The way people find information has changed, and the language for it is still being invented. This is a comprehensive resource to stay up to speed on proper AI Search, Answer Engine Optimization, and the agentic web terminology.

Every term below consists of a plain-English definition, and a short descriptor explaining why it matters to your brand.

Worth bookmarking. This vocabulary has a short half-life. Products get renamed, crawlers get split into three, proposed standards get abandoned mid-draft. We keep this page current, which makes it more useful to come back to than to copy.

How to read this glossary. Terms are grouped into 15 sections so you can learn the subject in order, and indexed A–Z at the bottom so you can easily find what you're looking for. If you're new to this, start with Core Concepts. If you're building a measurement program, start with Measurement and Metrics.

Contents

  1. Core Concepts
  2. AI and LLM Fundamentals
  3. How AI Answers Get Made
  4. Prompts and Query Behavior
  5. Citations and Sources
  6. Measurement and Metrics
  7. Platforms and Surfaces
  8. AI Crawlers and Bots
  9. Access Control, Licensing and Monetization
  10. The Agentic Web and AI Commerce
  11. Content Strategy for AI Search
  12. Entities and Structured Data
  13. Traffic, Analytics and Attribution
  14. SEO Carryover
  15. Risks and Failure Modes

1. Core concepts

AI Search

A type of search that uses artificial intelligence to understand what someone is asking and write them a direct answer, usually with links back to sources that influenced the answer. The system decides on your behalf which brands get named, which is the entire reason this glossary exists. Our AEO Guide walks through the whole discipline from first principles.

Answer Engine

Any AI system that responds to a question with prose rather than a page of ranked results. ChatGPT, Gemini, Perplexity, Copilot, Claude and Google's AI surfaces are all Answer Engines, and each one reaches its own conclusions from its own sources.

Generative Engine

A near-synonym for Answer Engine, preferred in academic work because it emphasizes that the answer is generated rather than retrieved. The term comes from the 2023 research paper that coined GEO.

Answer Engine Optimization (AEO)

The discipline of engineering content to become the cited source in AI-generated responses. Where SEO tried to win a position on a page, AEO tries to win a sentence inside an answer. See our full explainer on what Answer Engine Optimization is.

Generative Engine Optimization (GEO)

The same discipline as AEO under a different name. The term was coined in a 2023 academic paper, and was made popular by Andreessen Horowitz. Most practitioners now use AEO and GEO interchangeably, and we have written on why the distinction rarely matters.

LLMO (Large Language Model Optimization)

Another synonym for AEO, favored by people who want to name the technology rather than the interface. It has lost the naming race but still appears in job titles.

AI SEO

An umbrella term covering both optimizing for AI Search and using AI tools to do SEO work. The ambiguity is the problem, so prefer AEO when you mean the former.

Terminology war

The unresolved industry argument over whether to call this AEO, GEO, LLMO, GAIO, AIO or AI Search optimization. A 2025 survey of more than 200 SEO specialists found a 36% plurality preferred "AI Search optimization," with no option taking a majority, so there's no consensus to defer to.

Relevance engineering

A reframing of AEO that emphasizes engineering how a machine understands your content rather than chasing an algorithm. Useful as a mindset, and a signal that the person using it has read past the introductory posts.

AI visibility

Whether and how often your brand appears in AI answers to the questions your buyers actually ask. Visibility is the presence question, and it's the floor of any AEO program rather than the ceiling. Profound tracks it through Answer Engine Insights.

Zero-click search

A search query that ends right on the search results page. Profound's behavioral study with Kevin Indig and Clickstream Solutions found 92.8% of AI shopping tasks ended with no meaningful click to the open web.

Answer layer

The AI-generated response that now sits between your content and your buyer. Your content still does the work, but the answer layer decides how much credit you get for it.

On-site / On-page

The first-party content you control and can edit today. Your product pages, comparison pages, docs, blog and FAQs.

Off-site / Off-page

Everything said about you on properties you don't control. Review sites, Reddit, trade press, listicles and competitor comparison pages, which collectively drive far more of your AI visibility than most teams expect. Profound's citation-source analysis breaks the mix down by category and platform.

Third-party validation

Independent confirmation of your claims from sources the model already trusts. AI answers lean on outside corroboration, so being the only site that says you're the best is worth close to nothing.

Third-party signals

What independent sources such as review sites, trade press, forums and analysts say about you. AI answers lean heavily on outside confirmation, so your presence beyond your own site shapes the answer as much as your presence on it.

Brand SERP

What appears when someone searches your brand name directly. It remains the fastest read on whether machines understand who you're.

Agentic search

Search performed by software on a person's behalf, researching options, comparing them and sometimes completing the purchase. When the reader is a program, page design stops being a persuasion problem and becomes a parsing problem.

Conversational search

Search conducted as a dialogue, where follow-up questions replace new queries. It means your brand has to survive not one answer but a sequence of them.

AI assistant

The consumer-facing product wrapped around a model, such as the ChatGPT app or the Gemini app. Assistants are where most AI Search actually happens, which is why measuring the model isn't the same as measuring the surface.

Prompt

The question a person types or speaks to an AI system. In AEO the prompt replaces the keyword as the unit of measurement. It's longer and far more specific of intent. Profound's study of ChatGPT intent across more than 50 million prompts breaks down what people actually ask for.

2. AI and LLM fundamentals

Artificial intelligence (AI)

Software that performs tasks we associate with human thinking, such as understanding language and recognizing patterns. It's the umbrella over every other term in this glossary.

Machine learning

The approach behind modern AI, where systems learn patterns from data instead of following hand-written rules. It's why nobody at OpenAI can tell you exactly why the model named your competitor.

Generative AI

AI that produces new content rather than just classifying existing content. Answer Engines are generative AI pointed at search.

Large language model (LLM)

A model trained on enormous volumes of text that predicts language, and in doing so writes the answers your buyers read. Every Answer Engine sits on top of one.

Foundation model

A large, general-purpose model trained once at great expense and then adapted to many uses. GPT, Gemini, Claude and Llama are foundation models.

Frontier model

The newest and most capable models available at any moment. The label moves constantly, and the practical consequence is that your AEO baseline resets every few months.

Transformer

The neural network architecture behind modern language models, and the T in GPT. It lets a model weigh the relationships between all the words in a passage at once, which is what makes it good at meaning rather than matching.

Attention

The mechanism inside a transformer that decides which parts of the input matter most for the next word. Attention is unevenly distributed, which is why where a fact sits inside your page affects whether the model uses it.

Token

The unit of text a model reads and writes, roughly three quarters of a word. Pricing, speed, context limits and most AEO impression metrics are all counted in tokens.

Context window

The amount of text a model can hold in mind at once. Content that states its key facts early survives a crowded context window, and content that buries them doesn't.

Parametric memory

What the model knows from training, stored in its weights rather than looked up. This is the knowledge that persists when the model can't reach the web, and it's the reason stale facts about your company outlive the pages that caused them.

Non-parametric memory

Knowledge the model pulls in at answer time from an external source. Retrieval is non-parametric memory, and it's the part of the system you can influence this week rather than next year.

Training data

Everything the model learned before release. It's fixed until the company trains a new model, so anything you changed after that date exists for the model only if retrieval finds it.

Training data cutoff

The date after which a model's built-in knowledge stops. Your launches, rebrands and price changes since that date are invisible unless a live source carries them.

Effective cutoff

The point where a model's knowledge actually gets thin, which is usually earlier than the stated cutoff because recent events are underrepresented in training data. Assume the model's real picture of your category is older than the documentation claims.

Fine-tuning

Extra training that adapts a general model to a specific task or domain. It sharpens behavior without rebuilding the model, and it's not something brands can do to change how a public model sees them.

RLHF (Reinforcement Learning from Human Feedback)

Training a model against human preference ratings so its answers feel more helpful. RLHF is why models sound confident and agreeable, including when they're wrong about you.

Knowledge distillation

Training a smaller model to imitate a larger one, producing something cheaper and faster that behaves similarly. Distilled models power many of the fast, free tiers where a lot of AI Search actually happens.

Mixture of Experts (MoE)

An architecture that routes each input to a subset of specialized sub-networks rather than the whole model. It buys capability at lower cost, and it's one reason answers vary between sessions.

Multimodal AI

Models that handle images, audio and video alongside text. Your charts, screenshots, product photos and video transcripts are now readable inputs, not decoration.

Temperature

A setting that controls how much randomness the model uses when picking each next word. Higher temperature means more variety, which is part of why the same prompt doesn't return the same answer twice.

Non-determinism

The property that identical inputs can produce different outputs. Setting temperature to zero doesn't fix this: research from Thinking Machines Lab in September 2025 traced the cause to a lack of batch invariance in GPU kernels, and found one model produced 80 unique completions across 1,000 runs at temperature zero.

Answer variability

The natural spread in AI responses when you run the same prompt repeatedly. It's the single most important reason to measure AI visibility with repeated sampling instead of spot checks, because one manual check tells you almost nothing.

Synthetic data

Model-generated training data used to supplement or replace human-written text. It makes training cheaper and raises the long-term risk of models learning from their own output.

Inference

The act of running a trained model to produce an answer. Every AI Search response is an inference, and inference cost is why free tiers use weaker models than the ones in the benchmarks.

AGI (Artificial General Intelligence)

A hypothetical AI matching human ability across essentially all intellectual work.

3. How AI answers get made

Retrieval

The step where the AI reads live web pages while answering instead of relying only on training. You can watch it happen as the "searching the web" moment before the answer appears, and it's the mechanism that lets fresh content influence answers in days rather than years.

RAG (Retrieval-Augmented Generation)

The architecture behind AI Search, where the system fetches current documents and then writes its answer from them. RAG is the reason AEO is a real discipline and not just a waiting game on the next model release.

Grounding

Tying an answer to specific retrieved sources rather than to the model's general impressions. Grounded answers arrive with citations, and ungrounded answers arrive with risk.

Index

The store of pages an AI system has already ingested and searches through when it answers. Each company keeps its own, which is why ChatGPT and Claude can cite completely different sources for the same question.

Crawler

The bot that fetches pages across the web to build an index or a training set. If a crawler can't reach or render your page, the model never learns the page exists.

Re-indexing

When a page gets fetched again and the stored copy is replaced with the current version. Publishing a correction does nothing until a crawler returns, which is why fixing an outdated page is a two-step process you only control half of.

Pickup lag

The delay between a source going live or being updated and AI answers reflecting the change. Pickup lag is why AEO reporting needs a few weeks of patience before you conclude a tactic failed.

Query fan-out

The habit of expanding one user prompt into several background searches before answering. It's Google's own documented terminology, though the phrase doesn't actually appear in the patents it gets attributed to, which use terms like synthetic, related and implied queries. Profound's analysis of 10,000 prompts found roughly 1.4 to 2 fan-out queries per prompt, and our query fan-out tooling shows the searches your prompts actually trigger.

Synthetic query

A query the system generates itself rather than one the user typed. Your content competes in searches no human ever ran, which is why keyword-exact thinking breaks down here.

Implied query

A question the system infers the user actually meant. It's how "is this any good" becomes a search for reviews you've never seen.

Custom corpus

A curated document collection a system searches instead of, or alongside, the open web. Named in Google's published patent application "Query response using a custom corpus," and relevant because it means not all retrieval happens over the public index.

Passage retrieval

Fetching a specific chunk of a page rather than the whole page. Retrieval systems pull passages, so a page is really a collection of independently competing sections.

Chunking

Splitting a document into smaller pieces so a retrieval system can index and fetch them individually. How your page gets chunked determines what can be quoted from it, and you influence that through structure.

Semantic chunking

Chunking along meaning boundaries so each piece is self-contained. Writing sections that make sense on their own is the single most portable AEO tactic, because a section that can stand alone can be cited alone.

Late chunking

An embedding technique that encodes the whole document first and then splits it, so each chunk retains some surrounding context. It reduces the penalty for pages whose meaning depends on earlier sections.

Embeddings

Numerical representations of meaning that let a system judge how close a piece of content sits to a question. Relevance in AI Search is measured this way, not by keyword overlap.

Vector database

A store built to search embeddings by similarity rather than by exact match. It's the infrastructure underneath "find me content that means roughly this."

Cosine similarity

The standard math for scoring how close two embeddings are. It's the number that decides whether your passage makes the candidate set.

Semantic search

Retrieval based on meaning rather than exact wording. A page can surface for a question that shares none of its vocabulary and miss one that shares all of it.

Candidate set

The shortlist of retrieved passages the model actually considers before writing. Getting into the candidate set and getting cited are two different battles, and most content loses the first one.

Reranking

A second scoring pass that reorders retrieved candidates by relevance before the model reads them. Rerankers are more accurate and more expensive than the initial retrieval, and they're where borderline content gets cut.

Bi-encoder

A retrieval model that embeds the query and the document separately, which is fast and approximate. It handles the first pass over millions of documents.

Cross-encoder

A model that reads the query and document together for a more accurate relevance score. Slow, precise, and typically used for reranking the top candidates.

Synthesis

The step where the model writes one answer from multiple retrieved sources. Synthesis is where your carefully worded positioning gets compressed into half a clause, accurately or otherwise.

Position bias

The tendency of models to weigh information at the beginning and end of their context more heavily than the middle. Documented as the "lost in the middle" effect, and a practical argument for putting your key claim first.

LLM-as-a-judge

Using one model to evaluate another model's output at scale. Most AI visibility and sentiment scoring is built this way, which is worth knowing when you read a vendor's sentiment number.

Hallucination

Confident, fluent, false output, including claims about your company that no source ever made. Monitoring for it's now part of brand protection rather than a curiosity.

Citation recall

The share of generated sentences that are actually supported by the sources cited. A widely cited 2023 study measured 51.5%, which is frequently and wrongly repeated as "51.5% of AI answers are wrong."

Citation faithfulness

Whether a cited source says what the answer claims it says. Low faithfulness means you can be cited and still misrepresented, which no visibility metric will catch.

4. Prompts and query behavior

Prompt list

The defined set of questions you track because they're the ones your buyers ask. Choosing this list is the highest-leverage decision in an AEO program, and most teams rush it.

Branded prompt

A prompt that names your company or product, such as "is Profound the best AI analytics software?" Branded prompts test what the models believe about you, not whether they know you exist.

Unbranded prompt

A prompt about a problem or category that names no company, such as "best AI analytics software." This is where shortlists get formed and where visibility is genuinely won or lost.

Comparative prompt

A prompt asking the AI to weigh named options against each other. The answer functions as an instant analyst report assembled from whatever content the model has learned to trust.

Buyer-journey prompts

The full sequence of questions a buyer asks from early education through final evaluation. Brands that show up in the educational prompts tend to enter the comparison stage as the presumed favorite.

Commercial prompt

A prompt with clear buying intent, such as asking for the best option or a recommendation. Profound found commercial conversations rose from 13.9% to 19.2% of ChatGPT's mix across 7.5 million conversations between 2025 and mid-2026, which is the trend line that makes AEO a revenue conversation.

Transactional prompt

A prompt aimed at completing an action rather than learning something. These are the prompts agentic commerce is being built to close.

Navigational prompt

A prompt where the user already knows where they want to go. Small as a share of volume, and the one category where being the obvious answer is easy.

Informational prompt

A prompt seeking understanding rather than a purchase. It's the largest addressable category for content teams and the one most likely to be answered without a click.

Generative prompt

A prompt asking the model to make something rather than find something. Profound measured this as the single largest intent category in ChatGPT at 37.5%, which is a useful corrective to treating every prompt as a search.

Prompt volume

How much real-world search demand a given prompt represents. Note that vendors use this phrase two ways, sometimes meaning panel-derived consumer demand and sometimes just meaning how many times a tool sampled the prompt, so always ask which; Profound's Prompt Volumes means the former.

Prompt space

The full universe of ways a question can be asked about your category. You can't track all of it, so the discipline is choosing a representative sample rather than a flattering one.

Prompt clustering

Grouping semantically similar prompts so you can measure a topic instead of a hundred phrasings. It's how you stop reporting noise as movement.

Prompt engineering

Designing inputs to get better outputs from a model. In marketing it mostly describes building the prompt sets used to audit how models describe your brand.

Prompt tracking

Running a fixed set of prompts across AI platforms on a schedule to measure visibility, citations, sentiment and accuracy. Without a schedule you've anecdotes, and with one you've a baseline you can defend. This is the core of Profound’s Answer Engine Insights.

Personas

Configuring the tracked prompt to reflect a specific buyer type, region or context. The same question from a procurement lead and a developer can return different brands, and averaging them hides the finding.

Search variation

The difference in which sources a system retrieves across repeated runs of the same prompt. Distinct from answer variation, and the reason citation metrics need sampling too.

Execution

One individual run of one tracked prompt against one platform. Executions are the raw unit that every visibility percentage is calculated from, so the count matters when judging whether a number is meaningful.

5. Citations and sources

Citation

A reference an AI tool provides pointing back to a website. Citations reveal which content the engines trust in your category, which makes them the most actionable data AI Search produces. Profound's Citations view sorts them into owned, competitor, earned media, PR wire, social and institutional buckets.

Source

Any page or document the system retrieved to build the answer, whether or not it appears as a visible link. Retrieved-but-uncited sources still shape what the answer says about you.

Inline citation

A link embedded directly in the body of the answer, usually next to the claim it supports. Inline placement carries more weight with readers than the source list at the bottom.

Sources block

The list of references displayed separately from the answer text. SE Ranking found roughly 93% of review-platform links in AI Overviews appeared here rather than inline, so most citations are less visible than brands assume.

Cited domain

The root domain of a source that appears in an answer's citations. Rolling citations up to domain level is how you see who really owns the category's supply of evidence.

Brand mention

Any instance of your brand being named in an AI answer, with or without a link. Mentions are the raw material of share of voice, and a mention without a link still moves a shortlist.

Authentic brand mention

Accurate, relevant content about your brand inside a source the AI already cites. This is the unit of work in offsite AEO, because improving what a trusted source says about you is faster than becoming a trusted source.

New mention

Getting your brand added to an article the AI already cites for your buyers' questions but which currently omits you. It's the highest-leverage offsite play available, since the source has already earned the citation.

Mention refresh

Correcting an existing mention that's inaccurate, out of date, or positions you poorly. Cheaper than earning a new placement and usually more urgent.

Reverse mention

A reciprocal arrangement where two brands agree to mention each other in their content. Effective, and worth structuring carefully so it doesn't read as a link scheme.

Co-citation

Being cited in the same answer as another brand. Persistent co-citation is how models learn who your competitors are, and it can put you in a comparison set you never chose.

Co-occurrence

Your brand and a concept appearing together often enough that the model associates them. It's the mechanism behind category ownership, and it's built through repetition across many sources rather than one great page.

Brand-owned citation

A citation pointing to a domain you control. Profound's analysis of 11.84 billion citations found a median of roughly 57% of AI citations point to some company's own website, rising to 69% on Gemini, which means owned content matters more than the "you can't control it" narrative suggests.

Earned citation

A citation pointing to independent media, review sites or third-party coverage. Profound measured a median of about 30% of ChatGPT's citation mix and 17% in AI Overviews, so the balance shifts meaningfully by platform.

Social citation

A citation pointing to a community or social platform such as Reddit, YouTube or LinkedIn. Profound found social sources at roughly 15% in AI Overviews and 14% in AI Mode, which is enough to matter and not enough to build a strategy on alone.

Review-site influence

The weight AI answers give to platforms like G2, Capterra, Gartner Peer Insights and Trustpilot. SE Ranking found 34.5% of AI Overviews on 30,000 US commercial-intent keywords cited at least one review platform, ranging from 17.1% to 49% by query type, and vendors disagree on how much any single site contributes. Profound data behind Kevin Indig's regression across 500 G2 categories found the reviews-to-citations relationship real but weak.

Response

The complete answer an AI Search engine returns for one prompt. It's the container every other metric is calculated against, so agreeing on what counts as one response is step one of any measurement program.

Inline domain

The root domain of a link embedded directly in the answer text, as opposed to one listed in the sources block. Separating inline from sources-block domains tells you which sites are trusted enough to be quoted rather than filed.

Citation gap

Prompts and topics where AI engines cite your competitors but not you. This is the most directly actionable output of AI Search measurement, because it's a to-do list disguised as a report. Profound surfaces these in Answer Engine Insights.

Citation velocity

How quickly a source accumulates citations over time. Treat claims about it carefully, since at least three vendors define it incompatibly and no major platform exposes it as a standard field.

Citation hijacking

When an answer attributes your content, data or claim to a different source. It costs you the credit without reducing the effort, and it usually traces back to someone else republishing you more crawlably.

6. Measurement and metrics

Read this before you compare two vendors' numbers. Most AEO metric names aren't standardized. Profound publishes its own formulas in the Profound glossary, and we'd encourage you to demand the same of any vendor. The same brand's "share of ChatGPT" can be reported as 1.8%, 11.3%, 16.7% or 60% depending entirely on the denominator, and every one of those figures can be technically correct. Whenever you see a percentage, ask what it's a percentage of.

Visibility

Whether your brand appeared in a given AI answer, measured as a percentage of tracked answers. Profound defines it as a binary metric, so appearing once and appearing five times in the same response count the same.

Visibility score

A single number summarizing how findable you're across your tracked prompts. Some vendors report it as a straight presence percentage and others as a composite blending presence, position and prominence, so the label alone tells you nothing.

Presence rate

The share of tracked responses where your brand appears at least once. The cleanest and least manipulable visibility metric, and a good default for executive reporting. Some vendors call this brand presence.

Brand perception

How AI answers characterize your brand: the adjectives, the category they place you in, and the use cases they attach to you. It goes beyond sentiment, because an answer can be positive about you and still describe a company you don't recognize.

Mention frequency

A count of how many times your brand is named, including multiple mentions within a single answer. Distinct from visibility, and useful for spotting answers where you dominate rather than merely appear.

Share of voice

Your share of all brand mentions across your tracked prompts, compared against every competitor. Read it as a share and not a count, because two new entrants can shrink your percentage in a month where your actual visibility rose. See Profound's definition and worked example.

Citation share

How often a specific page or domain is cited as a proportion of all citations. Profound measures it across all citations, and it can be read at either page or domain level, which changes the number substantially.

Share of sources

The percentage of all sources in a set of answers that come from one domain. Similar in spirit to citation share and frequently used as a synonym, which is fine as long as you don't mix the two in one chart.

Brand position

Your rank relative to other brands named in the same answer. Being mentioned fourth in a list of ten isn't the same result as being mentioned first, and presence-only metrics erase that distinction.

Brand placement

How early in the answer your brand first appears, often expressed as a decile of the text. Earlier placement means more readers actually see it, since most people don't finish the answer.

Prominence

A combined read on position, placement and how much of the answer is about you. Prominence is where AEO reporting starts to resemble a real competitive metric instead of a checkbox.

Depth of mention

How substantively the answer discusses your brand rather than just naming it. A sentence of accurate positioning beats an appearance in a comma-separated list, and no standard formula for this exists yet.

First-mention rate

How often you're the first brand named in an answer. A strong proxy for default-choice status in a category.

Sentiment

Whether an answer portrays your brand positively, negatively or neutrally, usually reported as a percentage for each. Because models compress opinion from across the web, AI sentiment is a distilled reading of your reputation rather than a survey of it. Profound tracks it as Sentiment.

Aggregated sentiment score

The ratio of positive sentiment occurrences to all sentiment occurrences across your tracked answers, which is how Profound calculates it. Useful as a trend line, and shaky as an absolute number, because it's typically scored by another model.

Sentiment drift

A gradual shift in how models characterize you over time. Drift usually traces to a specific source changing or a new one entering the citation pool, which makes it diagnosable rather than mysterious.

Accuracy score

A percentage comparing your real pricing, positioning and product details against what AI actually says about you. It's the metric that turns AEO from a marketing exercise into a brand-risk one, and it's what Profound built FactCheck to measure.

Answer inclusion rate

The share of answers for a topic that include your brand. Effectively a scoped presence rate, and worth defining explicitly since no cross-vendor standard exists.

Competitive overlap

How often you and a specific competitor appear in the same answers. High overlap means you're being evaluated head-to-head whether or not your sales team frames it that way. Profound's study of AI shopping journeys found chosen brands held 24% share of voice against 11% for passed-over brands.

Source visibility

How often a given domain appears across a set of AI answers over time. This is the metric to watch when deciding which third-party sites deserve your PR budget.

Impressions

An estimate of how much exposure a mention actually received, usually by weighting mentions against search volume. Definitions vary enormously, and at least one major vendor uses Google search volume as the weighting input, which makes it a hybrid rather than a pure AI metric.

Position-Adjusted Word Count (PAWC)

An academic metric counting how many words of an answer are attributable to a source, weighted by where they sit. It comes from the original GEO paper, a 2023 preprint published at KDD 2024, and remains the most rigorous published attempt at measuring answer share.

Crawl-to-refer ratio

How many times an AI company crawls your site for every visitor it sends back. Cloudflare's 2025 measurements showed enormous spreads by operator, and the ratio has become the central exhibit in the publisher-compensation argument.

Data cleansing

Filtering out duplicate, malformed or non-representative responses before calculating metrics. Unglamorous, and the difference between a trustworthy baseline and a noisy one.

Pages versus domains

The choice of whether to attribute a citation to the specific URL or the root domain. Domain-level rollups make big publishers look dominant and page-level views show you which specific article to go fix.

Answer volatility

How much your measured visibility swings between sampling runs without any underlying change. Establishing your volatility band first is what stops you from reporting noise as a win.

7. Platforms and surfaces

AI Overviews

Google's AI-generated summary above the traditional results, assembled from multiple sources. Google reported more than 2.5 billion monthly users in May 2026, and Pew found people clicked a traditional result in 8% of visits with an AI summary against 15% without.

AI Mode

Google's full conversational search experience, where follow-ups replace new searches and cited sources replace the ten blue links. Google reported it passing one billion monthly users in May 2026, though independent measurement suggests actual query share remains small.

SGE (Search Generative Experience)

Google's Search Labs experiment, announced in May 2023 and superseded when AI Overviews launched publicly in May 2024.

ChatGPT

OpenAI's assistant and the largest single surface in AI Search, with OpenAI reporting more than a billion people using it as of August 2026 and a last-published weekly-active figure of 800 million in October 2025. Web search is now automatic rather than a separate mode, so the old distinction between ChatGPT and "ChatGPT Search" has effectively dissolved.

ChatGPT Atlas

OpenAI's browser, launched in 2025 and shut down in August 2026. Worth knowing only so you can date any strategy deck that still features it.

Gemini

Google's assistant, formerly Bard, reported at over one billion monthly active users in August 2026. Its citation mix skews unusually heavily toward brand-owned sources, which makes your own site disproportionately valuable there.

Perplexity

An Answer Engine built around transparent sourcing, which made it the favorite of people who wanted to see the citations. Small in traffic share relative to ChatGPT and Gemini, and consistently over-indexed in AEO discourse relative to its size.

Comet

Perplexity's agentic browser. Notable as an early example of the browser itself becoming the agent rather than a window onto one.

Microsoft Copilot

Microsoft's assistant across Bing, Windows and Office. Its practical AEO importance is enterprise reach rather than consumer query volume.

Copilot Search

Microsoft's AI answer experience inside Bing, sometimes still called Bing Copilot. Bing Webmaster Tools exposes AI-related data that no other platform offers publicly, which makes it useful out of proportion to its share.

Claude

Anthropic's assistant, with web search available. Anthropic is the only major operator that documents no robots.txt exemption for user-triggered fetches, which means blocking its bots genuinely removes you from its answers.

Grok

The assistant from SpaceXAI, formerly xAI, which SpaceX acquired in February 2026. Tightly coupled to X, and its answers lean unusually hard on real-time social content.

Meta AI

Meta's assistant that’s embedded across its apps. Enormous nominal reach, minimal referral traffic, and a separate crawler specifically for being cited in its answers.

DeepSeek

A Chinese lab whose efficient open-weight models reset industry cost expectations. Relevant to AEO mostly because open-weight models get embedded in products you'll never be told about.

Alexa for Shopping

Amazon's shopping assistant, which absorbed the former Rufus in May 2026 and hands off to Alexa+ for broader tasks. Amazon reported Rufus alone helped over 300 million customers in 2025, making this the largest AI surface most B2C brands aren't measuring.

Apple Intelligence

Apple's on-device AI layer, including the rebuilt Siri AI. Its AEO significance is distribution, since it reaches users who never open an AI app.

Brave Leo

Brave's built-in assistant, backed by its independent search index. Brave is one of the few players with its own index rather than a licensed one, which is why its answers can differ noticeably.

Duck.ai

DuckDuckGo's privacy-focused chat interface, alongside its Search Assist summaries. Small share, and a useful check on whether your visibility holds up outside the big four.

Gemini Notebook

Google's research and document tool, renamed from NotebookLM in July 2026. It fetches sources on the user's instruction, which is why it has its own crawler token.

AI browser

A browser with an assistant or agent built into the browsing experience itself. The category is churning fast, with several launches and shutdowns inside a single year, so build for the protocol rather than the product.

Answer Engine market share

The distribution of AI Search usage across platforms. Be careful which denominator is being used, since Similarweb's web-traffic share and SparkToro's clickstream share of all searches produce very different pictures, the latter putting AI tools at roughly 3.2% of US desktop search activity in late 2025. Profound's free Profound Index tracks brand visibility across engines weekly.

8. AI crawlers and bots

Why the token matters. Most operators now run three separate bots: one for training, one for building a search index, and one that fetches a page because a user asked. Blocking the wrong one can remove you from answers while doing nothing about training, so precision here's worth real money. Profound's Agent Analytics reads your server logs to show which bots actually reach you, and our crawlability audit flags the rules blocking them.

User agent

The identifying string a bot sends with each request. It's how you allow one purpose and deny another in robots.txt, and it's the level at which all of this is actually controlled.

Crawler purpose category

The classification of a bot by what it does rather than by who runs it. Cloudflare classifies verified bots by behavior across eleven values and notes that one bot can carry more than one, with Search, Agent and Training exposed as the presets you can actually act on.

Mixed-use crawler

A bot that serves more than one purpose, such as indexing for search while also collecting training data. Mixed-use bots are the reason "just block the AI bots" is bad advice.

GPTBot

OpenAI's training crawler. Blocking it limits training use and doesn't remove you from ChatGPT's answers.

OAI-SearchBot

OpenAI's crawler for building the search index behind ChatGPT's answers. This is the one to allow if you want to be cited, and OpenAI says opted-out sites won't appear in ChatGPT search answers though they can still show up as navigational links.

ChatGPT-User

OpenAI's fetcher for when a user or a GPT Action asks for a specific page. OpenAI documents that robots.txt rules may not apply to it.

OAI-AdsBot

OpenAI's crawler for validating advertiser landing pages. It only visits submitted ad destinations and isn't used for training.

ClaudeBot

Anthropic's training crawler, which honors robots.txt and supports crawl-delay. Straightforward, and the least ambiguous of the major training bots.

Claude-SearchBot

Anthropic's crawler for building the index behind Claude's web search. Allow it if you want Claude to be able to cite you.

Claude-User

Anthropic's fetcher for user-initiated retrieval. OpenAI, Google, Perplexity, Amazon and Meta all document robots.txt exemptions for their user-triggered fetchers, and Anthropic documents none for this one, warning instead that blocking it may reduce your visibility in Claude.

PerplexityBot

Perplexity's crawler for search and citation, explicitly not used for foundation-model training. Allowing it's a prerequisite for appearing in Perplexity's sources.

Perplexity-User

Perplexity's user-triggered fetcher, which its documentation states generally ignores robots.txt. A good illustration of why robots.txt is a request and not a wall.

Googlebot

Google's main search crawler, which also determines eligibility for AI Overviews and AI Mode. Blocking Googlebot removes you from Google entirely, so it's never the right tool for an AI-specific decision.

Google-Extended

A robots.txt control token that opts your content out of training and grounding for the Gemini apps and the Vertex AI Gemini API. It does not control AI Overviews or AI Mode, contrary to a great deal of published advice, and it's not a crawler with its own user-agent string. Google documents the actual controls on its AI features page.

GoogleOther

Google's general-purpose crawler for research and internal projects. Low-stakes, and included here because it turns up in logs and confuses people.

Google-GeminiNotebook

Google's fetcher for sources a user adds to Gemini Notebook, renamed from Google-NotebookLM in July 2026. The old token was honored only through August 2026, so log-parsing rules built before then need updating.

Google-Agent

Google's user-triggered fetcher for its hosted agents, added in March 2026. Like Google's other user-triggered fetchers it generally ignores robots.txt, and it marks the shift from crawling for indexes to fetching on behalf of a task.

Applebot

Apple's crawler for search and as a training source. If it finds no rules for itself, it follows your Googlebot rules, which catches a lot of site owners out.

Applebot-Extended

Apple's control token for opting out of AI training. Like Google-Extended it's a directive rather than a crawler. Separately, Apple treats an isAccessibleForFree: false signal in your structured data as a limit on AI-answer use.

Amazonbot

Amazon's crawler, which may be used to improve products and train Amazon's AI models. It honors robots.txt and respects noarchive and noindex.

Amzn-SearchBot

Amazon's retrieval crawler for search and Alexa answers, separate from training use. This is the one that matters for appearing in Amazon's AI shopping answers.

Amzn-User

Amazon's user-triggered fetcher, documented as possibly not following all robots.txt directives. Part of the same three-bot pattern as OpenAI and Anthropic.

meta-externalagent

Meta's crawler for training and direct indexing. Meta's own documentation writes it lowercase, which matters if you're pattern-matching.

meta-webindexer

Meta's retrieval crawler for citation in Meta AI answers. Allow it if you want to be quotable inside Meta's apps.

meta-externalfetcher

Meta's user-triggered and agentic fetcher, which its docs state may bypass robots.txt. Meta explicitly rejects the informal noai tag by name.

CCBot

Common Crawl's archiver, which builds the open dataset that dozens of third-party models train on. Blocking it affects far more downstream models than blocking any single lab's bot.

AI2Bot

The Allen Institute's crawler for open language model research. Nonprofit, transparent, and rarely worth blocking.

MistralAI-User

The only Mistral crawler token documented anywhere, and only via Cloudflare's third-party bot reference. Mistral publishes no crawler documentation of its own, so treat any MistralAI-Index or MistralAI-Training token you see in a blog post as unconfirmed.

Retired crawler tokens

Anthropic's Claude-Web and anthropic-ai no longer appear in its documentation, and ChatGPT-Agent doesn't appear in OpenAI's crawler documentation despite showing up in many blog posts. Audit your robots.txt against operator documentation rather than against other people's robots.txt, which is what our crawlability audit is for.

Signed agents

Cryptographically verifying that a bot is genuinely who it claims to be. Necessary because user-agent strings are trivially spoofable, and the reason Web Bot Auth exists.

Web Bot Auth

An emerging standard for bots to prove their identity using signed HTTP requests. It's what makes paid or permissioned crawling enforceable rather than aspirational.

Bot directory

A public reference listing known AI crawlers, their operators and their purposes. Useful, and worth cross-checking against the operator's own docs, since secondary directories carry tokens that were never real.

9. Access control, licensing and monetization

robots.txt

The plain text file at your site root that tells crawlers what they may fetch. It's a voluntary convention formalized as RFC 9309, which means it's a request rather than an enforcement mechanism. Profound’s technical AEO guide covers how to configure it for AI crawlers specifically.

Crawl-delay

A robots.txt directive asking a bot to slow down. Honored by some AI crawlers and ignored by others, including Google's.

noindex

A directive keeping a page out of a search index entirely. For Google, this is also what removes a page from AI Overviews and AI Mode eligibility.

nosnippet

A directive preventing any text excerpt from your page being shown. Along with max-snippet, this is the actual lever for limiting Google's AI features, not Google-Extended.

max-snippet

A directive capping how many characters of your page can be quoted. It gives you a dial rather than a switch, which is usually the more commercially sensible choice.

noarchive

A directive against storing a cached copy. Worth knowing that it's interpreted as "no AI training" by Microsoft and Amazon and ignored entirely by Google, which is a good example of how unsettled this area is.

noai

An informal meta tag asserting no AI use. It has no standards backing, Meta rejects it by name in its documentation, and treating it as protection is a mistake.

llms.txt

A proposed markdown file, now at v2, at your site root that points AI systems to the pages best representing you. Adoption by AI companies remains essentially unconfirmed: Google has stated Google Search does not use it and no major operator has documented using it, so treat it as low-cost insurance rather than a tactic. Our llms.txt explainer walks through the format.

llms-full.txt

A companion file containing your full documentation as plain text. It's a convention popularized by developer-documentation platforms rather than part of the original specification.

Content Signals

A proposed HTTP header and robots.txt extension letting you express permitted uses such as search, training or AI input separately. Cloudflare published it as a CC0 convention in September 2025 and extended it in July 2026, and the parallel standards work at the IETF's AIPREF working group remains active as of August 2026.

AI preferences

Ongoing standards work at the IETF to give publishers a machine-readable way to state how their content may be used by AI. It's the most likely source of an eventual real standard, and it's not finished.

Cloudflare AI Crawl Control

Cloudflare's tooling for allowing, blocking or charging individual AI crawlers, formerly called AI Audit. Announced in July 2026 and effective 15 September 2026, newly onboarding domains get training and agent crawlers blocked by default on ad-bearing pages while search crawlers stay allowed. The setting itself lives in Cloudflare's AI bot policies rather than in Crawl Control.

Pay-per-crawl

Charging AI companies for each page they fetch, typically enforced by returning an HTTP 402 response. Still limited in availability, and significant as the first credible mechanism for pricing crawler access.

Pay-per-inference

The proposed alternative of paying publishers when their content is used in an answer rather than when it's fetched. Better aligned with value and much harder to verify.

HTTP 402

The long-dormant "Payment Required" status code, now being revived as the handshake for paid crawling and agent payments. Its sudden relevance is a decent shorthand for how fast this space is moving.

RSL (Really Simple Licensing)

A standard for publishers to attach machine-readable licensing terms to their content. It aims to make "you may use this, on these terms, for this price" expressible in a file rather than a contract.

Content licensing deal

A direct commercial agreement between an AI company and a publisher for content access. Reddit's SEC filings disclosed an aggregate contract value of $203 million for data-licensing arrangements signed in January 2024, which is the only reliably documented figure in a space full of leaked numbers.

TollBit

A third-party marketplace for metering and monetizing AI crawler access. Its published research on bot-to-human traffic ratios is among the more useful public data on the crawl economy.

Generative features opt-out

Google Search Console's control for excluding your site from certain AI experiences. Read the scope carefully, because it doesn't cover everything people assume it does.

10. The agentic web and AI commerce

AI agent

An AI system that takes action rather than only answering, handling multi-step tasks with limited supervision. When your buyer is an agent, your site is an API whether you designed it that way or not.

Agentic browsing

An agent navigating and interacting with websites on a user's behalf. It breaks every assumption behind conversion-rate optimization, because there's no human to persuade.

MCP (Model Context Protocol)

An open standard, originated by Anthropic, for connecting AI models to external tools and data sources. It has become the default way assistants reach live systems, which makes an MCP server a legitimate distribution channel for your product.

MCP server

A service exposing your tools or data to any AI client that speaks the protocol. Publishing one is increasingly how a brand becomes usable inside an assistant rather than merely mentioned by it.

WebMCP

A Community Group draft, explicitly not a W3C standard nor on the standards track, for letting websites expose their own capabilities to agents directly through the browser. It's the "make your site agent-callable" answer to agents having to scrape you instead.

NLWeb

A project for turning existing websites into conversational, agent-queryable endpoints, originated at Microsoft and now developed independently. Early, and aimed squarely at the same problem as WebMCP.

Agentic Commerce Protocol (ACP)

An OpenAI and Stripe standard for AI agents to complete purchases with merchants, and the protocol behind ChatGPT's Instant Checkout. Its April 2026 release expanded the spec to cover carts, product feeds and orders.

AP2 (Agent Payments Protocol)

Google's standard for authorizing agent-initiated payments using signed mandates, contributed to the FIDO Alliance in April 2026. Its Checkout and Payment Mandates are how an agent proves it was actually authorized to spend your money.

UCP (Universal Commerce Protocol)

A competing industry standard for agentic commerce, explicitly designed to be compatible with AP2. OpenAI is absent from its 16-member Shopping Tech Council, which tells you the protocol war isn't over.

A2A (Agent2Agent)

A standard for agents to communicate and delegate to one another, donated by Google to the Linux Foundation in June 2025. It matters once buying involves an agent talking to your agent rather than to your site.

Perplexity instant checkout

Perplexity's in-answer purchase feature, descended from Buy with Pro and relaunched with PayPal in November 2025, now free to US users. An early proof that the answer itself can be the checkout page.

Buy for Me

Amazon's feature for completing purchases from merchants outside Amazon. It puts Amazon's agent between you and your own storefront.

Agent readiness

How well your site works when the visitor is software: clean markup, stable selectors, accessible labels and no critical content locked behind interaction. It's mostly a rediscovery of accessibility work you were already supposed to do.

Agentic readability

Whether an agent can reliably extract meaning and act on your pages. The practical test is simple: turn off JavaScript, read the page, and see whether the important facts survive.

Skills

Packaged instructions that teach an assistant how to perform a specific task well. Brand-authored skills are an emerging way to control how an assistant represents your product.

11. Content strategy for AI Search

Answer-first formatting

Putting the direct answer at the top and the supporting detail below. Retrieval systems extract passages and models weight early text more heavily, so burying the conclusion is an unforced error. See Profound’s guide to optimizing content for AI Search.

Inverted pyramid

The journalistic structure of leading with the conclusion and descending into detail. It happens to be the format retrieval systems reward, which is a rare case of good writing and good optimization agreeing.

Atomic content

Writing each section so it's self-contained and independently meaningful. If a section only makes sense after reading the one above it, it'll be cited badly or not at all.

Chunk optimization

Structuring headings, paragraph breaks and lists so that the pieces a retrieval system extracts are the pieces you'd have chosen. It's the most direct lever content teams have on retrieval.

Information gain

The unique data, original research or firsthand insight a page adds beyond what's already published. Google holds a patent on measuring it, granted in 2022, though the patent describes a session-relative and per-user calculation rather than a static corpus-wide originality score.

Original research

Publishing data nobody else has. It's the most reliable way to become a source that other sources cite, which is how you get pulled into answers indirectly.

Statistics page

A page collecting cited figures on a topic. These earn citations far out of proportion to their effort, because answers constantly need a number to point at.

Comparison page

A page weighing your product against named alternatives. Comparison prompts are where shortlists form, and Profound's behavioral study with Kevin Indig and Clickstream Solutions found comparison grids held attention longest in 35.9% of shopping tasks, second only to the plain answer text.

Alternatives page

A page addressing "alternatives to X" queries, whether X is you or a competitor. Uncomfortable to publish and consistently effective, because it's exactly what the buyer asked.

Listicle leverage

Getting your brand included in the third-party "best tools" lists that AI answers lean on. It's often faster than trying to rank your own page for the same query.

Content freshness

How recently a page was published or meaningfully updated. Answer Engines favor current sources, and stale pages quietly teach models stale facts about your pricing and product.

Content refresh

Systematically updating existing pages rather than only publishing new ones. In AEO this is usually higher return than net-new content, because the citation is already there and only the facts are wrong.

Topical authority

Deep, interlinked coverage of a subject that convinces systems you're a definitive source on it. Breadth without depth reads as a content farm, and depth without breadth reads as a single lucky post.

Topic cluster

A group of related pages organized around a central pillar page. It remains the cleanest way to build topical authority that both search engines and retrieval systems can see.

Pillar page

The comprehensive hub page a topic cluster links back to. It's the page you want retrieved when the question is broad and the specifics when it's narrow.

Machine-readable content

Content structured so software can parse its meaning without guessing. Clear headings, real HTML, plain text and structured data, rather than text baked into images or assembled by scripts.

Crawlability for AI

Whether AI bots can actually reach and parse your pages. Many AI crawlers don't execute JavaScript, so content that only appears after scripts run may be invisible to them regardless of how good it's. Profound's crawlability audit tests for exactly this.

Server-side rendering

Delivering fully-formed HTML from the server rather than assembling it in the browser. It's the single most reliable technical fix for AI crawlability.

Search gap

The distance between the questions AI engines actually run on a topic and the content you've published to answer them. Closing it means writing for the queries the machine generates, not only the ones people type.

Digital PR for AI

Earning coverage on the third-party sites that AI answers already cite. It's traditional PR retargeted at a citation graph instead of an audience.

Parasite SEO

Publishing on someone else's high-authority domain to borrow its standing. It works, it's fragile, and Google's site reputation abuse policy covers it.

Borrowed authority

Getting your message onto a domain the model already trusts. The honest version of parasite SEO, and a reasonable strategy when your own domain is young.

12. Entities and structured data

Entity

A distinct thing a system recognizes and tracks, such as a company, product or person. Establishing your brand as a clearly defined entity is the foundation under both visibility and accuracy.

Entity establishment

Getting recognized as a specific, unambiguous thing rather than a string of characters. Consistent naming, consistent descriptions and corroboration across independent sources are what do it.

Entity disambiguation

Helping systems tell you apart from other things with similar names. The work is boring and the alternative is being merged with a company you've never heard of.

Entity confusion

When a system mixes your brand up with a similarly named company, an acquired product or an old version of your business. It's the most expensive failure mode in AI Search because it damages you while looking like visibility.

Entity salience

How central your brand is to a given topic in a system's understanding. It's what separates being known to exist from being the answer.

Knowledge graph

A structured map of entities and the relationships between them, consulted by search and AI systems. It's how a machine knows your CEO is a person, your product is software, and the two are related.

Knowledge panel

The information box search engines show for a recognized entity. Its presence is a decent signal that your entity is established, and its contents are a decent signal of what the machine believes.

Wikidata

The open structured database behind Wikipedia and a common input to knowledge graphs. Easier to get into than Wikipedia and frequently overlooked.

Wikipedia presence

Having an article about your company on Wikipedia. Its influence is routinely overstated: on June 2025 data Wikipedia was about 47.9% of ChatGPT's top-ten domains but only around 7.8% of all ChatGPT citations, and Profound's later measurement put it lower still.

Notability

Wikipedia's threshold for whether a subject warrants an article, judged by independent coverage. You can't buy it, you can't write your own article, and attempting either goes badly.

Structured data

Machine-readable labels in your page code that state explicitly what the content is. It removes guesswork for parsers, and the evidence on whether it lifts AI citations is genuinely mixed.

Schema.org

The shared vocabulary most structured data uses. Organization, Product, Article and Person are the types that carry the most weight for brand understanding.

JSON-LD

The preferred format for adding structured data, embedded as a script block. Google recommends it over the older Microdata and RDFa formats.

sameAs

A schema property linking your entity to its other authoritative profiles. It's the cheapest, most direct entity disambiguation signal available.

Organization schema

Structured data describing your company, including name, logo, description and social profiles. If you only implement one schema type for AEO reasons, implement this one.

Product schema

Structured data describing a product, including price, availability and reviews. Populating it with concrete, specific attributes rather than the bare minimum is the part that helps a machine describe your product correctly.

FAQPage schema

Structured data marking up question-and-answer content. Note that Google's FAQ rich results stopped appearing on 7 May 2026, with reporting support withdrawn over the following months, so implement it for machine readability rather than for a visual result that no longer exists.

Author markup

Structured data identifying who wrote a page and establishing their credentials. It supports the experience and expertise parts of E-E-A-T, which AI systems appear to weigh when choosing sources.

The schema evidence problem

Controlled testing on whether structured data lifts AI citations has produced conflicting results, including a large diff-in-differences study finding effects indistinguishable from zero on ChatGPT and AI Mode and slightly negative in AI Overviews. Ablation testing has consistently found that only rendered, visible text was reliably retrieved, so treat schema as clarification for machines rather than a growth lever.

13. Traffic, analytics and attribution

AI referral traffic

Visits arriving at your site from an AI platform. It's still a small share of total traffic and it's growing fast. Profound's guide to connecting AI referral data to Google Analytics covers how to actually see it.

Referrer string

The identifier a browser passes indicating where a visitor came from, such as chatgpt.com or perplexity.ai. It's the raw material of every AI traffic report and it's frequently missing.

utm_source=chatgpt.com

The UTM parameter ChatGPT appends to outbound links. Convenient, inconsistent across platforms, and not a substitute for proper channel grouping.

GA4 AI Assistants channel

A default channel group Google added to GA4 in May 2026, triggered by a medium of "ai-assistant." Google's documentation names ChatGPT, Gemini, Claude, DeepSeek, Copilot and Grok as examples but hasn't published the full referrer list, so custom channel groups are still worth keeping.

Dark traffic

Visits with no usable referrer that land in direct traffic. Native desktop and mobile assistant apps are a major contributor, which means your real AI-driven traffic is higher than your dashboard shows. Profound's Agent Analytics works from server logs partly to get around this.

Attribution gap

The structural difference between the influence AI Search has and the influence your analytics can prove. It's why AEO reporting leans on visibility and citation metrics rather than sessions.

Generative AI performance report

Google Search Console's report covering AI Overviews and AI Mode, launched in June 2026. It aggregates the two surfaces, offers only Pages, Countries, Dates and Devices as dimensions, and reports impressions only. Clicks from AI Overviews and AI Mode are counted inside the main Performance report's web results, they're simply never broken out.

Branded search lift

An increase in people searching your brand name directly. When an answer names you without linking you, branded search is often where the effect actually shows up, though Profound has tracked ChatGPT sharply increasing the share of responses containing a link.

Zero-click rate

The share of searches ending without a click to any site. Profound's task-level study of AI assistants found 92.8% of shopping tasks ended with no meaningful click at all.

Conversion rate of AI traffic

How AI-referred visitors convert relative to other channels. The picture has genuinely reversed: Adobe measured AI-referred retail traffic converting 43% worse than non-AI traffic in July 2024, narrowing to 9% worse by February 2025, then flipping to 42% better in March 2026 and 54% better by May 2026.

Bot-to-human ratio

How many AI crawler visits your site receives per human visitor. TollBit measured bot visits per human visit across publishers on its own network moving from roughly 1:200 to 1:31 over 2025, and Profound's Agent Analytics measures the same ratio for your own domain.

Impression-only reporting

Metrics that tell you a page was shown but not whether anyone acted on it. It's the dominant reporting mode across AI surfaces, and getting comfortable with it's part of the job now.

14. SEO carryover

SEO (Search Engine Optimization)

The practice of earning visibility in traditional search results. It's not obsolete, since organic rank remains the strongest single predictor of getting cited in AI answers.

The rank-citation relationship

The consistent finding that pages ranking well in traditional search are far more likely to be cited in AI answers. Ahrefs found 76.10% of AI Overview citations came from top-ten results in July 2025, then reran the study in March 2026 and got 37.9%, a drop it attributes to Gemini 3-era query fan-out alongside its own improved parsing. Read them as before-and-after rather than as competing figures, and note Ahrefs separately characterizes the rank-to-citation relationship as positive but only moderate.

E-E-A-T

Experience, Expertise, Authoritativeness and Trustworthiness, Google's framework for describing credibility. Google says it's not a specific ranking factor while also deliberately weighting signals aligned with it, more so on health and finance topics, and it works well as a checklist for whether content deserves to be cited.

Featured snippet

The extracted answer box in traditional results. It was the training ground for answer-first writing, and the skills transfer directly.

Backlinks

Links from other sites to yours. Their AEO relevance is indirect, since they drive the organic rank that predicts citation, rather than being read as votes by the model.

Domain authority

A third-party estimate of a domain's overall standing. Useful for triage, and not a metric any AI system actually uses.

Canonical URL

The version of a page you designate as authoritative among duplicates. It determines which URL gets credited when you're cited, which is worth getting right before you start reporting on citations.

Core update

A significant change to Google's ranking systems. Because organic rank feeds AI citation, core updates now move your AI visibility too.

Site reputation abuse

Google's policy against publishing unrelated third-party content on an established domain to exploit its standing. Announced in March 2024 and tightened that November to remove the first-party-oversight defense, it applies regardless of scale.

Scaled content abuse

Google's policy against mass-producing low-value pages, including with AI. It's the guardrail that makes "just generate a thousand pages" a bad plan.

Keyword stuffing

Repeating terms unnaturally to signal relevance. The original GEO research tested it and found it scored below baseline on position-adjusted word count, classifying it as non-performing, so it fails in AI Search for much the same reason it failed in search.

15. Risks and failure modes

Knowledge decay

A model repeating facts about you that were true when it was trained and aren't true now: old pricing, retired products, abandoned positioning. Fresh crawlable content is the only real antidote, and it's slower than anyone wants. Profound built FactCheck to catch it, and found 47% of response content about a brand was unsolicited across 50,000 responses.

Brand fidelity

How accurately AI systems represent your brand beyond whether they mention it, covering category placement, current products and real differentiators. It names the gap between being visible and being described correctly, which is the gap most visibility dashboards are structurally unable to show you.

Content fidelity

How faithfully an answer reproduces the substance of your content when it draws on you, a term Adam Malamis of Gander claims coinage of. Low content fidelity means you get cited while your figures and claims arrive mangled, which no visibility metric will flag.

Prompt injection

Hiding instructions in content so a model reading it behaves differently. It's a security problem for anyone deploying agents and a reputational problem for anyone whose content can be poisoned.

Indirect prompt injection

Prompt injection delivered through a page the model retrieves rather than through the user's input. It's how a third-party site can influence what an assistant says about you.

Retrieval poisoning

Deliberately planting content designed to be retrieved and to distort answers. Academic work has demonstrated it repeatedly, and it's the adversarial edge of AEO.

Recommendation poisoning

Manipulating third-party listicles and comparison content to skew AI recommendations. Effective, cheap, and the reason competitor comparison pages deserve monitoring.

Strategic text sequence

A crafted string inserted into content to manipulate how a model ranks or describes it. Documented in research as a real attack, and firmly on the wrong side of the line.

GEO spam

Industrialized manipulation of AI answers through mass-generated or adversarial content. It's early enough that it often works, which guarantees the platforms will come for it.

Model collapse

The degradation that occurs when models are trained on increasing volumes of AI-generated text. A long-term risk to the whole ecosystem rather than something you manage quarterly.

AI slop

Low-quality, mass-produced AI content. Merriam-Webster made "slop" its word of the year for 2025, which is a reasonable indicator of how the public feels about the current supply.

Slopsquatting

Registering names, packages or domains that models tend to hallucinate. A genuine supply-chain risk and a strange new form of typosquatting.

AI washing

Overstating the role AI plays in a product or a marketing program. It attracts regulatory attention, and it makes credible AI claims harder for everyone.

Answer bubbles

The narrowing effect of everyone receiving similar synthesized answers to the same questions. It rewards incumbents and makes category entry harder than it was in traditional search.

Incumbent advantage

The tendency of AI answers to reinforce brands that are already well covered. It's the structural reason AEO for a challenger brand is an offsite problem before it's an onsite one.

Hallucinated claims about your brand

Confident, specific, invented statements about your pricing, features, customers or leadership. Detecting them requires sampled monitoring, because they don't appear in every run, which is what FactCheck does.

Circular sourcing

When AI answers cite content that was itself derived from AI answers. It's how one confident error becomes the consensus, and it's already visible in this industry's own statistics.

A–Z Index

All 309 terms, alphabetically.

A · B · C · D · E · F · G · H · I · J · K · L · M · N · O · P · Q · R · S · T · U · V · W · Z

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

Z

About this glossary

Built and maintained by Profound. Profound measures how brands appear across AI Answer Engines — visibility, citations, sentiment, accuracy and crawler activity — so marketing teams can work from data instead of spot checks.

Go deeper on any of this:

On sourcing. Every statistic here's attributed, and where a widely repeated claim doesn't survive contact with a primary source, we've left the correction in rather than quietly dropping it. Several figures in circulation across this industry trace back to nothing.

Corrections welcome. This field moves fast enough that some of the above will be wrong within a quarter, and we'd rather be told than stay wrong.