Six symptom patterns, six different causes. Diagnose from what you observe rather than from a theory about what is wrong.
Absence has symptoms, and the symptoms tell you the cause. A brand that is never mentioned has a different problem from one described inaccurately, which has a different problem again from one named only when the user names it first. This is a symptom-first diagnostic: six patterns of absence, what each one indicates, and the specific fix. Start with what you observe rather than with a theory about what is wrong.
Most brands arrive with a conclusion already formed — usually that they need more content, occasionally that something is technically broken. Both are sometimes right and neither is a diagnosis. The productive starting point is what you can actually observe: what happens when you and your buyers pose real questions to the model.
The tests take under an hour and they discriminate well, because the six patterns are genuinely distinguishable. A model that describes you inaccurately is doing something different from one that omits you entirely, and the underlying failures have almost nothing in common. Running the tests first is what prevents a year of well-executed work aimed at the wrong problem.
Pose four questions across each engine your buyers use, three times each because answers vary. First: who is [your brand]. Second: the category question, without naming yourself — what are the best options for [the problem you solve]. Third: the same question narrowed to your specific segment or use case. Fourth: a direct comparison between you and a named competitor.
Record what comes back. The combination of answers across the four questions places you in one of six patterns, and the pattern identifies the cause. What you are looking for is not whether the model likes you but which specific capability it lacks: identification, association, corroboration, or preference.
Six symptom patterns of AI invisibility, each pointing at a distinct cause. Identify your pattern from the four-question test before selecting any remedy.
Total silence
The model does not recognise the brand name at all, or returns nothing. Usually access is blocked or the brand has almost no independent footprint.
Cause: retrievability or complete absence of evidence.
Wrong brand
The description returned is about a different organisation sharing your name or operating adjacently.
Cause: entity collision. Resolution failure.
Vague or dated
You are recognised but described generically, or with facts that are years out of date.
Cause: weak or stale independent description.
Prompted only
Accurate when named; never raised when the category is the question.
Cause: no independent category association. The most common pattern.
Intermittent
Named on some runs and not others, for the same question.
Cause: boundary candidate. Evidence present but thin.
Named but ranked last
Present in the set, consistently framed as the weaker option.
Cause: preference. No documented reason to choose you.
If the model cannot say anything about your brand at all, the cause is almost always one of two things: its crawlers cannot reach your content, or nothing independent has ever been written about you. The first is a configuration problem and takes an afternoon. The second is a genuinely early-stage position and takes quarters.
Distinguishing them is straightforward. Check your robots directives per crawler and your server logs for arrivals. If the crawlers are being refused, that is your answer. If they are arriving and the model still has nothing, the content exists and the evidence does not, which is a different and slower problem.
Being confused with another organisation is more common than most brands expect, particularly with common-word names, recent rebrands, or a competitor operating in an adjacent market under a similar name. It is also the most damaging pattern, because every piece of evidence you accumulate may be attaching to the wrong entity.
The remedy is entity disambiguation: consistent full-form naming in formal contexts, structured data declaring your specific identity with links to authoritative profiles, and pairing your name with your category and location wherever you are described. It is cheap and it unblocks everything above it, which makes it the highest-priority fix when it appears.
Symptom to cause to remedy
Four-question test at the root, branching by answer combination into the six patterns. Each terminal node states the cause, the remedy, and a realistic timeline. This is the article’s primary artefact and should be built to be printed and used.
Prompted-only presence — the model describes you accurately when you name yourself and never raises you when the category is the question — is the pattern we see most, and it is the one most reliably misdiagnosed. From the inside it looks like the model does know you, which it does, so teams conclude the problem must be content quality or volume.
It is neither. The model has an accurate picture of what you are and no independent basis for offering you as an answer, because the only party connecting you to your category is you. That is an evidence problem with an off-site remedy, and no amount of publishing on your own domain addresses it. Recognising this pattern correctly probably saves more wasted budget than anything else in this article.
If the model describes you correctly but never volunteers you for your own category, you do not have a content problem. Nobody independent has stated which category you belong to — and that is not fixable on your own website.
Being named on some runs and not others for the same question means you are a boundary candidate: the evidence supports including you but not reliably enough to make you a consistent choice. This is genuinely good news, because it means the lower failures have cleared and the constraint is depth rather than kind.
The remedy is more of the same evidence rather than a different type of work: additional independent sources placing you in the category, greater consistency across them, and clearer documentation of what you are specifically good at. Brands in this pattern typically move to consistent presence within two to three quarters of sustained work, which is the fastest of the upper-pillar remedies.
Appearing in the set while consistently framed as the weaker option is a preference problem, and it is the one closest to a genuine competitive assessment. The model has evidence about you and evidence about the alternatives, and what it has read gives it a reason to prefer them.
The productive response is rarely to contest the general comparison. It is to find the specific dimension where the evidence genuinely favours you — a segment, a use case, a constraint — and get that documented independently. A model answering “best for X” needs a reason, and being the documented best for a narrower X beats being the arguable second-best for a broad one.
The tests are reproducible but only useful repeated. DUNkē tracks how you are named and described across eight AI engines — per prompt, against competitors — so the pattern is measured rather than sampled once.
Every one of these six patterns looks identical from the outside: the brand is not showing up. That is why the default response — produce more content — is applied to all of them, and why it works for approximately none. Content addresses extractability and coverage breadth, which are the constraint in only a minority of cases.
The cost is not just the wasted spend. It is the conclusion teams draw afterwards, which is usually that AI visibility is unmeasurable or not worth pursuing. A correctly diagnosed brand that fixes an entity collision in a fortnight and sees its description sharpen has a completely different experience of this discipline than one that publishes for a year against an association failure.
The remedies map cleanly. Patterns one and two are technical and identity work — days to weeks, internal, cheap. Patterns three and four are evidence work — quarters, external, and dependent on parties you do not control. Pattern five is more evidence of the same kind. Pattern six is positioning made public.
Work only the pattern you have. The most common error after a correct diagnosis is to address the identified cause and simultaneously start everything else, which dilutes effort and makes it impossible to tell what worked. Fix the one thing, re-measure, and let the result validate or falsify the diagnosis before expanding.
No single question identifies the pattern; the combination does. If the direct identification is accurate but the category question never surfaces you, that is pattern four. If the identification is accurate and the narrowed segment question does surface you while the broad one does not, you are a boundary candidate on the general question and secure on the specific one — which is a more advanced position than it feels.
If the comparison question frames you as the lesser option while the category question includes you, that is pattern six. And if the direct identification is about somebody else, nothing further matters until that is resolved. Reading the four together rather than reacting to any one is what makes the diagnostic reliable, and it is why the test is four questions rather than one.
Generated answers vary between runs, which produces a specific organisational problem: two people testing the same question get different results and reach different conclusions. One reports the brand is visible, another reports it is not, and the disagreement is treated as a measurement failure rather than as the finding it actually is.
The correct interpretation is that variability itself is data. Consistent presence and consistent absence are both stable states; alternation is a boundary state with its own diagnosis. Running each question at least three times and recording a rate rather than a verdict resolves the confusion and produces a more useful reading than either individual observation.
One outcome deserves special warning because it is routinely misread as a win: being described accurately and favourably when named. Teams see a flattering paragraph about their brand and conclude AI visibility is fine. What they have observed is that the model can retrieve information about a brand it was handed, which almost every established company clears.
The commercially relevant test is the one where you do not supply your name. Buyers asking assistants which options to consider are not typing your brand into the prompt — that is the entire problem. A brand that only appears when introduced has confirmed it is known and has learned nothing about whether it is recommended.
Because collisions block everything above them, the remedy deserves detail. Use your full legal or complete trading name consistently in formal contexts rather than the shortened version. Pair the name with your category and location wherever you are described, since that pairing is what disambiguates. Declare the entity in structured data with links to profiles that unambiguously belong to you.
Then work the third-party record: correct entries that describe the other organisation under your name, and ensure the sources most likely to be retrieved carry the disambiguating detail. This takes weeks rather than quarters and it is the highest-return fix available when the pattern applies, because every subsequent piece of evidence then attaches correctly.
Symptom, cause, remedy, timeline
Six rows, one per pattern. Columns: what you observe across the four questions, the underlying cause, the specific remedy, the realistic timeline, and the owning function. Designed as a one-page reference to sit alongside the decision tree.
Patterns three, four and five have quarter-scale remedies, which raises a practical question about the intervening months. The answer is to do the fast work that will amplify the slow work when it lands: ensure resolution is clean so accumulating evidence attaches correctly, and restructure key pages for extractability so that when you do become a candidate, your passages are usable.
Neither produces citation on its own in these patterns, and both increase the return on the evidence work when it arrives. That sequencing — do the amplifiers while the slow thing accrues — is more productive than either waiting or starting a content programme that addresses a constraint you do not have.
Running the diagnostic on a single engine produces a partial answer, because engines draw on different sources and resolve entities differently. A brand can present as pattern four on one engine and pattern three on another, which is informative: it usually indicates the underlying evidence exists but is concentrated in sources one engine favours and another does not.
Where the patterns differ sharply across engines, the constraint is usually source-mix rather than evidence volume. Where they are identical across engines, the constraint is the evidence itself. That distinction is only available by testing more than one engine, which is the main argument for the extra time it takes.
Occasionally the diagnostic returns a result worth acting on differently: the category is barely queried conversationally, the engines produce thin or uncited answers for the relevant questions, and the brand’s absence reflects an absence of demand rather than a failure of visibility.
That is a legitimate outcome and it should stop the programme rather than redirect it. Building AI visibility for questions nobody asks produces accurate measurements that go nowhere. Checking whether the category has moved into the answer layer at all is a five-minute precondition to the whole exercise, and it is skipped almost universally.
Each pattern lands with a different function, which is the practical reason diagnoses stall after they are made. Patterns one and two belong to engineering and technical marketing. Pattern three splits between whoever maintains third-party profiles and whoever owns communications. Pattern four belongs almost entirely to PR and brand. Pattern six belongs to product marketing and positioning.
Only one of those sits with the team that usually commissions the investigation. Establishing at the outset that the answer may belong elsewhere — and that the finding will need to travel — substantially improves the odds of anything happening. A correct diagnosis handed to a function with no mandate to act on it produces a well-informed absence.
Re-testing too soon produces noise and too late wastes months. The workable cadence follows the remedy: for access and entity fixes, re-test after two to four weeks, since the constraint was mechanical and resolution should follow the next crawl and re-index cycle. For evidence work, re-test quarterly, because nothing meaningful accumulates faster than that.
What to watch between tests is description quality rather than presence. Descriptions sharpening from generic to specific is the leading indicator that evidence is accumulating, and it appears before naming does. Teams that know to look for it stay funded through the lag; teams watching only for presence conclude nothing is happening for two quarters and stop.
Running the same four questions about a competitor is unusually informative and takes no extra tooling. It tells you which pattern they are in, which reveals whether their advantage is structural or evidential, and whether it is one you could plausibly close.
A competitor in confident presence with deep independent coverage is a long project. One who is merely resolvable and associated while you are not is a gap closeable within a year. And a competitor absent from the answers while ranking well is a signal that the whole category is still contested, which changes the urgency calculation considerably.
The diagnostic as described is observational, and its obvious weakness is that it establishes correlation between symptom and cause from audit experience rather than from controlled comparison. A stronger version would take brands presenting each pattern, apply only the indicated remedy, and measure whether the pattern resolves as predicted.
That is a design we can describe and have not run at the scale required to publish. Readers should treat the six patterns as a reproducible triage that has proven reliable in practice, not as a validated instrument — and should treat the falsifiable prediction attached to each remedy as the mechanism by which they can check it themselves.
It is worth quantifying, in effort rather than currency, what a wrong diagnosis costs. A content programme aimed at an association failure consumes a quarter of production capacity and produces no citation movement, because it adds to the side of the evidence that was never the constraint. A technical audit run against a preference problem finds nothing wrong and concludes the measurement must be faulty.
In both cases the second-order cost exceeds the first. The team concludes the discipline does not work, the budget moves elsewhere, and the actual constraint remains unaddressed for another year. An hour spent on the four-question test is the cheapest insurance available against that sequence, and it is skipped almost universally because the conclusion feels obvious before the test is run.
A mid-market B2B brand arrives convinced it needs more content. The test returns: accurate identification, no appearance on the category question, appearance on the narrowed segment question, and a competitor framed as the default in the comparison. That combination is pattern four moving toward pattern five — associated in a niche, unassociated in the category, present at the boundary.
The correct programme is not content. It is getting the category association documented by sources outside the niche where it already exists, and getting the specific strength that earns the segment appearance stated independently. That is a communications brief with a two-quarter horizon, and it is roughly the opposite of what the organisation was about to commission.
We do not claim the six patterns are exhaustive, that they are mutually exclusive, or that they occur in any measured proportion. Brands present combinations, and the diagnostic works by identifying the dominant one rather than by producing a clean classification.
We also do not claim the causal attribution is proven. The link between each symptom and its stated cause is inference from audit experience, tested by whether the indicated remedy resolves the symptom — which it usually does, and which is evidence rather than proof. The falsifiable prediction attached to each remedy is what lets a reader check the claim rather than accept it.
If you do only one thing after reading this, ask an assistant your category question without naming your brand, three times, and write down who it names. That single act converts an abstract concern into a specific competitive fact, and it is the observation everything else in this article depends on.
Most teams have never done it. They have asked about themselves, received a flattering paragraph, and drawn the wrong conclusion. The category question is the one your buyers actually ask, and the list it returns is your real competitive set in the answer layer — which is frequently not the set on your competitor slide.
How long until we see change after fixing the cause?
It tracks the remedy. Access and entity fixes can show in weeks because the constraint was mechanical. Association and corroboration take quarters because they depend on third parties publishing and on stale descriptions ageing out.
We block AI crawlers for content protection. Does that explain it?
Entirely, for the engines you block. Blocking removes you from those answers without recovering the traffic those answers displaced, which is a trade that only makes sense where the content itself is the product being sold.
Should we ask customers to mention us in their content?
Encouraging genuine advocacy is fine; orchestrating coordinated mentions is not, and communities detect it. What matters is that independent sources place you in your category for their own reasons, which is what makes the evidence worth anything.
Does this differ across engines?
The patterns are the same; which engine shows them can differ, because engines draw on different sources. Running the tests across each engine your buyers use is worth the extra time, and a brand present on one and absent on another usually has an access or resolution difference rather than a content one.
Absence from AI answers is six distinct conditions with six distinct causes, and they are distinguishable by a four-question test that takes an hour. Total silence points at access or a genuinely empty evidence base. Being described as another company is an entity collision. Vague or dated descriptions mean the independent record is thin or stale. Prompted-only presence — the most common pattern — means nobody independent has connected you to your category.
Intermittent naming means you are a boundary candidate and need depth. Being named but positioned last is a preference problem answered by specificity rather than by contesting the general comparison. Every one of these looks like “we need more content” from the outside, which is why the diagnosis matters more than the remedy: the remedies are known, and applying the wrong one costs quarters.
Methodology note: the six patterns are drawn from audit observation, not a measured population study. We state no proportion for how often each occurs, though we describe pattern four as most common on the basis of qualitative frequency in our own work. Readers should treat the diagnostic as a reproducible test rather than as a validated instrument.
“Every version of this problem looks the same from the outside — we are not showing up. Six different things cause it, and only one of them is fixed by writing more.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.