Not a blog, but a research library about how AI sees brands
Ansmeter articles, terms, and metrics — from foundational concepts to diagnostic research. Navigate by task, reading path, or reference.
25 materials · 7 reading paths
All materials
Material catalog — 25
Find by task
Practical task
Reading path
7 reading paths
Reference
The reference, metrics, and schemes are not hidden inside articles
| Material | Type | Level | Min |
|---|---|---|---|
| Foreword and guide to the updated Ansmeter corpusHow the Ansmeter research library is organized: article structure, material types, difficulty levels, reading paths, and navigation. | Guide | Introductory | 4 |
| Mini-research card for the Ansmeter libraryAn observation card template for recording data from each Ansmeter test run — so that individual responses build into a research history. | Observation template | Introductory | 4 |
| Why a strong brand can still be invisible to AI systemsExplains the central paradox: a brand can be well known to people and yet poorly distinguishable for AI at the moment of real choice. | Foundational text | Introductory | 7 |
| What AI really “knows” about a company: the brand’s internal representationExamines how a language model holds a brand internally: not as a card with a description, but as a probabilistic network of categories, attributes, and associations. | Foundational text | Intermediate | 7 |
| Which sources AI uses to form an opinion about a brand — and why the site is not the only heroThe layers from which AI assembles its opinion of a brand: the brand's own site, search context, independent reviews, user platforms — and why the site is no longer the sole arbiter. | Foundational text | Intermediate | 7 |
| From search engine to AI intermediary: how the customer path is changingHow the AI intermediary changes the customer journey: choice and comparison increasingly happen before the click, and the first synthesized answer becomes the frame for decision. | Foundational text | Introductory | 8 |
| What the market offers for AI visibility growth — and where the hidden costs liveA map of approaches the market uses to increase AI visibility: what genuinely helps and what merely creates an illusion of control. | Foundational text | Intermediate | 7 |
| The economics of invisibility: how a company loses demand before the first clickHow to translate the problem of AI invisibility from an abstract conversation about traffic into the language of early economic losses and manageable metrics. | Foundational text | Introductory | 7 |
| Mention, citation, and influence: three levels of brand presence in AI answersThree levels of brand presence in AI answers — mention, citation, and influence — and why a single metric is not enough for diagnostics. | Research article | Intermediate | 8 |
| The “answer bubble”: why the same brand looks different in ChatGPT, Google, Copilot, and other systemsWhy there is no single AI visibility: the same brand can look noticeably different across ChatGPT, Google AI Overviews, Copilot, and Perplexity. | Research article | Intermediate | 7 |
| Update lag: how quickly AI systems change their view of a company after news, a product launch, or a price changeWhy there is a time gap between a fact changing about a brand and its stable appearance in machine answers — and how to observe this lag in practice. | Research article | Advanced | 7 |
| Access economics: crawling, indexing, training, and the brand’s right to manage its presenceThe modes that make up AI access to brand content — crawling, indexing, training, licensing — and why this is already an economic question. | Research article | Advanced | 7 |
| Machine-readable commercial infrastructure: markup, product feeds, and catalogs as a language AI can understandThe data and markup layer that makes a brand and its products understandable to machines: catalogs, product feeds, structured descriptions, and their synchronization. | Research article | Advanced | 7 |
| External authority versus the brand’s own site: which sources really create the right to be recommendedWhich external signals and independent sources help a brand earn the right to be recommended in AI answers — and why the brand's own site without them is not enough. | Research article | Intermediate | 7 |
| Category drift: how a brand loses not only to a competitor, but to someone else’s frame of choiceHow a brand can lose not to a competitor but to a different choice frame: AI shifts the user's task into another category and assembles a different set of alternatives. | Research article | Intermediate | 7 |
| SEO and AI visibility: what carries over, what does not, and where familiar optimization can backfireWhat transfers from classic SEO to the AI answer environment, what stops working, and what new requirements emerge. | Foundational text | Introductory | 7 |
| Practical action map: how to strengthen a brand’s machine distinctnessSix sequential steps for improving AI visibility: from identity verification through language reassembly and trust contour to monitoring. | Guide | Intermediate | 8 |
| Observation from a run: how site language made a brand invisible in its own categoryAn observation from a real Ansmeter test run: a brand with strong SEO turned out to be invisible to AI because of a gap between the site language and the query language. | Field note | Intermediate | 4 |
| Visibility through the lens of language and geographyWhy the same brand looks different in AI answers across different languages and countries — and what practical consequences follow. | Research article | Intermediate | 7 |
| Multimodal distinctness: when a brand is searched not with wordsHow visual search, voice queries, and multimodal interfaces change brand visibility requirements — and what transfers from text optimization to the world of images and voice. | Research article | Advanced | 7 |
| ChatGPT Instant Checkout: purchasing without leaving the conversationOpenAI launched purchases directly inside ChatGPT — Instant Checkout. An analysis of what changed and how it affects brand visibility. | Update | Introductory | 2 |
| When the buyer is not a person but their agentHow brand visibility changes when an autonomous AI agent — one that searches, compares, and decides on its own — stands between the company and the buyer. | Research article | Intermediate | 7 |
| Wikipedia, Wikidata, and Knowledge Graph: the invisible foundation of AI visibilityWhy brand presence in Wikipedia, Wikidata, and Knowledge Graph has become a practical lever for AI visibility — and how to work with it. | Foundational text | Intermediate | 5 |
| Visibility Language Field: why the same brand lives in different competitive worldsWhen we ran the same brand across five languages, we expected noise — small score fluctuations. Instead, we found that when the language changes, what changes is not the brand's score but the entire market around it. | Field note | Intermediate | 7 |
| The competitive set: when measurement depends on the neighborsIn AI visibility, some metrics depend not only on the model's behavior but also on who stands next to the brand in the comparison set. The article explains why this set needs to be fixed between repeated runs, how the arithmetic of share-based metrics works, and when an update of the set is justified — and when it zeroes out any possibility of comparing against past reports. | Foundational text | Intermediate | 6 |
Material catalog — 25
Each path has its own logic. Founders need the economics of invisibility, marketers need sources and citation, technical leaders need data, indexing, and machine-readable infrastructure.
1
Where to begin
A route for the first entry into the topic
2
Route for founders
A frame for risk, demand, and brand position
3
Route for marketers
Sources, citation, and practical diagnosis
4
Route for technical leads
Metrics, infrastructure, and control over data availability
5
Route for researchers
The corpus as an observation system
6
Agency and consultant path
Methodological foundation for selling AI visibility services
7
Full course
All 25 materials from fundamentals to advanced diagnostics
Concepts
Terms used in Ansmeter articles and reports. Consistent vocabulary helps read the research and compare results.
- Machine distinctness
- An answer system's ability to consistently resolve a brand as the correct entity in the relevant scenario. Stronger than simple name recognition.
- Functional visibility
- A brand's participation in the actual moment of choice: in a shortlist, a recommendation, or a comparison. A brand may be well known yet functionally invisible.
- Source contour
- The set of owned, external, user-generated, and structured sources from which the system forms an opinion about the brand. The key object of audit.
- Mention
- The fact that the brand name or entity appears in the answer. The weakest level of presence.
- Citation
- A situation in which the brand or a source associated with it becomes evidentiary support for the answer. Stronger than a simple mention.
- Influence
- A brand's ability to define the frame of the answer: the criteria, the category, the list of alternatives, and the language of comparison. The most valuable and the rarest level.
- Update lag
- The delay between a change in a fact about the brand and its stable appearance in a machine-generated answer. It is made up of several stages.
- Machine readability
- The degree to which the properties of a brand and product can be reliably extracted from structured and synchronized data. Especially important in commercial scenarios.
- Category drift
- The displacement of a brand into someone else's choice frame, or the substitution of its market with another task category. It often happens before direct brand comparison.
- Loss before the click
- The brand participation that is lost in shaping choice before the user even clicks through to the site. A new economic unit of analysis.
- Answer environment
- An interaction mode in which the answer system does not return a list of documents, but immediately produces a synthesized answer to the user's question. It is in the answer environment that a brand competes not for rank in the results, but for a place inside the finished answer.
- GEO (Generative Engine Optimization)
- A set of practices aimed at making a brand appear more often and more accurately in the answers of generative answer systems. It differs from classic SEO in that it optimizes not for position in a list of links, but for participation in a synthesized answer. A market term already used by competitors and clients. Useful to know, but important to remember that behind the acronym stands the same task — machine distinctness.
- LLMO (Large Language Model Optimization)
- A synonym for GEO that emphasizes optimization specifically for language models, rather than for search systems with an AI layer. It appears in market materials less often than GEO, but it describes the same task.
- RAG (Retrieval-Augmented Generation)
- An architectural approach in which a language model retrieves relevant documents from an external source (web search, knowledge base, catalog) before generating an answer and uses them as context. RAG is precisely what explains why web sources and structured data affect the model's answer in real time, not only through training.
- Zero-click (zero-click)
- A situation in which the user gets a sufficient answer directly in the interface of a search or answer system and does not visit any external site. In zero-click, the brand either participates in forming the answer or loses contact with the user entirely.
- Knowledge graph (Knowledge Graph)
- A structured database of entities and the relationships between them, used by search and answer systems to identify, classify, and describe real-world objects. A brand's presence in a knowledge graph (for example, Google Knowledge Graph or Wikidata) increases the likelihood of correct entity identification in an answer.
- Entity disambiguation (entity disambiguation)
- A system's ability to distinguish one entity from another when names, categories, or contexts coincide or are similar. If a brand name is ambiguous or overlaps with a generic concept, the model may systematically confuse it with another entity.
- AI Overview
- A block of synthesized answer content that Google displays above the standard search results. It is generated by an AI model based on web sources. For many categories, AI Overview has already become the first point of contact between the user and the information. The brand either enters this block or ends up below it.
- Hallucination (hallucination)
- A situation in which a language model generates a claim that is unsupported by sources or factually wrong, yet sounds confident. A hallucination can either help a brand (the model "invents" a non-existent advantage for it) or harm it (it attributes someone else's weakness to it). Both are dangerous.
- Answer bubble
- An effect in which the same brand looks materially different across the answers of different answer systems because of differences in training data, retrieval architecture, and synthesis logic. There is no single, unified AI visibility. Diagnosis must be conducted across several platforms.
- Agentic choice
- A scenario in which an AI agent independently searches, compares, and makes a decision on behalf of the user — without showing an intermediate list of options. In an agentic scenario, the brand competes not for human attention but for machine preference. This is the next frontier after the answer environment.
- Confidence grade (Confidence Grade)
- An internal assessment of research run quality. It is calculated from the width of the confidence interval, the percentage of broken prompts, and the dispersion across scenario families. Grades: A (high) — narrow interval, stable result. B (good) — reliable result with moderate variation. C (moderate) — noticeable sensitivity to corpus composition. D (exploratory) — unstable result; a repeat run is recommended. The grade shows how stable the measurement is. At C or below, a repeat run is recommended before decisions are made.
- Native mode
- A mode in which the model answers only from its training data, without internet access. It shows how firmly the brand is embedded in the model's base knowledge. If the native score is high, the brand is already embedded in the model. If it is low, the model does not know the brand without external prompts.
- Web mode
- A mode in which the model additionally searches for information on the internet before answering. It shows the extent to which external sources strengthen or weaken the brand's position. The difference between web and native shows whether the site and the external environment help the brand or hold it back.
- Competitive set (competitive set)
- The fixed circle of brands against which the brand under study is measured. The term comes from classic marketing literature (Aaker, Keller, Kapferer) and refers to the neighboring brands within whose context a brand can meaningfully acquire share-based metrics. In AI visibility, the competitive set is usually assembled automatically through prompts such as "who else should be considered in this category," and therefore shifts slightly from one research run to the next. To ensure that repeated audits can be compared fairly, the set is fixed at the first measurement and reused in subsequent rounds until the client explicitly requests an update. visibility metrics split into those that are robust to the composition of the set (appearing in the answer, position, recommendation) and those that are sensitive to it (shares of mentions, scenarios, citations). Without explicitly fixing the set, comparing shares between two research runs of the same brand is methodologically incorrect: the denominator has been recalculated, and the arithmetic produces a new number even where the model's behavior has not changed.
- Visibility Language Field (VLF)
- A phenomenon in which an AI model assembles a different competitive set for a brand depending on the language of the prompts. When the prompt language changes, some competitors disappear from responses, others appear, and rankings shift radically. For a dominant category brand the headline score fluctuates moderately, but for less prominent competitors the spread can reach dozens of points — up to complete invisibility in certain languages. VLF arises from asymmetry of training corpora, differences in web sources, and the formation of separate associative graphs for each language. The model may know a brand equally well in every language yet recommend it differently: placing it first in one language while yielding the top spot to a competitor that did not exist in the first language at all. Brand knowledge does not guarantee a recommendation. A brand may be equally known to the model in five languages yet make the shortlist in only two. Competitors also differ — a strategy that works against one set of rivals may be useless in another language. A separate study is recommended for each target-market language.
- Query type
- A group of scenarios with a single semantic purpose — prompts that all test the same type of user question to an AI system. Ansmeter uses six query types in the core formula and three in the diagnostic layer. Breaking the report down by query type shows in which kinds of AI conversations the brand is more visible and in which it is less, and where the largest gap sits between being mentioned and making the shortlist.
- Solution discovery (organic_discovery)
- A query type in which the user is still looking for options and does not name any brand. The model decides on its own which names to surface first. Does the model name the brand when the user is just starting to search. The most neutral and most demanding type — the brand has nothing to offer here beyond its stable machine-distinctness.
- Shortlist (organic_shortlist)
- A query type in which the user is ready to compare and asks the model to build a short list of options. Does the brand make the shortlist when the user is ready to compare. A check of whether it holds its place among the first names rather than dissolving into the category-wide pool.
- Ranked listings (organic_ranking)
- A query type in which the user explicitly asks the model to order category players (“best of”, “top 5”, “who is stronger”). How high the model places the brand in an explicit category ranking. Shows not only visibility but also the level at which the model is willing to defend it.
- Comparison of options (comparison)
- A query type in which the user compares several solutions without naming specific brands. Does the brand hold up in comparative questions without name prompting. Tests whether the brand surfaces on its own in the comparison dialogue or whether the model ignores it in favour of competitors.
- Trust (trust)
- A query type in which the user asks about reliability, safety, and the soundness of a choice. Does the model associate the brand with reliability and sound choice. Directly measures whether the brand has reputational footing in machine answers.
- Expertise (authority)
- A query type in which the user looks for authority signals: experience, certifications, a long market history, deep subject-matter knowledge. Does the model see authority signals in the brand's domain. Shows whether the model has encoded the brand as a “domain-relevant” player worth listening to.
- Web search effect (web lift)
- The difference between the brand's score in Web mode and Native mode. A positive effect means external web sources reinforce the brand's position; a negative one means they dilute it; around zero means the brand is equally visible through both modes. Shows whether SEO and the external trust perimeter translate into machine visibility, or whether the model relies only on what it learned during training.
- AI Visibility Score
- The brand's final score in the Ansmeter report on a 0–100 scale. Calculated only from the neutral scenarios of the core formula; the diagnostic layer does not enter it. Reported with a confidence interval and a Confidence Grade. This is the canonical term — the same number appearing in different report sections is called "AI Visibility Score". Variants like "score", "rating", or "headline score" are colloquial shortcuts for the same concept and are used only in narrative prose when context is clear.
- Core-formula metrics
- Ansmeter calculates the AI Visibility Score from several metrics with fixed weights. Five of them form the diagnostic layer of "how the model describes the brand when named directly": - **Recommendation Rate** — the share of answers in which the model explicitly recommends the brand. - **Recommendation Strength** — how confidently the model phrases that recommendation. - **Centrality** — whether the brand is the main topic of the answer or mentioned in passing. - **Positive Tone** — the share of answers with explicitly positive sentiment when the brand is named. - **Argument Quality** — how well the model backs its recommendation with arguments and facts. The English names are kept as the canonical terms: they are used in academic literature on measuring AI visibility and in model API responses; translating them would distort the meaning and break the link with external sources. These five metrics describe not how often the brand appears but how the model describes it once it has. They affect only the diagnostic layer of the report — they are not part of the headline score.
- Aspirational tier
- A sub-set of the competitive cohort for the brands that dominate the category, regardless of whether the studied brand competes with them directly. Coexists with the peer tier: peers are the brand's actual market neighbours; aspirational brands are the category giants that set the ceiling. For Webflow in the CMS role the aspirational tier is WordPress and Adobe Experience Manager; the peer tier is Wix Studio, Framer, Squarespace. without the aspirational tier the report is structurally bimodal — you are either a recognised peer or you are not, and the metrics collapse to the "anchored / not anchored" binary. Adding the aspirational tier surfaces the honest gap to the category leader, which is not an indictment but a coordinate: the map of how far the brand is from category-dominant scale and the direction of growth.
- Sweet Spot kernel
- The geometric centre of the brand's consistent strength — the set of prompt families where the brand holds a strong position (position ≤ 3 and mention rate above the threshold) simultaneously across all 24 measurements of the 3 roles × 4 models × 2 lenses grid. A family strong only under the owner-lens but weak under the customer-lens does NOT enter the kernel — that is a positioning gap, surfaced in Section 5. The kernel is a mathematically clean point; the four supporting layers around it (satellites, emergent attributes, presence profile, sentiment + evidence) turn the point into a portrait. the kernel answers "where is the brand consistently strong regardless of lens and model". A narrow kernel means the strength is concentrated in one family; a wide one means the brand is perceived multi-dimensionally. Both shapes are signal, not verdict.
- Lens delta
- The gap in the AI Visibility Score between the owner-lens (marketer / industry-framed prompts) and the customer-lens (buyer-framed prompts). Both lenses run against the exact same roles × models × cohort grid — only the prompt framing differs. That is why the lens delta is structurally cleaner than cross-model or cross-role dispersion: it isolates positioning asymmetry without confounds from cohort composition or model bias. Three patterns: convergent (delta below threshold — self-positioning matches buyer mental models), owner-stronger (the brand reaches observers but not buyers articulating their jobs — messaging gap), customer-stronger (the brand is used or purchased but underrepresented in the marketer-framed narrative — positioning underclaim). a large lens delta is direct evidence of messaging asymmetry, not a hypothesis. The client validates the cause (real positioning weakness vs. audit artifact); the asymmetry itself is measured.
- Convergence level
- A qualitative read of how much different LLMs agree on the brand's ranking inside the competitive cohort. Derived from Kendall's coefficient of concordance (Kendall's W): high (W > 0.7) — models see the brand consistently, the portrait is consensus; moderate (0.4 ≤ W ≤ 0.7) — partial agreement; low (W < 0.4) — substantial disagreement, the brand has a contested identity. Convergence level answers not the accuracy of a single number but the robustness of the portrait: when agreement is high, Sweet Spot is consensus; when it is low, Sweet Spot is the midpoint between divergent views and the divergence itself becomes the story (Section 4). convergence level is a first-class signal, not a trust badge pinned next to the headline number. Low convergence is not "the report is inaccurate" but "the brand is unevenly anchored across LLM ecosystems" — and that is the diagnostic.
Report metrics
Quantitative metrics that appear in the Ansmeter report. Main score weights determine the AI Visibility Score. Diagnostic weights apply only to the explanatory layer.
Main score weights
| Metric | Weight | Description |
|---|---|---|
| Mention Rate | 28% | Share of neutral scenarios where the brand appears in the model's answer at all. A low Mention Rate means the model doesn't recall the brand without a direct hint. This is the most basic indicator: the brand either exists for the model or it doesn't. |
| Top-3 Rate | 14% | How often the brand lands in the top three of the answer rather than merely appearing near the bottom. High Top-3 with low Top-1 means the brand is visible but doesn't dominate. The difference between 'being on the list' and 'leading the list' is the difference between visibility and influence. |
| Top-1 Rate | 10% | How often the model names the brand first — making it the top pick. A consistent Top-1 means the brand has become the model's default recommendation in the category. This is the strongest dominance signal. |
| Avg Position | 15% | Weighted average position of the brand across all answers where it appears. The closer to 1, the higher the brand sits on average. Useful for tracking progress: a 0.5 improvement between runs is a real shift. |
| Prompt Coverage | 18% | In what share of all corpus scenarios the brand appears at least once. Unlike Mention Rate, this counts unique scenarios rather than individual answers. Low coverage means the brand is visible only in one type of question. |
| Response Share | 10% | What share of all mentions in answers belongs to this brand. If there are 10 players in the category and Response Share is 10%, the brand gets exactly its fair share. Above 15% means the brand pulls disproportionate attention. |
| Text Share | 5% | How much text the model devotes to the brand in its answer. A brand can be mentioned in one word or given a full paragraph — Text Share measures that difference. High Text Share with low Mention Rate means the brand appears rarely, but when it does, the model talks about it at length. |
Diagnostic weights
| Metric | Weight | Description |
|---|---|---|
| Recommendation Rate | 30% | Share of answers where the model explicitly recommends the brand, not just mentions it. 'I'd recommend X' and 'X exists' are different things. A high Recommendation Rate means the model considers the brand worth recommending, not just known. |
| Recommendation Strength | 25% | How convincingly the model phrases its recommendation. 'You could look at X' and 'X is the best choice for this task' carry different weight. This metric separates a polite mention from a confident recommendation. |
| Centrality | 20% | Whether the brand is the main topic of the answer or just one item on a list. High centrality means the model builds its answer around the brand. Low means the brand is mentioned but the answer isn't about it. |
| Positive Tone | 15% | Share of answers with explicitly positive tone toward the brand. A model can recommend neutrally or enthusiastically — Positive Tone captures that difference. Consistently negative tone signals a problem in how the model perceives the brand. |
| Argument Quality | 10% | Whether the model supports its recommendation with concrete arguments or sticks to generalities. Quality argumentation means the model can explain why this particular brand. This is the most mature level of visibility. |
Research scenarios
Each scenario tests a different type of visibility
- First appearance
- The user asks about a category for the first time — no names, no hints. The model must recall the brand on its own, relying solely on internal knowledge. This is the baseline test: does the brand exist for the machine as part of the category.
- Shortlist
- The user asks the model to suggest several options for comparison. The test shows whether the brand makes the consideration set — or the model compiles a short list without it.
- Comparison
- The model compares several solutions against user criteria. The test checks how convincingly the brand holds its position in a direct side-by-side — and what arguments the model finds in its favor.
- Ranking
- The model builds an explicit ranking by given criteria. The test shows where the brand lands in the hierarchy and how consistently it holds the upper part of the list.
- Trust
- The user asks whether a solution can be trusted — is it reliable, safe, proven. The test measures whether the model associates the brand with dependability and is willing to recommend it as a safe choice.
- Expertise
- The model answers a question that requires deep product or industry knowledge. The test checks whether the model sees expertise signals in the brand — specific details, specialization, unique competencies.
- Task search
- The user describes a specific task or problem without naming products. The model selects a fitting solution on its own. The test shows whether the model recalls the brand as an answer to a practical need.
- Topic navigation
- The user explores a topic in several steps — from general questions to specific ones. The test checks whether the brand surfaces as the conversation deepens and in what context the model mentions it.
- Agent choice
- An AI agent (automated system) selects a solution without human involvement — for example, for integration or automation. The test simulates a scenario where the brand must be chosen by a machine for a machine.