because it happens inside an assistant

The research step that decides candidacy now happens before any analytics tool can see it, because it happens inside an assistant. A buyer with a problem opens Claude or a similar assistant and asks the unscoped question first: not "best HVAC contractor near me" but "we are a small property manager, is it a mistake to use a regional HVAC contractor instead of a national one for multiple sites, and how would I tell". The assistant answers with criteria. Those criteria become the buyer's checklist. By the time the buyer searches, clicks, and possibly contacts you, they are not forming an opinion; they are checking you against a checklist they built in a conversation you will never read.

This is why a modern assistant matters as a force in the buying decision, not as a tool you use. Claude models, accessed through the Claude API or a Claude assistant, are now where a meaningful share of pre-purchase research actually happens for SMB buyers: the buyer describes their situation in plain language, asks what could go wrong, asks how to evaluate a vendor, and gets a structured answer that frames the rest of their search. You do not control that answer. You can only influence whether your kind of business shows up well in it, and you can only do that if you know what the buyer is asking. Without that, the rest of marketing is optimizing for a search behavior the assistant already pre-empted: you are tuning the visible half of the funnel while the decision gets made in the half you cannot see.

How to define your real ICP and do buyer research that is not theater

The procedure has three parts and they go in order: define the ICP from who you actually serve best, replace stated preference with observed behavior, and reconstruct the assistant questions the buyer asks before contact. None of it requires a research budget or a tool. It requires being honest about which customers you are good for, willing to look at behavior instead of opinion, and willing to write the assistant questions down even though you cannot see them directly. Done in this order it produces a buyer picture you can hand to the value proposition, the positioning, and the channel work without re-deriving it each time.

Define who you serve best, the job, and the disqualifiers

Start from your own best customers, not from the market. The ICP is not the largest possible audience; it is the description of the buyer you serve better than anyone else, where the work is profitable, the relationship is good, and the result is something you can stand behind. The fastest way to find it is to look at the customers you already have and ask which ones you would clone if you could. Then ask what those customers have in common that is not demographic: the same trigger, the same job, the same fear, the same constraint. That commonality is the ICP. The demographic facts are how you describe it afterward, not how you find it.

The disqualifiers matter as much as the qualifiers and are usually missing. A real ICP states who you are not for, in plain terms: the job you are bad at, the buyer whose constraint you cannot meet, the situation where you are the wrong choice and saying so is the honest answer. A niche industrial-supply shop that is excellent for certified critical-line material and mediocre for bulk commodity orders should say that, because the buyer who needs commodity bulk is a buyer it will lose anyway and lose expensively. An ICP with no disqualifiers is a wish, not a definition. Write the ICP as three short statements: who you serve best, the job they hire you for stated as the outcome and the regret, and who you are explicitly not for and why.

Watch behavior, not stated preference: what to look at and what to ignore

People are reliable about what they did and unreliable about why, and customer research that forgets this measures the unreliable part. What a buyer says they want in a survey is a theory of themselves, often the flattering one. What they actually did, which competitor they chose and what they said in the moment they were deciding, is the data. The instruction is narrow: treat stated preference as a hint to investigate and treat observed behavior as the conclusion. When the two disagree, the behavior is right.

There are five sources of real behavior an SMB already has and usually ignores. Sales-call notes and lost-deal reasons, in the buyer's own words, not the sales rep's summary. The exact questions buyers ask before they buy, collected verbatim over a month from email, calls, and forms. The reasons customers actually gave for choosing you, asked in the first week when the memory is fresh and the answer is not yet a story. The reasons buyers gave for choosing a competitor, which requires asking the ones who said no and is worth more than any won-deal data. And the language buyers use for their own problem, which is almost never the language the business uses for its product. None of this is a research project. It is reading what you already have with the question "what does the behavior say" instead of "what do I already believe".

What to ignore is equally specific. Ignore survey questions that ask buyers to predict their future behavior or rank hypothetical features, because the answers are confident and wrong. Ignore the loudest customer when they are not representative, because volume is not signal. Ignore your own explanation of why a deal was lost until you have the buyer's, because the internal story is almost always kinder to the business than the truth. Ignore demographic correlations that have no mechanism, because age and company size correlate with plenty of things and explain none of them.

Reconstruct the assistant questions: what the buyer asks an AI before they reach you

You cannot see the buyer's assistant conversation, so you reconstruct it from what you can see. The method is direct. First, collect the unscoped questions buyers ask once they do reach you, the ones that start with "is it normal that" or "how do I know if" or "should I be worried about", because those are the same questions they already asked the assistant, just asked again to confirm. Second, take your own ICP's job and fear and write the five questions a person in that situation would ask an assistant before they would trust anyone selling, phrased the way a worried non-expert phrases things, not the way the industry phrases them. Third, and this is the part most owners skip, actually run those questions through a capable assistant the way the buyer would, and read the answer the buyer would have read.

That third step is uncomfortable and necessary, because the answer the assistant gives is the checklist your buyer arrives with. Open Claude, describe the buyer's situation in the buyer's words, ask the buyer's deciding question, and read what comes back: the criteria it lists, the warnings it gives about small vendors in your category, the way it tells the buyer to evaluate someone like you. For agentic research at scale, Claude Code can run a structured pass over many such buyer questions and collect the patterns, but the manual version is enough to start and you should do the manual version first because reading one real answer changes how you see the buyer more than a report ever will. The output of this step is a written list of the questions your buyer asks an assistant and the checklist the assistant hands them. That list is the part of customer knowledge that did not exist five years ago and now governs whether you are considered.

Customer knowledge versus the things it gets confused with

Customer knowledge gets impersonated by three things that feel like it and are not, and one thing it is actually made of. An owner with a persona slide, a demographic profile, and a recent survey can sincerely believe the customer is known, and be wrong on all three at once. Each of the three near-neighbors has a real but narrow use; the failure is mistaking the narrow use for the whole, and the fix is knowing exactly what each one is and is not for.

Customer knowledge vs demographics

Demographics are descriptive labels: age, company size, industry, revenue band, location. They are useful for one thing, deciding roughly where to spend money to reach people, and useless for the thing owners use them for, predicting what a buyer will decide. Two fabrication shops of the same size in the same region can be opposite customers because one foreman was burned by a bad substitution and the other was not. Demographics describe the container; the job and the fear are the contents, and the decision is in the contents. Use demographics to estimate reach and segment spend. Do not use them to explain a decision; they cannot, because they do not contain the job.

Customer knowledge vs persona theater

Persona theater is the named, stock-photo persona: "Marketing Mary, 42, enjoys yoga and efficiency". It is a prop that produces the feeling of knowing the customer without any of the substance, and it is dangerous precisely because it is satisfying. A persona is only worth anything if every line in it is a documented behavioral fact with a source, which the stock-photo kind never is. The honest version of a persona has no photo, no invented name, and no hobbies; it has the job, the trigger, the fear, the disqualifier, and the actual questions, each traceable to something a real buyer did or asked. If a line in your persona cannot be traced to evidence, it is theater, and theater aimed at a buyer is just confident guessing with a headshot.

Customer knowledge vs surveys

A survey captures what people say, which is a theory of their own behavior and usually the flattering theory. It has one legitimate use: generating hypotheses about language and concerns that you then verify against behavior. It has one dangerous use, the common one: treating the answers as conclusions. "Would you pay more for faster delivery" gets a yes from people who then buy the cheaper slower option, because the survey measured the aspiration and the purchase measured the person. Use surveys to find questions worth investigating and the words buyers use for their problem. Never use a survey as the reason for a decision; promote a survey finding to a conclusion only after behavior confirms it, and discard it without guilt when behavior contradicts it.

Customer knowledge vs watching real behavior (the source of truth the others impersonate)

Watching real behavior is the actual source of truth, and the previous three are each mistaken for it because each is cheaper to obtain. Behavior is what the buyer did when it cost them something: which vendor they chose, what they asked before they committed, why the lost deals were lost in the buyer's own words, what the assistant told them and what they did with it. It is harder to collect than a demographic profile, less satisfying than a persona, and less tidy than a survey table, which is exactly why the other three get used in its place. Customer knowledge is the disciplined version of watching behavior, written down and kept current. The other three are inputs to it at best and substitutes for it at worst, and the entire skill is refusing the substitution.

What knowing your customer feeds

A finished buyer picture feeds four downstream pieces of work, and each consumes a different part of it. It feeds the message, which has to be true for the real buyer. It feeds the position, which is sharper when you know exactly who you are best for. It feeds the channel choice, which is only answerable once you know where this buyer actually is and what they are doing there. And it feeds the site and content, which is where what the buyer asks becomes something they can find. Each is a distinct downstream discipline with its own guide; what the buyer picture has to give them is a true description, not a guessed one.

A message is only as true as the buyer it was written for

A message can only be true for a buyer you have actually described. The job and the fear from the ICP are the raw material of a value proposition: the claim has to be about the outcome the buyer is hiring for, and the proof has to answer the fear that would make them regret the choice. Customer knowledge supplies what the message must be true about; it does not supply the message. Turning the documented buyer into a testable claim a buyer believes is its own procedure, covered in how to write a value proposition buyers believe.

A position is only as sharp as the ICP behind it

Positioning is the decision about which buyer you are the obvious choice for, and that decision is impossible without a real ICP and a real set of disqualifiers. The clearer you are about who you serve best and who you are explicitly not for, the sharper the position you can credibly take, because a position is a claim of being the best choice for a specific someone and customer knowledge is what makes that someone specific. Choosing and holding a position a small business can win is covered in positioning: how a small business wins by being specific.

The buyer picture narrows the channels worth running

A small team can run two or three channels well and will run all of them badly if it tries. Which two or three depends entirely on where this specific buyer actually is and what they are actually doing there, including the fact that a large part of their research now happens inside an assistant before they touch any channel at all. The ICP and the reconstructed assistant questions tell you which channels are even plausible for this buyer. Picking and committing to the few a small team can sustain is covered in choosing the two or three channels a small team can run.

How it becomes answers buyers actually find

The last thing customer knowledge feeds is the most concrete: the questions your buyer asks, written down honestly, are the exact questions your site and content should answer in language the buyer used, so that when they search after the assistant conversation, they find you answering the thing they were actually worried about. This is also where the assistant-research step connects to AI-visibility: being the source an assistant draws on when it answers your buyer's deciding question is a strategy of its own, covered in getting your business recommended by AI assistants, and customer knowledge is its prerequisite because you cannot be the answer to a question you have not identified. Doing this well is not a one-time task. Turning what buyers actually ask into answers they find is ongoing content and search work most SMBs do not staff, which is the work behind Iron Goo's SEO service; the customer knowledge in this guide is what makes that work aimed at the right questions instead of guessed ones.

The buyer picture the rest of the pillar is built on

Modern marketing for a small business in the AI era is the work of generating demand and building a brand when the buyer does most of their deciding before you ever hear from them, and every part of that work assumes you know who the buyer is and what they ask. Customer knowledge is not one topic among the pillar's topics; it is the input the message, the position, the channels, and the AI-visibility work all consume, which is why it sits this early in the pillar and why a small team that gets it right can punch above its size. The rest of the pillar is downstream of this one document: get the buyer wrong here and every guide after it inherits the error.

The next move is concrete, not conceptual. Open a document and write the three statements: who you serve best, the job they hire you for stated as the outcome and the regret, and who you are explicitly not for and why. Then collect one month of the verbatim questions buyers actually ask you, and run your buyer's deciding question through a capable assistant the way they would, and read the checklist it hands them. That document is your real ICP, and it is the thing the rest of this pillar builds on. With it written, the value proposition guide gives you the message it makes true, and the positioning guide gives you the position it sharpens; without it, every guide after this one is advice aimed at a customer you have not actually met.