Glossary
Website assistant glossary
85 terms used when building, buying or operating an AI assistant on a website. Each one says what it is, and why it changes a decision.
- Private launchOnboarding selected teams now
- Answers from your pages onlyIt refuses when your content does not cover it
85 terms
A
- AbuseDeployment
Deliberate misuse of an assistant: scripted floods, attempts to extract the system prompt, or using it as a free model endpoint.
The cost of abuse is metered per token, so it shows up on an invoice before it shows up in monitoring.
- AI website assistantRetrieval
A chat interface on a website that answers visitor questions from that site's own published content, rather than from a model's general knowledge or a scripted decision tree.
The distinction from a chatbot is grounding. A scripted bot follows branches you wrote; an assistant retrieves passages and phrases an answer from them, which means it can handle phrasing you never anticipated and can also be wrong in ways a script cannot.
RelatedRetrieval augmented generationGroundingWebsite trainingRefusal
- Answer engine optimisationSearch and discovery
Writing and structuring content so that AI answer systems can find, quote and attribute it. In practice it means leading with a direct answer, being specific enough to quote, and sourcing factual claims.
The same structure that makes a page quotable by an external answer engine makes it retrievable by your own assistant. One piece of work, two payoffs.
RelatedStructured dataQuotablellms.txtRetrieval augmented generation
B
A structured data type describing where a page sits in the site hierarchy.
Cheap to add, and it is what lets a search result show a path rather than a bare URL.
RelatedStructured dataSchema.org
C
- CanonicalSearch and discovery
The URL declared as the authoritative version of a page when several URLs serve similar content.
A canonical pointing at a URL that then redirects is reported as a page with redirect and quietly wastes the signal.
RelatedSitemapStructured data
- ChunkingRetrieval
Splitting content into retrievable passages. Where the boundaries fall matters more than how large the pieces are, because a boundary that cuts an answer in half produces two passages, neither of which answers the question.
Chunking on document structure rather than character count is most of the difference between an assistant that answers and one that returns something adjacent to the answer.
- CitationRetrieval
Naming the page an answer came from, shown to the visitor alongside the answer.
It lets the visitor verify, and it is the cheapest debugging tool available: when an answer is wrong you immediately know whether retrieval or phrasing failed.
- ConfidenceModels
A signal, either from the model or derived from retrieval scores, indicating how well supported an answer is.
Confidence is useful for routing and for triggering handoff, and misleading if shown to visitors as a percentage, because it measures support rather than correctness.
- Content gapAnalytics
A topic visitors ask about that your content does not answer.
Gaps found in chat logs are stronger evidence than keyword tools, because someone typed the question on your site with intent to act on the answer.
- Context windowModels
The amount of text a model can consider at once, including the supplied passages, the conversation history and the answer.
Larger is not automatically better, because irrelevant context dilutes the relevant part. A big window is permission to be lazy about retrieval.
- Conversation analyticsAnalytics
Aggregate analysis of chat transcripts: what was asked, what was answered, what was refused, and what happened next.
The useful output is not a dashboard of volume. It is a ranked list of topics your content failed to cover, sorted by how many people needed them, which is a content roadmap drawn from your own visitors.
- Conversation rateAnalytics
The share of visitors who start a conversation.
Low is not automatically bad. On a page that answers its own questions well, a low conversation rate is the site working.
- Conversion rateConversion
The share of visitors who take the action a page exists to produce.
An assistant can raise it by answering objections and lower it by competing with the primary action. Both effects are real and need measuring separately.
- Cosine similarityRetrieval
A measure of how close two vectors point in the same direction, used to rank stored passages against a question.
It measures similarity of meaning, not relevance to the asker's intent, which is why a high scoring passage can still be the wrong answer to the question asked.
- Cost per conversationPerformance
The fully loaded cost of one visitor conversation: retrieval, prompt tokens, answer tokens and any reranking.
It is the only cost number that scales with success. Modelling it per conversation rather than per month is what makes pricing decisions honest.
RelatedTokenModel routingRate limit
- CoverageAnalytics
The share of real visitor questions your indexed content can actually answer.
Coverage is measurable without labelled data: count refusals and low confidence answers by topic and you have a content roadmap.
- CrawlDeployment
Reading the pages of a site to collect their content for indexing.
What the crawler can reach defines what the assistant can answer. Content behind a login, inside an image or rendered only by client side script is invisible to it.
- CreditPerformance
One unit of billed usage. On Creobot's pricing, one credit covers one visitor conversation regardless of how many messages it contains.
Charging per conversation rather than per message matters because per message pricing creates pressure to make the assistant terser, which is the opposite of what a visitor needs.
D
- Data boundaryData and privacy
A plain statement of what data goes where: what is stored, for how long, who processes it and what leaves your control.
Buyers ask this before they ask about accuracy. An honest boundary described in plain language beats a compliance badge with no detail behind it.
RelatedRetentionSubprocessorDPAPII
- DeflectionSupport
A question the assistant answered that would otherwise have become a support ticket.
Deflection is the metric most often inflated. Counting every answered question as a deflection assumes every visitor would otherwise have written in, which is not true.
- DPAData and privacy
A data processing agreement, the contract governing how a processor may handle personal data on a controller's behalf.
It is the document a buyer's legal team asks for. Not having one yet is defensible for an early-stage product, provided you say so plainly.
E
- EmbedDeployment
The snippet of script added to a site that loads the assistant.
One embed on every page is the fast path. It is also how an assistant ends up on checkout and login pages where it does not belong.
- EmbeddingRetrieval
A list of numbers representing the meaning of a piece of text, produced by a model trained so that texts with similar meaning end up with similar numbers.
Embeddings are what make semantic search possible, but they encode similarity, not truth or recency, so an embedding will happily match a confidently wrong passage.
RelatedVector databaseCosine similarityChunkingSemantic search
- Escalation triggerSupport
The written rule that decides when a conversation goes to a person.
Triggers should be auditable rules you can review, not a judgement the model makes fresh each time. Repeated failure, explicit request and high value intent are the reliable three.
F
- Fallback modelModels
A second model used when the primary is unavailable, rate limited or too slow.
Without one, a provider incident becomes a visible outage on your marketing site. With one, it becomes a slightly different tone of answer.
- FAQPageSearch and discovery
A structured data type describing a page's questions and answers.
Only valid when the questions and answers are actually visible on the page. Hidden FAQ schema is a policy violation, not a shortcut.
RelatedStructured dataSchema.org
- File trainingDeployment
Uploading documents so their contents become retrievable alongside crawled pages. Handbooks, specification sheets, price lists.
Files are often the highest value sources on a site because they hold answers that were never published as web pages. A text based PDF extracts cleanly; a scanned one has no text layer and yields nothing.
- Follow up rateAnalytics
How often a visitor asks again immediately after an answer.
A high follow up rate on one topic usually means the answer was technically correct and practically useless.
RelatedAnswer qualityContent gap
G
- Generative engine optimisationSearch and discovery
Structuring content so that generative AI systems can find, use and attribute it when composing an answer. Frequently shortened to GEO.
In practice it overlaps almost entirely with writing quotable, well sourced, directly answered content. The techniques that make a page usable by an external generative system are the same ones that make it retrievable by your own assistant.
RelatedAnswer engine optimisationLLM visibilityQuotableStructured data
- GroundingRetrieval
Constraining an assistant to answer only from supplied passages, and to say it does not know when the passages do not contain the answer.
Grounding is a refusal mechanism before it is an accuracy mechanism. An assistant that cannot refuse cannot be trusted on a page where wrong answers become commitments.
RelatedHallucinationRefusalCitationRetrieval augmented generation
- GuardrailModels
A rule constraining what an assistant will discuss, promise or reveal, enforced in the system prompt, in code, or both.
Guardrails written only in the prompt are suggestions. The ones that matter commercially, such as never quoting a price, belong in code as well.
H
- HallucinationRetrieval
An answer that is fluent, confident and not supported by any source. On a marketing site the dangerous version is not obviously absurd, it is plausible and slightly out of date.
The plausible hallucination is the expensive one, because nobody catches it. An invented price or policy is a commitment your company did not make, in writing, to a buyer.
- Human handoffSupport
Passing a conversation from the assistant to a person, with the transcript and context intact.
Handoff quality is judged on what the person receives. Handing over a bare email address recreates the work the assistant was supposed to save.
- Hybrid searchRetrieval
Running keyword and semantic retrieval together and merging the ranked results.
It covers the two failure modes that each method has alone: semantic search missing an exact code, and keyword search missing a paraphrase.
I
- IndexRetrieval
The processed, searchable form of your content: chunks plus their embeddings plus the metadata needed to attribute and filter them.
An index is a snapshot. The moment a page changes, the index is wrong until it is rebuilt, which is why staleness is an operational problem rather than a one time setup problem.
- IntegrationDeployment
A connection that moves data between the assistant and another system, for example sending a captured lead to a CRM or an escalation to a chat channel.
The word is used loosely. A script embed is not an integration, and a webhook carrying the same name as a product is not the same as native support for it. Ask what actually moves and in which direction.
RelatedWebhookEmbedHuman handoff
- IntentAnalytics
What the visitor is actually trying to accomplish, as distinct from the words they typed.
Phrasing carries intent reliably enough to route on. A question containing how much has different urgency from one containing how does.
K
- Keyword searchSearch and discovery
Search that matches literal words and phrases.
Still better than semantic search for exact identifiers such as SKUs, error codes and plan names, which is why serious systems run both and merge the results.
RelatedSemantic searchHybrid search
L
- Large language modelModels
A model trained to predict text, used here to phrase an answer from passages it was given rather than from what it memorised.
Model choice matters less than most buyers expect. If retrieval returns the wrong passage, the best available model answers confidently and wrongly.
RelatedRetrieval augmented generationTokenContext windowModel routing
- LatencyPerformance
The time between a visitor sending a question and seeing the answer begin.
Retrieval adds time on a page where people are impatient. Budget it explicitly: time to first token matters more than time to complete answer.
- LauncherDeployment
The visible button that opens the assistant.
Launcher placement is a conversion decision, not a styling one. On mobile it competes with your primary call to action for the same thumb.
- Lead captureConversion
Collecting contact details inside a conversation.
Asking too early kills the conversation. Asking after the assistant has demonstrably helped converts far better and annoys far fewer people.
- LLM visibilitySearch and discovery
Whether and how a language model system surfaces your content when answering a question in your domain.
It is not directly measurable the way a search ranking is, because answers vary by phrasing, model and session. What can be measured is whether your pages are crawlable, quotable and specific enough to be cited at all.
RelatedGenerative engine optimisationAnswer engine optimisationllms.txtQuotable
- llms.txtSearch and discovery
A plain text file at the root of a site that tells language model systems what the site contains and where the useful pages are.
It is a proposal, not a standard, and support is uneven. Low cost to publish and no harm if ignored.
M
- Model routingModels
Sending different questions to different models based on difficulty, cost, latency or availability.
Routing is mainly an economics and reliability decision. Simple lookups do not need an expensive model, and no single provider is up all the time.
RelatedFallback modelCost per conversationLatencyLarge language model
O
- One click deployDeployment
Publishing an assistant to a live site without a rebuild or a developer.
It is only one click if training, guardrails and escalation rules were decided beforehand. The click is the last step, not the whole job.
P
- PIIData and privacy
Personally identifiable information: anything that identifies a specific person, including an email address typed into a chat.
Visitors will paste things you did not ask for. Handling that gracefully is a design requirement, not an edge case.
- PrecisionRetrieval
The share of returned passages that are actually relevant.
Precision is what protects the context window. Every irrelevant passage you include is budget spent making the answer worse.
- Programmatic SEOSearch and discovery
Generating many pages from a structured data set, typically one page per combination of entity and attribute.
It works when each generated page answers a real question that somebody actually asks. It fails, and is penalised, when the pages are permutations with no distinct value, which is the usual outcome when the data set is chosen for size rather than for demand.
- Prompt injectionData and privacy
Content that tries to override an assistant's instructions, for example text on an indexed page telling the assistant to ignore its rules.
It matters more than teams expect for site assistants, because the attack surface is your own indexed content and anything a visitor can type.
Q
- QualificationConversion
Establishing whether a visitor is a plausible customer, and for what.
Qualification inside a chat works when it feels like helping and fails when it feels like a form. Two questions is usually the ceiling.
- QuotableSearch and discovery
Written so a specific sentence can be lifted and attributed without losing its meaning.
Quotable writing leads with the claim, keeps the sentence self contained, and avoids pronouns pointing at a previous paragraph.
R
- Rate limitDeployment
A cap on how many requests can be made in a period, applied by a model provider or by you to a visitor.
Provider limits cause outages. Your own limits are the main defence against a single visitor running up a bill.
- RecallRetrieval
The share of relevant passages that retrieval actually returns.
High recall with low precision floods the prompt with noise and dilutes the answer. The two have to be traded against each other deliberately, not tuned by feel.
- RedactionData and privacy
Removing or masking sensitive values from a transcript before it is stored or shown.
Redaction at write time is the only kind that protects you, because redaction at read time means the raw value was already stored.
RelatedPIIRetentionTranscript
- RefusalRetrieval
The assistant declining to answer because the supplied passages do not cover the question.
Refusal rate is a health metric, not a failure metric. An assistant that never refuses is either exceptionally well sourced or quietly making things up.
- ReindexingDeployment
Rebuilding the index after content changes, either on a schedule or triggered by a publish event.
- RerankingRetrieval
A second pass that reorders the passages retrieved by vector search, usually with a slower model that reads the question and each passage together.
It fixes the common case where the right passage was retrieved but ranked fourth, which matters because most prompts only include the top two or three.
RelatedRetrieval augmented generationCosine similarityRecallPrecision
- Response timeSupport
How long a visitor waits for a person once a conversation has been escalated.
Promising live handoff you cannot staff is worse than not offering it. An honest queue time beats a broken promise of immediacy.
- RetentionData and privacy
How long transcripts and related data are kept before deletion.
Retention is a product decision with legal consequences, not a setting. Shorter retention costs you the ability to debug and to find content gaps.
RelatedData boundaryTranscriptDPA
- Retrieval augmented generationRetrieval
Looking up relevant passages from your own content at question time and handing them to a language model along with the question, instead of expecting the model to already know your content.
It turns the model's job from recall into reading comprehension, which is a far easier job and the reason this approach works at all for content the model has never seen.
S
- Schema.orgSearch and discovery
The shared vocabulary of types and properties used by structured data across search engines.
Using the right type matters less than using it honestly. Marking a thin page as an Article does not make it one.
- Search engine optimisationSearch and discovery
Structuring a site so search engines can crawl, understand and rank it. The long standing discipline that answer engine and generative engine optimisation build on rather than replace.
The foundations still apply: crawlable pages, correct canonicals, a maintained sitemap, and content that answers the query. What changed is that a second class of reader now consumes the same pages and quotes from them rather than linking to them.
RelatedAnswer engine optimisationCanonicalSitemapStructured data
- Semantic searchRetrieval
Search that matches on meaning rather than exact words, so a question about cost can match a passage about pricing.
It removes the keyword brittleness of older site search, but it also removes the predictability, which is why answers need citation.
- SessionAnalytics
One continuous visit, which may contain several conversations.
Session level analysis is what connects a chat question to whether the visitor then went to pricing, which is the only link that shows commercial value.
- SitemapSearch and discovery
An XML file listing the pages of a site worth indexing.
It is also the cleanest input for training a site assistant, because it is a list you maintain of pages you consider canonical.
RelatedCrawlrobots.txtIndex
- Source citationRetrieval
Naming the page or file an answer was drawn from, shown alongside the answer so the reader can check it.
It is the cheapest debugging tool available. When an answer is wrong, the citation tells you immediately whether retrieval fetched the wrong passage or the model misread the right one, and those need completely different fixes.
- Source limitDeployment
The number of pages or files a plan allows an assistant to answer from.
It reads as a restriction and behaves as a guardrail. More sources produce worse answers unless each one earns its place, so on lower tiers the limit is often doing useful work rather than getting in the way.
- Source selectionRetrieval
Deciding which pages and files an assistant is allowed to answer from.
- StalenessData and privacy
The gap between what a page says now and what the index still believes it says.
Staleness produces the worst class of wrong answer: the confidently outdated one, phrased with total certainty because retrieval did its job on a stale passage.
- StreamingPerformance
Sending the answer word by word as the model produces it, rather than waiting for the whole response.
It changes perceived speed far more than it changes real speed, which is why it is worth doing even when total time is unchanged.
RelatedTime to first tokenLatency
- Structured dataSearch and discovery
Machine readable markup describing what a page is, most commonly JSON-LD using schema.org types.
It has to restate the visible page, never a more optimistic version of it. A FAQ schema that disagrees with the visible FAQ is a defect in both.
- SubprocessorData and privacy
A third party that processes data on your behalf, such as a model provider or a hosting platform.
The list has to be public and current, because a customer's own compliance obligations depend on knowing who is in the chain.
RelatedDPAData boundaryRetention
- System promptModels
The standing instruction given to the model before any visitor message: who it is, what it may answer, what it must refuse, and how to behave when unsure.
T
- TemperatureModels
A setting controlling how much randomness the model uses when choosing words.
For a grounded website assistant it should be low. Creativity is not a feature when the job is repeating what a page already says.
RelatedSystem promptGrounding
- TicketSupport
A support request tracked to resolution by a person.
Pre sale questions and post sale tickets look similar in a chat log and need completely different routing. Conflating them corrupts both queues.
- Time to first tokenPerformance
How long until the first word of the answer appears.
- TokenModels
The unit a model reads and bills in, roughly three quarters of a word in English.
Tokens are the currency of both cost and latency. A prompt stuffed with marginal passages costs money twice, once at the meter and once in answer quality.
- TranscriptSupport
The full record of a conversation, including what the assistant retrieved and what it refused.
A transcript without the retrieval trail tells you what happened but not why, which makes it useless for fixing the underlying content.
V
- Vector databaseRetrieval
A store designed to hold embeddings and answer the question of which stored vectors are closest to a query vector, quickly, across a large collection.
The database is rarely the bottleneck. Source selection and chunk boundaries decide answer quality long before the index does.
W
- WebhookDeployment
An HTTP request the assistant sends to a URL you control when something happens, such as a lead being captured or a conversation escalating.
Webhooks are the honest general purpose integration. They require somewhere to receive them, which is why they are often described as an integration with tools that merely accept them.
- Website trainingDeployment
Pointing an assistant at pages on your site so their content becomes retrievable. The pages are crawled, extracted, split and embedded; nothing about the model changes.
Calling it training invites the assumption that the model learns your business. It does not. It gains the ability to look your content up, which is why a page edit takes effect at the next index rather than at a retraining run.
- Widget placementDeployment
Which pages show the assistant, and where on those pages.
The right answer is usually not everywhere. Pages with a single intended action, such as checkout, are usually better without it.
RelatedLauncherEmbedConversion rate
No term matches that. or read the blog.
About this glossary
85 terms. The test for inclusion is whether knowing it changes a decision when you are building, buying or running an assistant on a website. Terms that only matter to people training models are deliberately left out.
Because a definition alone rarely helps. Knowing that an embedding encodes meaning is trivia. Knowing that it encodes meaning but not truth or recency is the thing that explains why an assistant confidently quoted a page from two years ago.
Yes. Search matches the term, its definition and its category. The category chips and the A to Z rail combine with the search, and letters with no terms are disabled rather than clickable into an empty result.
Every term here is one that changes a decision when you are putting an assistant on a website. Terms that only matter to people training models are left out.
No. They describe how website assistants work in general. Where Creobot makes a particular choice, the product pages say so rather than the glossary.
Yes. Every term has its own anchor, in the form /glossary/#t-chunking, and the link opens the page with that term in view.
Signal
Still deciding what to ask a vendor?
Creobot opens to teams in order, through the waitlist. Ask a direct question and we will answer it, including where Creobot is the wrong fit.