Deltas over a period, filtered hard. Most of what gets published as movement is answer variance with a narrative attached.
Levels tell you where things stand; movement tells you what is happening. This report tracks deltas — who gained citation share, who lost it, and what changed to cause it — because a brand moving from twelfth to fifth is a more informative event than a brand sitting at third. The discipline it requires is distinguishing genuine movement from the answer variance that makes almost every short-term change look like a story.
A standings table shows an accumulated position, which is the product of years of evidence and says little about what is happening now. A delta isolates a period, which makes it possible to ask what changed in that window — a brand’s work, a competitor’s, or the environment’s.
That specificity is what makes movement actionable. A brand that gained share over a quarter did something, or something happened to it, and the window narrows the candidate explanations enormously. Position alone offers no such handle, which is why a table of leaders is interesting and a table of movers is instructive.
The central methodological difficulty is that generated answers vary between runs, which means share figures move without anything real having changed. A report that treated every fluctuation as movement would produce a dramatic monthly narrative composed almost entirely of noise, which describes a good deal of published AI-search commentary.
Two rules exclude it. A persistence rule: a change is reported only where it holds across repeated observation rounds rather than appearing once. And a minimum-magnitude rule: movements below a stated threshold are treated as indistinguishable from variance regardless of persistence. Together they remove most apparent movement, which is the correct outcome and makes for a shorter report.
Four tests a change must pass before it is reported as movement. Most apparent monthly change fails at least one.
Persistence
Does the change hold across repeated observation rounds rather than appearing in one?
Excludes: single-round fluctuation.
Magnitude
Does it exceed the stated threshold for the category’s baseline variance?
Excludes: movement within normal noise.
Breadth
Does it appear across multiple prompts rather than on one question?
Excludes: single-question artefacts.
Surface check
Is it isolated to one surface or present across several?
Distinguishes: engine change from brand change.
The fourth test does the most diagnostic work. A brand losing share on one surface while holding on others has almost certainly encountered something surface-specific — a retrieval change, a source-mix shift, an access issue with one crawler. A brand losing across all surfaces simultaneously has more likely changed something itself, or been overtaken.
That distinction is unavailable from blended data and it changes the response entirely. It is also why this report cannot be produced from a single-engine measurement, and why any winners-and-losers commentary based on one surface should be read as describing that surface rather than the answer layer.
Gains are pleasant to report and decline is where the transferable content sits. A brand losing citation share has almost always done something identifiable — a migration that broke rendering, a rebrand propagated halfway, a content restructure that removed the passages being cited, a competitor’s evidence programme landing.
Those are recognisable, repeatable failure modes, and reporting them is more useful to a reader than another account of a brand doing well. It is also less commercially comfortable, which is part of why most industry reporting skews toward winners. We report both and expect the losses to be the section people actually use.
Answer variance produces share changes with nothing behind them. A persistence rule, a magnitude threshold, and a breadth test remove most of what other reports present as this month’s story.
We can observe that a brand moved and we cannot observe why, because engine internals are not public and brand-side changes are not disclosed. Where a cause is offered it will be labelled as a hypothesis, supported by whatever observable evidence exists — a visible site migration, a rebrand, a competitor’s publication — and stated as inference.
This is the discipline most commentary in this space abandons. Attributing a citation change to a specific tactic reads as expertise and is usually unfounded, since multiple things change simultaneously and none is isolable. A report that says clearly that it does not know why is more useful than one that guesses confidently.
For readers calibrating their own tracking, genuine movement has a recognisable shape. It appears gradually rather than in a single step, since evidence accrues rather than switching on. It shows on multiple related prompts rather than one. It usually appears in description quality before it appears in share. And it persists across observation rounds rather than reverting.
Movement that appears suddenly, on one prompt, on one surface, and reverses next round is variance. Recognising the difference is the single most useful skill for anyone reading their own citation data, and it is why this report publishes its filter rather than only its conclusions.
Movement is only detectable against a stable baseline sampled repeatedly. DUNkē tracks citation share across eight AI engines over time — per prompt, against competitors — which is what makes a delta trustworthy.
The four reports in this series answer different questions and are designed to be read together. The trends report covers the environment. The benchmark covers distributions — is your number unusual. The index covers standing — who holds the category. This covers movement — what changed and for whom.
The intended sequence when investigating a change in your own data: check trends for an environmental explanation, check this for whether competitors moved in the same window, check the index for whether the category structure shifted, and check the benchmark only when asking whether your level is reasonable. Four questions, four instruments, deliberately not combined into one number.
Stating the failure modes in advance. It will miss slow movement that never crosses the magnitude threshold in a single period, which means gradual erosion can go unreported for several editions. It will occasionally report movement that later proves to be a longer variance cycle. And it will attribute nothing confidently, which some readers will find unsatisfying.
The first is a genuine trade-off: a lower threshold would catch gradual change and would also fill the report with noise. We have chosen the conservative error, and readers tracking their own data should apply a lower threshold to themselves than we apply to a published report, since they can investigate a false positive cheaply and we cannot.
Across the declines we can attribute with reasonable confidence, a small set of causes recurs. Site migrations that changed rendering, so content that was being cited stopped being readable. Rebrands propagated to owned properties and not to third parties, fracturing a coherent record. Content restructures that removed or reworded the specific passages engines were drawing on.
And competitive: a rival’s evidence programme reaching the point where its corroboration exceeded yours on shared questions. The first three are self-inflicted and preventable; the fourth is the market working. Distinguishing them matters because only one requires a strategic response rather than a fix.
One failure mode deserves separate treatment because it is counterintuitive and increasingly common. A brand improves its content — consolidates thin pages, rewrites for clarity, redesigns templates — and loses citations, because the specific passages that were being extracted no longer exist in the same form.
This is a genuine risk of doing good work, and the protection is knowing which passages are being cited before you change them. A brand with citation tracking can check whether a page is currently a source before restructuring it; a brand without one discovers the cost afterwards and usually attributes it to something else.
The profile of a real gain is recognisable and slower than expected. Description quality sharpens first. Presence appears on narrow, specific questions. It becomes consistent on those before appearing at all on broader ones. And it holds across observation rounds rather than appearing and reverting.
A brand appearing suddenly at the top of a broad category question, with no prior movement on narrower ones, is almost always variance. That pattern is worth internalising because it prevents both false celebration and the wrong conclusion about what caused it.
The shape of a genuine gain versus variance
Two overlaid lines over six months. One shows gradual movement — description quality first, narrow questions next, broad questions last — holding at each stage. The other shows a spike and reversion. Annotate the points at which the filter would exclude the second.
A movement report necessarily names sources that lost share, which raises a fairness question we handle by rule: we report observed measurement without commentary on the source’s quality, we offer causes only as hypotheses, and we correct any error a source demonstrates.
We also do not contact sources for comment before publication, because a source explaining its own movement introduces exactly the self-reporting bias the method excludes. That means occasionally publishing a decline whose cause we do not know, which is less satisfying than a narrative and considerably more honest.
There will be editions where nothing passes the filter. That is not a failure of the report; it is the correct output of a period in which little changed, and it is a useful signal in itself — a quiet quarter tells brands that competitive positions are stable and that observed movement in their own data is more likely internal.
Publishing an empty edition rather than lowering the threshold to fill it is the test of whether a research series means what it says about method. We expect this to be the least popular editorial decision in the series and the one that most determines whether it is worth reading.
If a brand appears as a loser, the sequence is: check whether the loss is surface-specific, which points at engine or access issues; check whether it coincides with a change on your side, particularly a migration or restructure; and check whether a competitor gained the equivalent share, which points at competitive rather than technical cause.
If a brand appears as a winner, the useful response is to identify what landed, because the same lever probably has more to give. Movement reports are most valuable to the brands in them, and the brands in them are usually the ones who can least easily explain why.
This report is a market reference and not a substitute for monitoring your own position, which should run continuously and at a lower threshold. You can investigate a false positive cheaply; a published report cannot, which is why our filter is conservative and yours should not be.
The intended interaction: your monitoring alerts you to a change, you check this report to see whether the market moved with you, and the combination distinguishes an internal cause from a competitive or environmental one. That triangulation is the practical value, and it requires both instruments rather than either.
We will miss gradual erosion that never crosses the magnitude threshold in a single period, which is the most consequential blind spot and an accepted cost of a conservative filter. We will occasionally report a movement that proves to be a longer variance cycle. And we will decline to attribute causes readers want attributed.
The first is worth stating most clearly, because a brand declining slowly will not appear here until the accumulated change is large, by which point it has been happening for some time. Readers should not treat absence from this report as evidence that nothing is happening to them — that is what their own continuous monitoring is for.
The most valuable use for most readers is not seeing who won but recognising the failure patterns before they apply to you. A migration planned for next quarter, a rebrand in progress, a content consolidation on the roadmap — each is a documented cause of citation loss, and each is preventable with a check that takes an hour.
The check is knowing which of your pages are currently being cited before you change them, and knowing where your entity description appears before you change it. Both require measurement you should have anyway, and both convert a common loss into a non-event.
It would be easy and popular to attribute movements confidently, and every attribution would be unfounded. Multiple things change simultaneously, engine behaviour is not disclosed, brand-side changes are not announced, and no isolation is possible. Confident attribution in this field is a rhetorical choice rather than an analytical one.
Where we can point at an observable, timed, plausible cause — a visible migration, a public rebrand, a competitor’s major publication — we say so and label it inference. Where we cannot, we report the movement and say we do not know. That will be less satisfying than commentary elsewhere and is the only defensible position available.
This report tracks movement rather than standing, filtered by persistence, magnitude, breadth, and a surface check that separates engine change from brand change. Most apparent movement fails those tests, which makes it shorter than comparable commentary and considerably more reliable.
Losses are reported as carefully as gains, because decline carries the transferable failure modes — broken rendering, half-propagated rebrands, restructures that removed cited passages. Causes are offered as hypotheses because nobody can currently isolate them. And some editions will be empty, which is the correct output of a quiet period rather than a failure of the method.
Diagnosing your own movement
Decision flow starting from "citation share changed". Branches: is it surface-specific (engine or access cause), does it coincide with an internal change (migration, rebrand, restructure), did a competitor gain equivalent share (competitive cause), or none of these (check environmental trends). Each terminal node states the next action.
Of the four in this series, movement reporting has the strongest pull toward bad practice. It is the most read, the most shareable, and the one where a compelling narrative is easiest to construct from noise. Every incentive points toward reporting more movement and attributing it confidently.
The filter exists to resist that, and it will make this the least dramatic movement commentary available. We think that is the correct trade and we are aware it may be the wrong commercial one. Readers should judge the series partly on whether we publish the empty editions we have promised.
Four reports, four questions, deliberately uncombined: what changed in the environment, is my number unusual, who holds the category, and who is moving. Each requires a different sampling design, and collapsing them into a single index would compromise all four.
Used together they let a brand distinguish environmental change from competitive change from internal change — which is the attribution problem at the centre of this discipline and the reason almost every organisation currently reacts to the wrong cause. That is the whole purpose of publishing them.
Fourth of four, and the one that depends most on the others. Movement is only interpretable against an environmental reference, a distribution, and a standings table — otherwise a delta is a number with no context and any explanation for it sounds equally plausible.
Used together the four let a brand answer the attribution question: did the environment change, did I change, or did a competitor change. That question sits underneath almost every decision in this discipline, and almost no organisation can currently answer it.
Before you conclude that a change in your citation data means something, check whether it persists, whether it appears on more than one prompt, whether it is confined to one surface, and whether it exceeds your own baseline variance. Most changes fail at least one of those tests.
Applying that filter to your own data is more valuable than anything in this report, because you can run it continuously and at a lower threshold than a published series can. The filter is the transferable part; the report is the market reference that makes your results interpretable.
Not readership. The measure that matters is whether brands start distinguishing environmental, competitive, and internal causes before responding to a change — which would show up as fewer misdirected diagnostics and fewer teams concluding that AI visibility is unmeasurable after chasing a variance signal for a quarter.
That is not observable to us directly, which is an awkward property for a measurement publication to have. What we can control is publishing the filter, publishing the empty editions, and refusing to attribute causes we cannot support. Whether that changes anyone’s practice is out of our hands.
Most of what circulates as AI search movement analysis is variance with a story attached, and the stories are compelling precisely because they are constructed after the fact to fit whatever moved. That is an easy and profitable format, and it degrades the field’s ability to distinguish signal from noise.
The filter published here is the whole of our objection to it. Apply persistence, magnitude, breadth, and a surface check to any movement claim — ours included — and most of what gets reported as a trend does not survive. That test costs nothing to apply and is the most useful thing in this article.
We have committed to publishing editions where nothing passes the filter, and it is worth being explicit that this will look like a failure to some readers. A movement report with no movements reads as a report with no content.
It is not. It is the correct output of a stable period and it carries a genuine finding: competitive positions did not shift, which means changes in your own data are more likely internal. That is directly useful and it is the kind of usefulness that does not perform well, which is why almost nobody publishes it.
Movement rather than standing, filtered by persistence, magnitude, breadth, and a surface check that separates engine change from brand change. Most apparent movement fails those tests, which makes this shorter and more reliable than comparable commentary.
Losses are reported as carefully as gains, because decline carries the repeatable failure modes: broken rendering after a migration, a rebrand propagated halfway, a restructure that removed the passages being cited. Causes are hypotheses, because nobody in this field can isolate them — and saying so is more useful than guessing confidently.
The four tests here — persistence, magnitude, breadth, surface — apply to any claim about AI search movement, not just to ours. Applied to industry commentary generally, most of what circulates as a trend fails at least two of them, usually persistence and breadth.
That makes the filter more valuable than the report. A reader who takes nothing else from this series but the habit of asking whether a reported change persisted, exceeded variance, appeared broadly, and held across surfaces will discard most of the noise this field produces — which is a considerably better outcome than reading our editions.
Why is the report short some months?
Because little passed the filter. A long movement report is usually a sign that variance is being presented as news, and we would rather publish a short honest edition than a padded one.
Can a brand appear as both a winner and a loser?
Yes, across different surfaces or different question groups, and that combination is among the most informative outcomes — it usually indicates a source-mix or category-association shift rather than a general change.
How long before a brand’s work would show up here?
Structural and entity work typically shows in description quality within weeks and in share over a quarter or more. Evidence work takes longer. A brand appearing as a mover has usually been working for two or three quarters.
Do you contact brands about their movement?
No. Positions and movements are computed from observation, and asking brands to explain their own results would introduce exactly the self-reporting bias the method is constructed to avoid.
This report covers movement rather than standing: who gained and lost citation share over a period, filtered by persistence, magnitude, breadth, and a surface check that separates engine changes from brand changes. Most apparent movement fails those tests, which makes the report shorter than comparable commentary and considerably more reliable.
Causes are offered as hypotheses because nobody in this field can currently isolate them, and losses are reported as carefully as gains because decline carries the transferable failure modes — broken rendering, half-propagated rebrands, restructures that removed the passages being cited. Read it alongside the trends, benchmark, and index reports, each of which answers a different question and none of which should be collapsed into a single number.
Methodology note: this edition publishes the filter and reporting design rather than a data run, since movement requires prior editions to measure against and this is the first. Thresholds for magnitude will be stated per category once baseline variance has been established through repeated observation, because a threshold set before variance is measured would be arbitrary. We would rather publish the rules now and the movements when they are trustworthy.
“Most of what gets published as this month’s AI search movement is answer variance with a narrative attached. The filter that removes it makes for a much shorter and much more useful report.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.