The convention. The reference glossary in this category runs to roughly eighty-four terms with no citations, no dates and no attribution on any entry. That is the category norm, and it is the opening. Every entry here carries four things it does not:
| Field | Why |
|---|---|
| Origin | LBOX, a named originator, industry, or a vendor product name we define but do not adopt. Borrowed frameworks are credited to their author, never to whoever we read them in (Terminology Standard Rule 4) |
| Evidence | Where a term encodes a claim about how engines behave: its evidence class and an as-of date. Model behaviour changes between generations; an undated effect is a liability |
| Often confused with | The near-miss that actually causes errors. The most useful field here, and no competitor glossary has it |
| Status | settled, contested, deprecated, or do not adopt |
Audience. Three readers, not one: clients, our commercial team, and engineers. Sections 1-6 are category vocabulary and publish. Sections 7-10 are internal - identity keys, flag semantics, integration contracts, operating discipline. Every entry in those sections has cost someone real time; the definition is the smaller half, the failure it prevents is the point.
1. Evidence and sourcing
Evidence class · LBOX · settled — The strength label on every claim we publish: DOCUMENTED (stated by the party who would know), REPORTED (asserted by a credible third party), SAMPLED (measured by us, with n and window), ESTIMATE (modelled or inferred). Always paired with an as-of date. Often confused with: source class. Evidence class is how strongly shown; source class is who is speaking.
Source class · LBOX, Adam 2026-08-20 · settled — What kind of speaker a claim comes from: primary_vendor, independent_research, practitioner, popularizer, aggregator_echo. Each is authoritative about a narrow thing. A vendor statement is strong evidence of stated policy and no evidence of observed effect.
Laundering · LBOX · settled, observed 2026-08-20 — When a vendor claim passes through a third party and gains credibility while losing its warning label. A commercial assertion restated by a commentator scores better than the vendor saying it directly, because the conflict flag keys on the speaker rather than the origin. Attach conflict to the root origin.
Root origin · LBOX · settled — The earliest identifiable source of a figure: the study, the press release, the vendor post. Not the article restating it.
Corroboration · LBOX · settled — Count of distinct root origins, never count of mentions. Fifty articles tracing to one press release is corroboration of one. Agreement across different source classes outranks agreement within one, because the incentives oppose. Often confused with: consensus. Consensus counts voices; corroboration counts origins.
Unattributable figure · LBOX · settled — A number with no reachable root origin. Not downgraded, refused. May be recorded as circulating and unattributed; renders nowhere as fact.
Contested · LBOX · settled — A published document that new validated evidence disagrees with. It renders as contested, showing what disputes it. It never silently continues to read as settled.
2. How engines behave
Answer engine · industry · settled — Any system that answers directly by synthesising sources, rather than returning a ranked list of documents to read.
Retrieval substrate · LBOX · settled · REPORTED, as-of 2026-08-22 — The index an engine actually reaches when it searches, frequently not its parent company's. Crawlability on one search engine does not imply reachability by an assistant sitting on a different substrate. Often confused with: the model. The substrate decides what is available to cite; the model decides what gets cited.
Query fan-out · Google (Robby Stein, blog.google, 2025-03-05) · settled · DOCUMENTED — Google's own published term for decomposing one question into multiple concurrent sub-searches before synthesising. Attribute to Google. At least one vendor uses it as a product feature name without attribution; we do not.
Multi-turn session · industry · settled — A real conversation where constraints accumulate across turns and narrow what is reachable later. The unit almost no measurement instrument reproduces. Often confused with: a repeated prompt. Firing one question a hundred times measures a conversation nobody is having.
Single-turn probe · LBOX, descriptive · settled — Asking an engine a cold, contextless question and counting the answer. The standard industry instrument, synthetic by construction, and a limitation the category's leading vendors have published and not resolved.
Retrievability · LBOX · settled — Whether the HTML a crawler actually receives contains the answer. Judged on served bytes, not on what a browser renders after JavaScript.
Empty shell · LBOX · settled · SAMPLED, as-of 2026-08-22 — A retrievability verdict: a client-rendered page that served no content before JavaScript. Found on lbox.com itself, which is why we use it as the worked example rather than a client's.
Comprehension problem · LBOX naming · settled — The engine does not know you have a capability. Solved on-page, by publishing.
Trust problem · LBOX naming · settled — The engine knows you have the capability and still declines to recommend you. More of your own content cannot fix this; only third-party corroboration can. Often confused with: a visibility problem. Identical in a mention count, opposite remedies. Distinguishing them is the point of pairing discovery with validation prompts.
Reachable / Chosen / Lifted · LBOX naming · settled — Three sequential gates from crawl to citation, failing for different reasons: access, credibility, structure. Ours carries drop-off volumes; naming gates without volumes diagnoses nothing.
3. Measurement, and the words that mislead
Visibility score · industry · contested — The proportion of tracked answers mentioning a brand. Real, and not comparable across vendors: a same-domain, same-window test found one tool reporting 1.9x more citations than another because each counts differently. Track the trend, never the absolute, and never mix two tools.
Verified mention · LBOX · settled — A brand mention confirmed because the string was found in the answer text by code, not because a model asserted it. Every metric filters on this. A model self-reporting which brands it named is an unaudited claim about its own output.
Citation share · industry · settled — Proportion of cited sources attributable to a domain. Distinct from visibility: you can be mentioned often and cited never.
Citation source class · LBOX · settled · SAMPLED, as-of 2026-08-22 — Classification of citing domains derived from measured distribution in our own corpus rather than borrowed, which is why video and community are separate classes here. Rule: classes are defined by relationship to the brand, not domain type. A founder's or employee's personal site is owned, wherever hosted. Filing it as earned media inflates the one column the category is weakest in.
Factual accuracy · LBOX · settled — Whether a claim an engine makes is true against external reality: pricing, certifications, integration support. A wrong answer is a defect.
Positioning conformance · LBOX · settled — Whether an engine's description matches the brand's declared messaging. A divergence is a finding, not an error - sometimes the outside world is right and the brand should update. Often confused with: factual accuracy. At least one vendor blends the two and reports the result as an accuracy percentage using brand-supplied ground truth. Blending hides which of a client's problems is a correction and which is an argument they are losing.
Frozen baseline · industry practice · settled — A measurement preserved unchanged so later runs are comparable. Without it, movement is unprovable.
Held-out cohort · standard experimental practice · settled — A matched set deliberately not optimised, so the treated set's delta means something. The rarest discipline in this category: one controlled test found an untouched control page rising more than the edited pages, making the intervention indistinguishable from noise.
Observed, not attributed · LBOX · settled — The mandatory label on any before/after where no control was held. It went up; we are not claiming we caused it.
Latency curve · LBOX · settled · ESTIMATE, as-of 2026-08-22 — The expected order and timing of signals after a change: crawl, then citation appearance, then visibility score, then rank, then sentiment last. Includes a defined too-early-to-read period. Every published case study in this category spans four to five months while its re-measure step reads as something you do next week, which is how clients conclude failure and churn a week before results arrive.
4. The program
Positioning canvas · LBOX; lineage below · settled — The artifact fixing what a brand wants said before anything is measured: category frame, named competitive alternatives, differentiator, proof, who it is for, who it is explicitly not for. Held as inferred/confirmed pairs: the system proposes from page content, the client confirms, confirmed always wins. Lineage: the For/Who/The/That/Unlike sentence form is Geoffrey Moore's (Crossing the Chasm). April Dunford explicitly rejects positioning statements - her 2019 piece is headed that they are harmful and you should never write one. Attributing a Moore-shaped template to Dunford inverts her stated position; do not.
Category frame · April Dunford, competitive frame of reference · settled · SAMPLED, as-of 2026-08-22 — The category label you choose to be evaluated within. Measured swings of 25 to 50 percentage points in visibility from changing it alone, on published five-brand data. Often confused with: a fill-in-the-blank. It is a decision, and the one most often made by default.
Anti-target · LBOX naming · settled — Who you are explicitly not for. Derived from the frame and audience rather than asserted, because a brand that hedges across segments gets described inconsistently, and a synthesis engine reads inconsistency as evasion.
Attribute · industry · settled — A specific thing a brand wants to be known for, at the grain measurement rolls up to. Prompts nest beneath attributes; scores read at attribute level, never prompt level.
The six prompt types · industry · settled — Discovery (must return brand names), Competitor (must return a verdict), Validation (must return yes or no), Accuracy (checkable against source of truth), Sentiment (must return pros and cons), Market perception (must return category criteria, not brands). Rule: a prompt not producing its type's expected output is not failing. It is acting as a different type.
Trigger word · industry · settled — A category noun (software, tool, platform, provider, solution) without which a discovery prompt returns advice instead of brands.
Coverage · industry · contested — How many prompt variants an attribute needs before the reading stops being hostage to phrasing. Universally prescribed, and no published source attaches a number, stopping rule or saturation test.
Spine / depth · LBOX, Adam 2026-08-22 · settled — The spine is the recommended path through the program; depth is everything else: available, valuable, explicitly not an unfinished step. Neither is a permission. Nothing is locked. A client who takes the spine and skips the rest has completed a program, not abandoned one.
Fix class · LBOX · settled — Which party can close a gap: content_fixable (we do it), supporting_content (we draft, you approve), brand_work (only you can, never auto-fabricated). The third is the honest core of a prescription: the part no vendor can sell you out of.
4a. The LBOX proprietary frame
Governed by the Terminology Standard, which is the naming authority. That document decides which word goes in front of whom; this glossary defines what the concept is. Where they touch, the Terminology Standard wins.
All entries in this section are DRAFT. The Terminology Standard is marked draft for Adam's approval, nothing published, with three open decisions. Nothing below goes in front of a buyer until those are settled.
| Market-facing | Internal | What it is |
|---|---|---|
| Evidence Footprint | WES | Where proof of your claims exists online, and how strong each source is |
| AI Visibility | ADS | Whether AI brings you up unprompted when a buyer asks an unbranded question |
| Citation Readiness | CRS | Whether your evidence is complete, consistent and independent enough for AI to repeat safely |
| The Gap | Opportunity Gap | The distance between the visibility your evidence deserves and the visibility you get |
| Reputation drag | new | Negative material that may be suppressing recommendation. Provisional, unproven, tied to CP-02. No causal claim until measured |
The four-step diagnostic · LBOX market framing; underlying ladder is published research · settled as language, contested as mechanism — A buyer locates themselves without being taught anything: AI never finds you, finds you but doesn't mention you, mentions you but doesn't recommend you, recommends you. Internally S1-S4. Critical constraint: per the source authors' own paper, only the first step's causal reading is strong. The middle two labels are working interpretations. Never state them as mechanism in copy.
Prominence · LBOX naming; ladder is published research · settled — Awareness footprint, not size. Category leader, Established challenger, Mid-market, Specialist, Regional (internally L1-L5). A profitable company with real customers can sit at Specialist if the authority lists don't carry it. Say that explicitly; it defuses the offence a founder takes at being called long-tail. Never use tier tokens in client copy.
Terms we deliberately decline
Not avoided because awkward. Declined because using them would make a claim we cannot stand behind.
Share of voice · declined — uninterpretable when computed from retrieval. Say share of retrieved conversation, or omit.
Sentiment score · declined — implies a measurement we do not stand behind. Use named issues with evidence.
AI ranking factors · declined — nobody outside the labs knows these, and claiming them is precisely the thing we criticise competitors for. Say what the published research measures.
Internal tokens in client copy — L1, S3, Jaccard, WES, ADS, CRS. Rule 1: never lead with an acronym. A buyer who must learn your vocabulary before understanding your value has already left.
5. Approaches, named for mechanism never for brand
Full definitions with what each can and cannot fix are in the strategy taxonomy. Headline entries:
Differential agent-serving · LBOX naming · settled — detecting AI agents at the CDN and serving a parallel, token-light, JavaScript-free version of the page. Fixes legibility; fixes nothing about what third parties say; carries cloaking-adjacent policy exposure and the highest vendor lock-in in the taxonomy.
Gap-filled owned content generation · LBOX naming · settled — detecting topics you are absent from and publishing your own content into them. Fixes genuine coverage absence. Cannot fix a trust problem, which is roughly two of three routing outcomes.
Third-party inclusion outreach · LBOX naming · settled — getting added to the ranked lists engines actually cite. The strongest-evidenced remedy in the category, and sold by essentially nobody.
Earned-media distribution · LBOX naming · settled · REPORTED, as-of 2026-08-22 — placing identical content on third-party publisher domains rather than your own. Changes the identity of the citing domain without changing the content. A published measurement recorded an 8% to 34% citation rate, and neither the measuring vendor nor any other sells the distribution.
Self-directed AI instruction page · LBOX naming · do not adopt — a page addressed to AI systems with directives about how to describe you. Untested, unenforceable, legible to any analyst who reads it. Two vendors in the category publish one. Catalogued because it exists and is spreading, not because it works.
Monoculture · LBOX, Adam 2026-08-22 · settled · SAMPLED, as-of 2026-08-22 — Over-reliance on one or two approaches. The category's actual weakness: roughly 82% of shipped remedy capability sits in owned media, against a routing logic published by the leading vendor that sends two of three outcomes off-page.
6. Terms we define but do not adopt
Defined so a reader meeting them elsewhere is not lost. Marked so nobody borrows them by accident.
Promptologie · vendor chapter title · do not adopt — appears as a chapter heading and nowhere in the body of the document that coins it. Never defined, never used. Citing it as a framework cites an empty set.
llms.txt · community proposal · DISPROVEN for citation impact, as-of 2026-08-22 — a well-known file intended to guide AI readers. Independent studies found the overwhelming majority of published files receive no requests at all, and Google has confirmed no support. Harmless to publish; do not sell it as a remedy.
Prompt volume · vendor product term · contested — a modelled estimate of conversational demand. No engine publishes such data and no vendor discloses its model. The least verifiable number in the category and among the most used to justify budget. We decline to produce our own.