Four pillars, strictly ordered: Access, Identity, Evidence, Expression. Work above a failing pillar produces nothing measurable.
Most AI visibility advice is a list of tactics with no theory underneath, which is why it produces activity without result. What follows is the structure we use instead: four pillars that determine whether a brand can appear in a generated answer, in the order they bind. It is not a checklist. It is a diagnostic architecture, and its main claim is that these four are strictly ordered — effort spent on a later pillar while an earlier one fails produces nothing measurable.
Checklists assume independence: do these twenty things, in any order, and outcomes improve. AI visibility does not work that way, because the requirements are conditional. A model cannot corroborate a brand it cannot resolve, and cannot extract from content it cannot reach. Sequence is not a nicety here; it determines whether work produces any observable effect at all.
That is the practical argument for a structure with an order. It converts an unbounded list of possible activities into a single question — which pillar is currently binding — and the answer to that question makes almost all of the list irrelevant until it is resolved. Most of the value of the framework is in what it tells you not to do yet.
Four pillars determining whether a brand can appear in a generated answer, ordered by dependency. Diagnose the lowest failing pillar; work above it produces no measurable result.
Access
The engine can reach and read you. AI crawler permission, server-rendered content, no technical barrier between the model and your material.
Owner: engineering. Timeline: days. Cost: low.
Identity
The engine can resolve what you are and distinguish you from everything similarly named. Consistent naming, declared entity, clean disambiguation.
Owner: marketing + technical. Timeline: weeks. Cost: low.
Evidence
Independent sources connect you to your category and describe you consistently. Corroboration the engine can rely on when asserting.
Owner: PR, brand, partnerships. Timeline: quarters. Cost: high.
Expression
Your content answers real questions in passages that survive extraction. Answer-first, self-contained, format matched to intent.
Owner: content. Timeline: weeks. Cost: moderate.
Access is binary and cheap, which makes it the right place to start even though it is rarely the binding constraint for established brands. Either the engines’ crawlers can reach your content and read it without executing client-side JavaScript, or they cannot. There is no partial state and no sophistication to it.
It earns its position as pillar one because failure here is total and silent. A brand that blocks AI crawlers by default configuration, or serves its content only after client-side rendering, has forfeited every downstream possibility without any signal that it has done so. Checking takes an hour and eliminates the most embarrassing category of failure.
Identity is the model’s ability to resolve you as a specific entity. It fails far more often than teams expect, and it fails invisibly, because nobody thinks to test whether the systems they want to appear in actually know who they are. Common-word names, recent rebrands, acquisitions, category changes, and collisions with organisations in adjacent markets all degrade it.
The work is unglamorous: describe yourself identically everywhere, declare your entity in structured data, connect it to your authoritative profiles, and correct the descriptions that have drifted. It is cheap, internal, and fast — and it is a hard precondition for everything above, because corroboration has to attach to an entity the model can identify.
The four-pillar diagnostic sequence
Vertical flowchart, one decision node per pillar with its diagnostic test. Failing branches exit to the pillar-specific remedy with owner and timeline. Passing branches continue upward. Terminal state at the top: eligible for citation. This is the framework’s primary artefact.
Evidence is where most brands actually stop, and it is the pillar least amenable to anything a marketing team can execute alone. It asks whether independent sources connect you to your category and describe you consistently — corroboration a model can rely on when deciding whether to assert something about you.
It is slow, expensive, and dependent on parties outside your control, which is precisely why it is defensible once built. It is also the pillar most commonly skipped, because the alternatives are faster and feel more productive. A brand that publishes prolifically while no independent source states what category it belongs to has invested heavily in pillar four while pillar three remains unbuilt.
Expression is whether your content answers real questions in passages that survive being lifted out. Answer-first structure, self-contained sections, format matched to the shape of the question, no important information trapped where a machine cannot read it. It is the pillar most closely resembling traditional content work, which is why it receives disproportionate attention.
It is genuinely important — a resolvable, well-corroborated brand whose content cannot be extracted from will be passed over for one whose can. But it is fourth for a reason. Excellent expression on a brand the model cannot resolve or corroborate produces nothing, and that combination describes a large share of the content programmes currently running under the banner of AI visibility.
Access, Identity and Expression are fast, internal and satisfying. Evidence is slow, external and hard to attribute. The framework exists to stop you skipping it, because it is where most brands actually stop.
There is a structural reason organisations get this wrong, and it is worth naming. Three of the four pillars are fast, internal, and attributable. The third is slow, external, and diffuse. Any team choosing work on the basis of what it can execute and demonstrate will systematically underinvest in exactly the pillar that most commonly binds.
This is not a failure of intelligence but of incentive, and the framework’s practical function is to make the omission visible. Once the diagnosis names Evidence as the binding pillar, continuing to fund content work becomes a decision rather than a default — which is usually enough to change it.
Pillar economics
Rows: Access, Identity, Evidence, Expression. Columns: owner, typical timeline, relative cost, attributability of result, frequency as binding constraint. The pattern to make visible: the pillar that binds most often is the slowest, most expensive, and least attributable.
Readers of our earlier work will recognise the Recommendation Ladder™, and the relationship is worth stating so the body of work stays coherent. The Ladder describes the sequence a specific recommendation passes through: retrievability, resolution, association, corroboration, preference. The Framework describes the four capability areas a brand builds and owns.
They map onto each other cleanly. Access supports retrievability; Identity supports resolution; Evidence supports association and corroboration; Expression and Evidence together support preference. Use the Ladder to diagnose a specific failure, and the Framework to organise the programme that fixes it — one is a diagnostic, the other an operating model.
The framework only pays once you know which pillar is failing, and that requires measurement. DUNkē tracks whether you are cited across eight AI engines, per prompt, against competitors — the evidence the diagnosis rests on.
The temptation with any four-part framework is to score each pillar and produce a composite. We advise against it. A composite hides the only thing that matters — which pillar is binding — and produces the reassuring outcome that a brand failing Identity entirely can still score respectably by excelling elsewhere. That is precisely the misreading the ordering exists to prevent.
The output that works is a profile: pass or fail per pillar, with the lowest failure flagged as the constraint. It is less satisfying than a number and considerably more actionable, because it produces a single instruction rather than four areas of partial improvement. If a score is required for reporting, report the binding pillar alongside it and treat the number as context.
Used over time, the framework organises the programme rather than just the diagnosis. Access and Identity become maintained states with periodic verification, since both regress silently through deployments and organisational change. Evidence becomes a standing programme with its own budget and cadence, because it accumulates rather than completing. Expression becomes an editorial standard applied at publication.
That division also clarifies ownership, which is usually the practical obstacle. Access sits with engineering, Identity spans marketing and technical, Evidence sits with PR and brand, Expression with content. Naming the owner per pillar is what turns the framework from a model into a plan, and it is why the four-way split is useful beyond diagnosis.
A framework that cannot be tested is a diagram. Each pillar has a direct test that takes minutes and returns an unambiguous answer. Access: fetch your priority pages as a crawler would and confirm the content is present without executing JavaScript, then check logs for each AI crawler arriving. Identity: ask each engine who you are and read for accuracy, specificity, and hedging.
Evidence: search the category without your brand name and see whether independent sources place you in it, then collect ten third-party descriptions and check whether they agree. Expression: copy three passages from key pages into a blank document and read them cold — if they need surrounding context to make sense, they are not extractable. Four tests, one afternoon, one binding pillar.
An underappreciated property of the lower pillars is that they degrade without anyone deciding they should. Access regresses through deployments that alter rendering, security policies applied without visibility review, and platform migrations that change how content is served. Identity regresses through rebrands that propagate incompletely, acquisitions, and third-party descriptions ageing.
This means the first two pillars are maintained states rather than completed tasks, and they need periodic re-verification tied to the deployment cycle rather than to an annual audit. The characteristic failure is a brand that fixed access eighteen months ago, changed platforms since, and has been invisible for six months without anyone checking the thing they considered done.
The pillars are ordered but not isolated, and two interactions matter practically. Identity amplifies Evidence: corroboration attaches to a resolved entity, so the same volume of independent description produces more usable evidence for a brand the model can identify cleanly than for one it confuses with others. Fixing Identity retroactively increases the value of Evidence you already had.
Expression amplifies everything above it: a brand that is resolvable and well corroborated but whose content cannot be extracted from will be described rather than quoted, which is a weaker form of presence. The practical implication is that Expression work has its highest return on brands that have already cleared the first three — which is the opposite of how it is usually sequenced.
The four pillars with dependencies and owners
Vertical stack, Access at the base through Expression at the top, each band annotated with owner, timeline, cost, and its diagnostic test. Show amplification arrows: Identity into Evidence, Expression across the upper three. This is the framework’s canonical illustration.
The pillars are engine-agnostic in structure and engine-specific in detail, which matters when a brand cares disproportionately about one surface. Access is checked per crawler, since permissions are granular. Identity is tested per engine, because resolution genuinely differs between them. Evidence is largely shared, since the underlying corroboration is the same, though which sources each engine favours varies.
Expression is shared entirely — a well-structured passage serves every retrieval system. So the framework applied to a single engine changes what you test rather than what you build, which is convenient: one programme, verified per surface. It also explains why brands strong on one engine and weak on another usually differ on Access or Identity rather than on the slower pillars.
Honesty about scope prevents overreach. The framework describes eligibility — whether a brand can be cited. It does not describe preference among eligible candidates, which is where positioning, specificity, and genuine differentiation operate, and which the Recommendation Ladder™ addresses at its top rung.
It also does not describe demand. A brand that clears all four pillars for a category nobody asks about has achieved eligibility for an empty query set. Prompt research sits logically before the framework, defining the questions worth being eligible for. The framework answers whether you can appear; it does not answer whether appearing there is worth anything.
An underrated use is settling arguments about where to spend. Marketing wants content, engineering wants technical work, PR wants coverage budget, and each case is reasonable in isolation. The framework converts that into an empirical question: which pillar is binding, and the answer determines the allocation rather than the relative persuasiveness of the advocates.
This works because the tests are cheap and the results are hard to dispute. A brand that fails the identity test in front of the leadership team has settled the question of whether entity work is a priority. That is a considerably better basis for allocation than an argument, and it is why we run the tests live rather than reporting them.
The four-way split also clarifies capability. A team with strong Expression and no Evidence capability will produce excellent content and no citation movement, which describes a large share of in-house setups. Recognising that gap as structural rather than as underperformance changes what you hire for or which agency you brief.
The same applies to briefing external partners. An agency briefed to improve AI visibility will default to what it is good at, which is usually Expression. A brief that names the binding pillar and asks specifically for Evidence work — earned association, description correction, citable assets — produces a different engagement and a different set of candidates.
The framework lands better demonstrated than explained. The sequence that works is to run the identity test live — ask an assistant who the company is, in the room — then show the description spread across third-party sources, then show which competitors are named for the category question and which are not.
Three artefacts, none of which requires accepting the framework in advance, and each of which is uncomfortable in a productive way. The framework then arrives as an explanation of what was just observed rather than as a model to be adopted, which is a considerably easier sell and produces faster agreement on where to spend.
Three recur. Diagnosing correctly and then funding the comfortable pillar anyway, which is the most common and produces a year of activity with no citation movement. Treating a pillar as complete when it was verified once, particularly Access and Identity, which regress. And building a composite score despite the warning, which reliably hides the binding constraint within two reporting cycles.
The fourth, subtler failure is applying the framework without a prompt set — achieving eligibility for questions nobody asks. That produces genuine improvement on every pillar and no commercial result, which is the hardest version to diagnose because everything looks like it worked.
Because each pillar has a different natural owner, the framework doubles as an argument about structure. Brands that make progress typically create a single objective spanning technical, content, and communications, with one person accountable for the citation outcome rather than for their function’s output.
Where that accountability does not exist, each function optimises its own pillar and the binding one — usually Evidence, owned by whoever owns PR — receives attention proportional to how much it is asked about, which is rarely. The framework does not solve that, but naming the owner per pillar makes the gap visible enough to be addressed.
Frameworks should state their own weaknesses. The clearest in this one is that Evidence bundles two things that behave differently: the presence of independent association and the consistency of independent description. They have different remedies, different timelines, and different difficulty, and treating them as one pillar occasionally obscures which is failing.
We keep them together because splitting them makes the framework five pillars and harder to hold, and because in practice the diagnostic distinguishes them within the pillar. But readers applying it should test the two separately, and we would not be surprised to publish a revised version that separates them formally.
Compressed for the reader who will remember one thing: four pillars, ordered by dependency. Access — can the engine reach and read you. Identity — can it tell what you are. Evidence — does anyone independent agree. Expression — can it lift a usable passage. Diagnose the lowest failure, work only there, re-measure.
The three claims that matter are that the order is real, that Evidence usually binds, and that a composite score hides both. Everything else in this article elaborates those. A team that internalises only the ordering will still allocate better than one that has memorised a longer list without it, because the ordering is what converts a list of activities into a decision.
There is an obvious argument against publishing the framework an agency uses to diagnose clients. We publish it because the framework is not the scarce part. Running it properly requires measurement infrastructure, the patience to work a slow pillar, and the organisational authority to act on a finding that lands outside the function that commissioned it.
Those are the constraints, and none of them is solved by knowing the model. A team that reads this and correctly diagnoses its own binding pillar has done something genuinely useful, and will either fix it themselves or brief a partner far better than they otherwise would. Both outcomes are preferable to another year of undiagnosed content production.
Run the four tests in order and stop at the first failure. Fetch a key page as a crawler sees it. Ask three engines who you are. Search your category without your name and see whether anything independent places you in it. Read three of your passages cold.
Whichever fails first is your programme for the next quarter, and everything else on your current roadmap should be explicitly deferred rather than run in parallel. That instruction is uncomfortable precisely because it usually cancels work already scheduled — which is the clearest sign the framework is doing its job rather than validating what you were going to do anyway.
Frameworks are useful because they compress, and compression always loses something. This one loses the fact that the pillars interact, that categories differ in which binds, and that a brand can fail a pillar partially rather than completely. Applied mechanically it will occasionally give the wrong answer.
The protection is to treat the diagnosis as a hypothesis with a testable prediction rather than as a verdict. If remedying the named pillar does not move citation within its expected window, the framework was wrong for that case and the next step is investigation rather than more of the same work. A model that cannot be wrong is not a diagnostic.
Can we work Evidence while fixing Identity?
You can start it, and you should, because it takes quarters. Just do not expect measurable citation movement from it until Identity clears — corroboration has to attach to an entity the model can resolve.
How do we know when a pillar has cleared?
Each has a direct test: crawler access and rendered content for Access; asking engines to identify you for Identity; checking whether independent sources state your category and agree for Evidence; reading passages in isolation for Expression.
Which pillar gives the fastest visible result?
Expression, where the brand already clears the first three. That is why it feels productive and why it misleads — the speed is conditional on the pillars beneath it already being sound.
Does this apply outside AI search?
Access and Expression map onto conventional technical and content SEO, and Identity increasingly affects knowledge panels and entity understanding generally. Evidence has always driven authority. The framework reorganises rather than replaces.
The AI Visibility Framework™ orders four pillars by dependency: Access, Identity, Evidence, Expression. A model must be able to reach you, resolve what you are, find independent agreement about you, and extract a usable passage — in that sequence. The ordering is the substance of the framework, because it means work above a failing pillar produces no measurable result regardless of how well executed it is.
The pattern it exposes is consistent: three pillars are fast, internal, and attributable, while the one that most commonly binds is slow, external, and diffuse — so teams systematically build around it. Diagnose the binding pillar, report a profile rather than a composite so the failure stays visible, assign an owner to each, and treat Evidence as a standing programme rather than a project. That is the difference between a visibility strategy and a list of tactics.
Methodology note: the ordering of the pillars is a diagnostic claim drawn from audit practice, not a measured hierarchy. We do not claim a measured distribution of which pillar binds across a population, and no percentage appears in this article. A study establishing the ordering empirically would hold three pillars constant while varying the fourth across matched brand cohorts and track citation outcomes — a design we describe rather than claim to have run.
“Four pillars, one of them slow, external and hard to attribute — and it is the one that usually decides the outcome. The framework exists mainly to stop you skipping it.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.