LLM Visibility: Present or Absent, Not Ranked

It is binary, and that changes everything

Your name is either in the answer or it is not. There is no eleventh place, no gradual climb, and no position to defend. That single property invalidates most of the intuition carried over from ranking.

Under ranking, effort produces visible incremental progress. Moving from twenty to twelve is real, reportable and encouraging even though it delivers almost no clicks. The feedback arrives early and continuously.

Here there is no partial credit. You do a quarter of work and you are still absent, right up until you are not. That is genuinely harder to fund and harder to stay committed to, and it is the main reason programs in this area get abandoned two months in.

The partial credit that does exist is in frequency rather than position. Across many askings of many questions, your name appears some proportion of the time, and that proportion can move. That is the only continuous quantity available, and it is what any sensible measurement watches.

Why the same question gives different answers

Anybody who tests this properly discovers within an hour that asking twice can produce different names, and that fact has to be built into everything downstream.

The process is not deterministic. Two runs of an identical question can gather overlapping but different material and can select differently from it. This is a property of the system rather than a fault in your test.

The consequence is that a single reading is a sample, not a measurement. A business that asks once, sees a competitor named, and reorganizes a quarter around it has built a plan on one draw from a distribution.

It also means improvement is probabilistic rather than binary at the level of the program even though it is binary at the level of one answer. You are not switching yourself on. You are increasing how often you come up, which is a slower and less satisfying thing to watch.

Practically, ask each question at least several times across several days before recording anything, and treat the spread as data rather than noise. A question that names you three times in ten is telling you something quite different from one that names you never.

Three things it is not

Clearing away three common misreadings makes the rest of this considerably easier.

It is not traffic. Being named frequently does not necessarily produce visits, because the answer often satisfies the question. Presence and sessions are different quantities and conflating them leads to a disappointing conversation with whoever holds the budget.

It is not brand awareness. Being named in an answer is a mechanical outcome of being reachable, quotable and deep on a subject. A widely known brand with a thin site on a subject will lose to an obscure specialist with a thorough one.

It is not a score. A single percentage is a summary of whichever questions somebody chose. Change the questions and it changes without anything about you changing. Useful watched over time against a fixed set, meaningless as a verdict.

How to measure something with no position

The measurement problem here is real and it is solvable, but not with tooling you already own.

Nothing in your analytics will show this. A person who reads your name and then searches for you arrives as direct traffic with no trace of the assistant. There is no referrer to read and no attribution to build, so the event has to be observed at the source.

That means asking. Fix a set of between six and ten questions in the language a buyer uses rather than the language you use about yourself. Ask each one several times. Record whether you were named, what was named instead, and which sources were credited where the assistant shows them.

Then repeat on a schedule and watch the frequency rather than any individual answer. Monthly is usually right. Weekly measures noise, quarterly loses the ability to connect a change to a cause.

Take the first reading before you change anything. This is the item most often skipped and the only one that cannot be produced retrospectively, which is why programs so often end unable to say whether they worked.

What actually moves it

Three things move this, in a fairly reliable order of cost.

Being reachable. The cheapest and most binding. Answer crawlers are separately named clients governed by rules frequently set by somebody solving an unrelated problem. A site can be entirely open to search and closed here, which produces total absence with no other symptom.

Being quotable. A conclusion in the first two sentences and claims that survive being lifted out of their paragraph. Days of writing, weeks of approval, and it improves the pages for readers too.

Being deep on one subject. The expensive one. Reaching for you is a judgment about whether you are a reasonable source on this subject, and one page is thin evidence however good it is. This runs on quarters and does not compress.

What does not move it is volume across many subjects, keyword repetition, or writing pages in a supposed machine dialect. All three are commonly attempted and none of them addresses any of the three gates.

What a realistic result looks like

Knowing the shape of a genuine improvement stops you reading noise as success or success as failure.

The first thing to move is questions that name you directly. These come back accurate rather than empty or garbled, which is a low bar and a genuine signal that retrieval is working.

The second is specific long-tail questions matching pages you rewrote. Fewer sources address these precisely, so they are the easiest to win and the earliest evidence that an editorial change landed.

The last, if it moves at all, is the broad category question with the most obvious phrasing. That one is contested by everybody and usually answered from long-established sources. Treating it as the measure of success means concluding nothing worked for several quarters while a great deal is working.

Track those three groups separately. A single average across all of them moves slowly and tells you almost nothing about which of your changes mattered.

The honest trade-off

This is a channel with no attribution, non-deterministic measurement, no position to report, and a payoff that arrives in steps rather than gradually. Those are real reasons to be cautious and they should be said plainly rather than buried.

Against that, the work that moves it is largely work that makes a site better regardless. Being fetchable, stating conclusions clearly, and covering one subject properly are all defensible on their own terms. The failure case leaves you with a clearer, deeper site rather than a wasted quarter, which is unusual.

The genuine risk is opportunity cost. A quarter spent on depth in one subject is a quarter not spent elsewhere, and if the subject turns out not to be what buyers ask about, the work was done correctly and bought nothing. Deriving the subject from language customers actually used is the only real mitigation.

There is a second risk that gets less attention, which is the reporting relationship. Because this cannot be attributed, whoever runs it is asking to be trusted on a measurement they collected themselves, and that is an awkward position to hold for several quarters. The way to survive it is to fix the question set publicly at the start, publish the readings whether they are flattering or not, and never present a single answer as evidence. A program that only reports the good months will not be believed in the month it needs to be.

Reading it against competitors

The instinct on seeing a rival named is to study that rival, and it is usually the wrong first move.

What a competitor's presence tells you is that the question is answerable and that somebody cleared all three gates for it. It does not tell you which gate you are failing, and that is the only thing you need to know. Studying their pages when your own crawler access is refused teaches you nothing you can act on.

The genuinely useful comparison is the set of sources being named across many askings, not one competitor in one answer. If the same handful of names keeps appearing regardless of phrasing, that is a category with entrenched sources and your realistic move is a narrower subject rather than a frontal attempt.

If instead the names shift substantially between phrasings, the category is unsettled and there is room. That is a considerably more useful read than any individual competitor's page structure, and it takes an hour to establish.

Where competitor study does earn its place is in the selection gate. If they are quoted and you are retrieved but never used, comparing the actual quoted passage against your equivalent paragraph is direct, specific feedback of a kind this discipline rarely offers. Our page on benchmarking covers choosing the comparison set properly.

Which businesses this suits, and which it does not

This is not equally worth pursuing for everybody, and the honest version of that is worth stating.

It suits businesses whose buyers research before purchasing, particularly where the research involves comparing approaches rather than comparing prices. Those questions are exactly what people put to assistants, and a specific, honest source gets reached for.

It suits specialists more than generalists. Depth on one subject is the binding factor, and a business that genuinely knows one narrow thing well is better placed than a larger one covering everything moderately.

It suits businesses that can say something only they can say. If your pages contain nothing that could not appear on a competitor's site, the selection gate will not open however good the structure is, and no amount of technical work substitutes for that.

It suits you less if your buyers do not research, if your category is dominated by sources decades older with far more written about them, or if you need attributable traffic this quarter. All three are legitimate reasons to spend the money elsewhere, and a supplier who does not raise them is not being straight with you.

Four mistakes this framing prevents

Understanding that this is binary and non-deterministic prevents four expensive decisions.

Reporting it as a rank. There is no position, so a report showing one has invented it. What can honestly be reported is frequency across a fixed question set.

Reading one answer as evidence. The same question can answer differently between runs, so a single reading is a draw from a distribution rather than a finding about your business.

Abandoning at two months. Because there is no partial credit at the level of a single answer, a program genuinely working can look like nothing happening. The frequency across many questions is where early progress shows, and only if somebody set it up to be visible.

Expecting traffic. Presence and sessions are different quantities. An answer that satisfies the question produces no visit, and a program sold on traffic will be judged a failure even when it did exactly what it should. Our page on visibility monitoring covers setting the measurement up so this does not happen.

Bottom line

Stop thinking in positions. There is no rank here, only present or absent, and the only continuous quantity is how often you are named across a fixed set of questions asked repeatedly.

Measure by asking, start before you change anything, and ask each question several times because the same question can answer differently. A single reading is a sample.

Then work the three gates in cost order. Reachable first because it is cheap and binding, quotable second because it is fast, deep third because it is the one that actually decides the outcome and the one that takes a quarter.

And decide before you begin what result would count. A program without that agreement will be judged against whatever the person asking happens to expect, which in this channel is usually traffic, and traffic is not what this produces. Being named more often across questions your buyers actually ask is the outcome. If that is not worth something to the business on its own, this is the wrong quarter to spend here and it is better to know that in advance than in month five.