What changed in the answer layer itself — coverage, source mix, citation stability. Published method first, because a trends report without one is an opinion with a chart.
This report tracks what changed in the answer layer itself — coverage, citation behaviour, source preferences, and interface shifts — rather than which brands moved. It exists because most teams discover an environmental change months after it affected them, having attributed the resulting decline to their own performance. This edition sets out the methodology in full, because a trends report whose method is undisclosed is an opinion with a chart attached.
The specific problem it solves is misattribution. When citation presence drops, the first question is whether something changed on your side or in the environment, and almost no organisation can answer it. Without an environmental reference, every movement gets attributed to your own work — which produces panic when the cause was external and complacency when it was not.
A trends report is the reference. If coverage expanded into your category this month, that is the explanation for a click-through decline, and no amount of investigating your own site will find it. If nothing environmental moved and your presence fell, the cause is yours and worth pursuing. The report exists to make that distinction available.
It does not rank brands, report who is winning, or benchmark categories. Those are separate reports with different methods and different sampling requirements, and blending them would produce something that does all three badly. This one answers a single question: what changed about how the answer layer behaves.
It also does not forecast. The variables here are set by platform product decisions rather than by trends with momentum, and coverage has already moved in both directions. A report extrapolating from three months of observations would be projecting a line through points that are not on a trajectory.
The fixed observational design behind this report. Published so readers can assess what the findings can and cannot support.
Prompt panel
A fixed set of prompts spanning intent types and categories, held constant across editions so changes reflect the environment rather than a changed sample.
Changes to the panel are disclosed in the edition they occur.
Engine coverage
Each prompt run across the major answer surfaces separately, never blended, because surfaces demonstrably disagree.
Surfaces reported individually.
Repeat sampling
Multiple runs per prompt per edition, because generated answers vary between observations.
Single-run observations are not reported.
Persistence threshold
Changes are reported only where they hold across repeated observations and multiple prompts.
Most month-to-month movement fails this and is excluded.
Stated limits
What the panel cannot support is published alongside what it can.
No claim is made beyond the sample.
Five things. Answer coverage: what proportion of the panel returns a generated answer, tracked per surface, which detects expansion or contraction. Source count: how many sources are cited per answer, which indicates whether the field is widening or narrowing. Source mix: what kinds of sources appear — reference, community, editorial, commercial — which is the most strategically informative measure.
Citation stability: how consistent the cited set is across repeated runs, which indicates how settled the retrieval is. And interface changes: observable shifts in how sources are presented, which affect the value of a citation independently of whether you have one. Together they describe the environment a brand is operating in.
Source mix is the one worth reading closely, because it tells you what kind of thing the engines currently reward in a category. A shift toward community sources indicates that experiential judgement is being favoured over authoritative statement. A shift toward reference sources indicates the opposite. A shift toward commercial sources indicates brands are becoming credible enough to cite directly.
Those are strategically actionable in a way coverage percentages are not. A brand losing citations in a category whose source mix has shifted toward community discussion has a diagnosis available immediately, and it is not a content quality problem. This is the measure we would keep if the report were reduced to one.
Almost no organisation can currently tell. Without an environmental reference, every citation movement gets attributed to internal performance — which produces the wrong response roughly half the time.
Generated answers vary between runs, which means most apparent month-to-month movement is noise. We apply a persistence threshold — a change is reported only where it holds across repeated observations and multiple prompts — and readers should apply the same scepticism to any trends reporting that does not.
The practical guidance is to treat single-month movements as unconfirmed and multi-month directional consistency as signal. A coverage figure that moves and returns is measurement variance. One that moves and stays has told you something. Most published AI search trend data does not make this distinction, which is why it reports dramatic changes that quietly reverse.
The panel is fixed and finite, which means it represents the categories and intent types it contains rather than the web. Coverage figures from it apply to that panel and not to search generally — a point we make repeatedly because prevalence is notoriously sample-dependent and every published figure differs for exactly that reason.
We also cannot report causes. We can observe that source mix shifted; we cannot say why, because engine internals are not public. Where we offer an interpretation it will be labelled as interpretation. And where a section has no reliable signal in a given month, it will say so rather than being filled with something that looks like a finding.
This report tells you what moved for everyone. DUNkē tells you what moved for you — citations across eight AI engines, per prompt, against competitors, on your own question set.
The intended workflow: when your citation tracking shows a change, check this report before investigating internally. If the environment moved in a way that explains it, the finding is contextual and your response is different — possibly no response at all. If nothing environmental moved, the cause is yours and worth the investigation.
That sequence saves considerable wasted diagnostic effort, and it is the entire practical purpose of publishing environmental data. A brand with its own measurement and no environmental reference will misattribute regularly. A brand with both can separate the two, which is the difference between reacting and responding.
Because a trends report without a disclosed method is unfalsifiable, and this publication has argued repeatedly that unfalsifiable claims are the field’s central problem. Publishing the panel design, the sampling protocol, the persistence threshold, and the limits means a reader can assess whether the findings support the claims.
It also means we can be checked. A reader running the same protocol on a similar panel should reach compatible conclusions, and where they do not, one of us has learned something. That is a considerably better position than the industry norm, which is confident percentages with no sampling frame attached.
Blending coverage across surfaces would produce a figure describing none of them. Surfaces expand and contract independently, respond to different product decisions, and serve different query mixes, which means a single coverage number averages behaviours that frequently move in opposite directions.
Reported separately, divergence between surfaces becomes informative. One expanding while another holds indicates a product decision at one platform rather than a general shift in the answer layer, and a brand seeing a change on one surface only can immediately rule out its own performance as the cause.
The number of sources cited per answer is a quiet but consequential measure. A widening set means more citation slots per question, which lowers the bar for inclusion and favours challengers. A narrowing set means the field is concentrating, which favours established, heavily-corroborated sources and makes entry harder.
Neither direction is good or bad in itself; both change strategy. A brand pursuing citation in a narrowing field needs to focus its evidence on fewer questions where it can genuinely lead. In a widening field, broader coverage becomes viable. Watching this measure over several editions tells you which game you are playing.
The fourth measure is how consistent the cited set is across repeated runs of the same prompt. High stability indicates settled retrieval, where positions are earned and durable. Low stability indicates the engine is uncertain, which means the field is contestable and evidence is not yet decisive.
This is the measure most useful for timing. A category with low citation stability is one where a brand can plausibly enter, because nothing has consolidated. High stability means the positions are held by sources with enough corroboration that the engine returns them consistently, and entering requires accumulating comparable evidence rather than being opportunistic.
The fifth measure is not about citation at all but about presentation: whether sources are displayed prominently or subordinated, whether links are provided, and how much of the answer surface the sources occupy. These changes affect what a citation is worth independently of whether you have one.
A platform reducing source prominence lowers the value of every citation on that surface without changing anyone’s standing. A brand tracking only citation share would see no movement and experience a decline in downstream effect it could not explain. Including presentation in an environmental report is what makes that legible.
The five monthly measures
Rows for coverage, source count, source mix, citation stability, and interface changes. Columns: what it measures, what an increase indicates, what a decrease indicates, and what strategic decision it should inform. A single reference page for reading each edition.
Structurally: a short summary of anything that passed the persistence threshold, then the five measures reported per surface with the previous edition alongside, then a section on category-level differences where the panel supports them, then an explicit statement of what showed no reliable change.
That last section is deliberately prominent. Reports that only publish movement create an impression that something significant happens monthly, which is untrue and trains readers to expect drama. Stating plainly that four of five measures were flat is honest reporting and calibrates expectations correctly.
Any fixed panel represents itself rather than the web, and the composition determines what the report can detect. A panel weighted toward informational queries will register coverage changes early, because that is where coverage moves first. One weighted toward commercial queries will register them late.
We publish the composition for that reason, and we would rather readers discount findings that fall outside the panel’s coverage than assume the report speaks for the whole of search. A trends report that does not disclose its panel is asking readers to generalise from a sample they cannot see, which is the failure mode this entire series is constructed to avoid.
Platforms announce changes selectively and describe them favourably, which makes announcements a poor basis for understanding what actually changed. Independent observation catches changes that were not announced, measures the magnitude of ones that were, and occasionally establishes that an announced change had no observable effect on the panel.
That third outcome is more common than expected and is worth reporting when it occurs. A feature announced with fanfare that does not move any measure on a reasonable panel has told you something useful about how much attention to pay to the next announcement.
The measures we would like and do not currently produce: per-category coverage at a granularity that supports individual verticals rather than broad groupings, which requires a much larger panel. Latency measurement — how quickly new content becomes citable — which requires controlled publishing. And citation persistence, tracking how long a source holds a position.
Each is tractable and expensive. Stating what we would measure with more resource is more useful than implying the current measure set is complete, and readers should treat the gaps as gaps rather than as areas where nothing is happening.
How to use an edition in five minutes: check whether any measure moved on the surfaces your audience uses; if so, check whether the direction explains anything you observed in your own data; if not, your movement is internal and worth investigating. That is the entire intended workflow.
What should not happen is treating the report as a source of tactics. It describes an environment; what to do about that environment depends on your position within it, which this cannot see. Reports that conclude with recommendations are usually generalising from a panel to readers whose circumstances they do not know.
A trends series measures change, and change requires a prior state. A first edition reporting current values as though they were findings would be publishing a snapshot dressed as a trend — which is precisely the failure this series is constructed to avoid, and doing it in edition one would undermine everything after it.
So this edition establishes the panel, the protocol, and the thresholds, and subsequent editions report movement against it. That is slower and it means the series becomes useful in its second edition rather than its first, which we consider the correct trade for a report whose entire value rests on being trustworthy about what it can support.
This report tracks the environment — coverage, source count, source mix, citation stability, and interface changes — on a fixed panel, per surface, with a persistence threshold that excludes the variance dominating most published trend data. Its purpose is attribution: distinguishing what changed around you from what changed about you.
Read source mix most closely, because it indicates what kind of source engines currently reward. Discount single-month movements. Do not import coverage figures, which describe the panel rather than your category. And require a published method from any trends data you rely on, including this one.
What each measure indicates
Rows for the five measures. Columns: what an increase indicates strategically, what a decrease indicates, which decision it should inform, and the typical lag before it affects brand-level outcomes. Designed as a one-page companion for reading each edition quickly.
The field currently has confident commentary and almost no disclosed method. Figures circulate without sampling frames, month-to-month variance is reported as trend, and readers have no way to assess whether a finding survives its own methodology because the methodology is not published.
This series is an attempt to do the opposite: publish the design first, state the limits prominently, report what showed no change, and accept shorter and less dramatic editions as the cost. Whether that is commercially sensible is genuinely uncertain. It is the only version we would be willing to defend in three years.
The panel design published here is reproducible, and we would rather be checked than trusted. A reader running a comparable protocol on their own panel who reaches different conclusions has found something worth knowing — either about their categories or about ours.
That is the practical reason for publishing method rather than protecting it. A field where several parties measure openly and compare notes reaches reliable knowledge faster than one where each holds a proprietary dataset and asserts conclusions from it. We would like this to become the norm, and the only way to argue for it is to do it.
First of four. This covers the environment; the benchmark covers distributions, the index covers standings, and the movement report covers deltas. Each answers a different question and each requires a different sampling design, which is why they are separate publications rather than sections of one.
The intended sequence when something changes in your own data is to read this first, because an environmental explanation makes the other three unnecessary. Ruling out the environment is the cheapest diagnostic step available and almost nobody currently has the reference to do it.
Before investigating a change in your own citation data, check whether the environment moved. Half the time it did, and the investigation you were about to run would have found nothing while consuming a week.
That is the entire practical function of an environmental reference, and it is why we publish one openly rather than holding it as a client benefit. A field where everyone can rule out environmental causes cheaply spends its diagnostic effort considerably better than one where nobody can.
Stated for transparency, since a research series that is expensive to run is one that can be quietly abandoned. The panel requires repeated sampling across surfaces monthly, which is largely automatable, plus manual assessment of source mix and interface changes, which is not.
That combination makes it sustainable but not trivial, and it means the panel will grow slowly rather than comprehensively. Readers should expect coverage to expand gradually and should treat any month where the panel changed as a break in the series rather than as a continuous trend, which is why panel changes are disclosed prominently.
There is a commercial argument for keeping this proprietary: knowing what changed before competitors do is worth something. We publish it because the alternative is a field where every organisation misattributes its own performance and nobody can check anyone’s claims.
Environmental reference data is closer to infrastructure than to competitive advantage. It makes everyone’s measurement better, including our clients’ and including our competitors’, and a discipline where attribution is possible is one where good work can be distinguished from luck. That seems worth more than the edge we give up.
Monthly is the fastest defensible cadence for environmental measurement and it is already faster than most of what it measures. Coverage decisions are made on platform timelines, source mix shifts gradually, and interface changes are episodic.
We publish monthly because attribution needs a recent reference and a quarterly report would leave brands unable to explain a change for up to twelve weeks. The trade-off is that most editions will report little, and readers should treat a quiet edition as information rather than as an absence of it.
This report measures the environment so you can tell whether a change in your data was caused by you or by something around you. Five measures, fixed panel, per surface, with a persistence threshold that removes the variance most published trend data reports as news.
Read source mix most closely, discount anything that moved once, and never import a coverage figure into your own planning — it describes this panel, not your category. And require a disclosed method from any trends data you rely on, including this one, because a report you cannot check is an assertion with a chart.
Everything here is reproducible and we would encourage it, particularly for a single category where a focused panel beats a broad one. Fix your prompts, run them across the surfaces your buyers use, sample each several times, and record the five measures monthly.
Your panel will be more sensitive than ours for your category because it is built for it, and you can run a lower persistence threshold than a published report should. The design here is a starting structure rather than a specification, and a category-specific version of it is more useful to you than any general report can be.
How large is the panel?
Fixed and finite, spanning multiple categories and intent types, sized to be sampled repeatedly across surfaces every month. The exact composition and any changes to it are published with each edition, because a panel that changes silently invalidates the trend.
Why not report a single AI visibility index?
Because surfaces disagree and categories differ enough that a blended index would describe no actual condition. We report per surface and per category, which is less tidy and more accurate.
How soon after an engine change would this detect it?
Within an edition where the change affects coverage or source mix on panel prompts, and not at all where it affects categories the panel does not cover. That limitation is inherent to any fixed panel and is why we publish its composition.
Can we suggest additions to the panel?
Yes, and category coverage requests are the most useful. Panel changes are disclosed in the edition they take effect, since an undisclosed change is indistinguishable from a trend.
This report tracks the environment rather than brand performance: coverage, source count, source mix, citation stability, and interface changes, measured on a fixed panel across surfaces separately, with a persistence threshold that excludes the month-to-month variance that dominates most published trend data.
Its purpose is attribution — letting you distinguish a change in the environment from a change in your own performance, which almost no organisation can currently do. Read source mix most closely, discount single-month movements, and do not import coverage figures into your own planning, because prevalence is sample-dependent and this panel describes itself rather than the web.
Methodology note: this edition publishes the design rather than a full data run, because a trends series is only meaningful with prior editions to compare against and this is the first. Subsequent editions report measured change against the panel established here. We would rather publish the method now and the findings when they mean something than publish a first-month snapshot presented as a trend.
“A trends report with no disclosed method is an opinion with a chart attached. If you cannot see the panel and the sampling protocol, you cannot assess whether the finding survives them.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.