Every week I sit in a meeting where two people use the same word and mean different things. One person says "we got cited in ChatGPT" and means the assistant named the brand. Another hears it and assumes a link, a visit, maybe a lead. Those are four different outcomes, and mixing them up makes reports look better or worse than they really are.
I wrote this glossary to fix that. It covers the standard vocabulary of AI search visibility, then goes further into the terms I track every week but rarely see defined anywhere: crawl-to-referral ratios, crawler-purpose controls, answer volatility, model version drift, sponsored answers, and agentic commerce protocols.
A quick disclosure: I'm Growth Lead at XLR8 AI, a generative engine optimization (GEO) company that helps brands get named and cited in AI search. I've kept this glossary vendor-neutral, and I flag where a definition is my own working convention rather than an industry standard.
AI search is any search experience where a model writes the answer instead of returning a list of links. ChatGPT search, Perplexity, Claude, Gemini, Copilot, Google AI Overviews, and AI Mode all fit this definition.
An answer engine responds with a written answer, usually with sources attached. I use the term when I want to stress that the user may never see a results page.
I'll be honest: nobody has standardized these. Google Search Central says optimizing for its generative features is still SEO. I pick one label per team and define it once, rather than running four strategies with four names.
AI Overviews are Google's AI-written summaries above organic results. AI Mode is Google's separate conversational search experience for complex, multi-step questions. Google says the two can use different models and techniques, so I track them separately. Appearing in one tells me nothing certain about the other.
A zero-click search ends without the user visiting any website, because the answer appears on the page. AI answers push this further. I now treat brand presence inside the answer as a result in its own right, not just a path to a click.
Conversational search lets people ask natural-language questions and follow-ups while the engine keeps context. In a multi-turn conversation, my brand needs to survive the follow-up ("which is cheapest?", "which works for a team of five?"), not just the opening question.
An LLM is a model trained on large volumes of text to understand and generate language. It powers every assistant on this list.
An AI agent plans steps, uses tools, browses, and completes tasks with little human input. Agentic search is the research version: the agent searches, reads, and searches again until it can answer. Agents compare options on a buyer's behalf, so I make sure the facts they need (pricing, specs, policies) are easy to find and consistent.
RAG retrieves relevant external content, adds it to the model's input, and writes the answer from that material. For visibility work, I separate two steps: getting my page retrieved, and getting my brand named or cited in the final answer.
Grounding ties an answer to supporting sources. It reduces guessing, but a grounded answer can still be wrong, so I never treat it as a guarantee of accuracy.
Query fan-out splits one question into several related searches run in parallel. Google says AI Overviews and AI Mode may use it. I plan content around the sub-questions a buyer's prompt implies, not just the literal prompt.
A passage, or chunk, is a section of a page that a retrieval system can match on its own. I write sections that make sense out of context because that helps readers, but I don't follow a magic paragraph length. Google says special chunking isn't required.
Embeddings turn text into numbers that capture meaning. Vector search uses them to find conceptually similar content without exact keyword matches. Hybrid search combines vector and keyword matching. Together, they explain why clear topical coverage often beats keyword repetition.
An entity is a distinct, recognizable thing: a person, a company, a product. A knowledge graph stores entities and how they relate. I keep a brand's core facts (name, category, founders, locations, products) identical everywhere so models resolve it as one entity, not three.
A token is the unit a model processes, which can be a word, a part of a word, or a character. The context window is how much material a model can handle in one request or conversation. The knowledge cutoff is the date its training data ends. Retrieval can supply newer information, so a cutoff doesn't decide how fresh every answer is.
A hallucination is a confident, false statement from an AI model, such as an invented price or a feature a product never had. I monitor for these on branded prompts and trace them back to the source that seeded them.
Non-determinism means the same prompt can produce different answers on different runs. It's the reason I never trust a single screenshot. More on this under answer volatility below.
The biggest mistake I see is treating a mention, a citation, and a visit as the same thing. Kevin Indig's July 2026 analysis of Semrush data, covering 1,094 US categories and more than 600,000 ChatGPT citations, found that the most-cited domain was also the most-mentioned brand in only 20.8% of cases. Being cited and being named are different outcomes.
| Term | My working definition | What I watch for |
|---|---|---|
| Brand mention | The answer names the brand, with or without a link. | A neutral listing isn't a recommendation. |
| Consulted source | A URL the system read while building the answer. | OpenAI's web-search documentation separates consulted URLs from the smaller set of citations it displays. |
| Citation | The answer shows a link to a source. | The cited domain may not be the brand that gets named. |
| Recommendation | The answer explicitly suggests choosing the brand. | I require an endorsement, not just an appearance. |
| AI referral traffic | A visit to the site from a link in an AI answer. | GA4 added an AI Assistant channel on May 13, 2026. It counts visits, not answer exposure. |
| Mention rate | Answers naming the brand ÷ all sampled answers. | I always state the prompts, models, dates, and runs behind the number. |
| AI visibility score | A vendor's composite summary of visibility or readiness. | I check the inputs. A score out of 100 isn't a probability of appearing. |
| Share of model / AI share of voice | The brand's mentions ÷ all tracked brands' mentions in a fixed sample. | One answer can name several brands, so this differs from mention rate. |
| AI sentiment | Whether the answer describes the brand positively, neutrally, or negatively. | Negative doesn't always mean inaccurate. I check the claim first. |
| Branded vs. unbranded prompts | Branded prompts name the company. Unbranded prompts describe a need. | Unbranded prompts show whether AI recommends me when buyers don't know my name yet. |
| Prompt tracking | Running a fixed prompt set on a schedule to track mentions and sources over time. | The prompt set must reflect real buyer questions, or the trend means little. |
If you want to see these measures for your own brand, the free AI visibility report from XLR8 AI covers brand descriptions, queries, share of voice, and competitors.
E-E-A-T stands for experience, expertise, authoritativeness, and trustworthiness. Google says trust matters most and that E-E-A-T isn't a single ranking factor. I use it as a review lens, not a score.
Information gain is the new information a page adds beyond what already exists on the topic. Original data, firsthand testing, and specific examples give a model a reason to use my page over the ten that say the same thing.
BLUF means "bottom line up front." I put the direct answer in the first sentence or two of a section, then add detail. Readers skim, and retrieval systems often pull from the top of a passage.
Proof points are specific, verifiable facts: statistics, named results, certifications, dates. The original GEO research paper (2023) found that adding statistics, quotations, and citations improved one visibility measure by roughly 30–40%. That's a benchmark result, not a guarantee on today's platforms, but it matches what I see.
Freshness signals show a page is current: visible update dates, recent data, current product details. I refresh pages when facts change, not just to change a date.
Topical authority is how thoroughly a site covers a subject across all its pages. One great article rarely beats a site that answers every related question well.
Structured data is machine-readable markup, usually schema.org in JSON-LD format, that labels page content. I'll be direct about the evidence: Ahrefs' May 2026 study compared 1,885 pages that added JSON-LD with about 4,000 control pages and found no meaningful citation lift. That study looked at pages already earning citations over 30 days, so it doesn't prove schema never helps. I still use schema for products and organizations, but I don't sell it as a citation lever.
An unlinked brand mention names my brand without linking to me. Earned media is unpaid coverage: press, reviews, forum threads, comparison articles. AI engines read both. I've seen a single well-placed Reddit thread or industry roundup change how an assistant describes a brand.
A named author is a real, identifiable person credited on the page. Google recommends clear authorship. I add it because it builds trust with readers, not because a byline guarantees citations.
AI companies now run separate bots for separate jobs. Blocking the wrong one can quietly remove you from AI search.
| Term | What it does | What I watch for |
|---|---|---|
Training crawlerGPTBotClaudeBot |
Collects content that may train future models. | Blocking it doesn't require blocking the same company's search crawler. |
Search crawlerOAI-SearchBotClaude-SearchBotPerplexityBot |
Indexes pages for the assistant's live search results. | Block this and you can disappear from that assistant's cited answers. |
User-triggered fetcherChatGPT-UserClaude-UserPerplexity-User |
Fetches a page when a user's request needs it. | OpenAI and Perplexity say robots.txt may not apply to these fetches. |
| Google-Extended | A robots.txt token controlling some Gemini training and grounding uses. | It doesn't control inclusion in Google Search or its AI features. |
| robots.txt | Tells crawlers which parts of a site they may visit. | It controls crawl access, not snippets or indexing. |
| WAF rules | Firewall rules that can block bots before robots.txt matters. | Perplexity advises checking WAF rules and its published IP ranges. |
| Server-side rendering (SSR) | Serves finished content in the HTML instead of building it with JavaScript. | Vercel reports most AI crawlers don't execute JavaScript. |
| llms.txt | A proposed markdown file summarizing a site for language models. | Experimental. Google says Search ignores it. |
| IndexNow | Notifies participating search engines that a URL changed. | Submission doesn't guarantee indexing or citations. |
Most glossaries stop at the list above. These are the terms that actually explain the surprises in my reports. A few are my own working definitions, and I've marked them that way.
The crawl-to-referral ratio is how many pages an AI company's bots fetch for every visitor it sends back. Cloudflare's July 2025 data put Anthropic at about 38,000:1, OpenAI at about 1,100:1, Perplexity at about 195:1, and Google at about 5:1. Cloudflare notes these are upper bounds, because many apps don't pass a referrer. I use the ratio to decide which bots earn access. It's also a reminder that AI visibility mostly pays off in brand presence, not clicks.
Crawler-purpose controls let site owners allow or block bots by what they do (search, agent, or training) instead of by name. Since July 1, 2026, every Cloudflare plan, including Free, has separate switches for all three. From September 15, 2026, newly onboarded domains block training and agent crawlers on ad-supported pages by default, while search stays allowed. Cloudflare also changed how it treats mixed-use bots like Googlebot, Bingbot, and Applebot, but a blunt "Block" setting can still stop them. I audit these settings on every new client site, because one toggle can undo months of work.
Answer volatility is how much answers to the same prompt change from run to run. SparkToro's January 2026 study collected 2,961 runs of 12 prompts from 600 volunteers and found less than a 1-in-100 chance that ChatGPT or Google's AI returned the same brand list twice. I report visibility as a percentage across many runs, never as a "rank."
Prompt sampling means running each prompt enough times to estimate a mention rate. Statistical confidence describes how certain that estimate is. No single run count works for every case, and small samples need extra care. My rule: if a "10-point gain" falls inside the noise of the last three weeks, I don't celebrate it.
Model version drift is a visibility change that lines up with a model or system update, not with anything I changed. I annotate dashboards with model release dates so my team doesn't credit or blame our content for a provider's update.
Persona and memory effects are the ways personalization changes answers for different users. OpenAI documents that ChatGPT memory can refine searches using location or preferences. I log account state and test with persona-framed prompts ("I'm a CFO at a 200-person company…") so I know which audience a result represents.
The citation core is the small set of domains an assistant cites again and again for a category. The citation gap is the list of sources where competitors appear and my brand doesn't. Mapping both tells me where earned coverage will matter most. Tools like XLR8 AI Insights surface these cited sources by prompt and model.
Answer position is where the brand appears in the answer. First-mention rate is how often it appears first, divided by all sampled answers. Visibility-to-recommendation rate is how often a mention turns into an explicit recommendation. Being listed fifth with a caveat is not the same win as being recommended first.
Narrative flattening happens when AI describes a distinctive brand in generic category language that could fit any competitor. A perception gap is the difference between verified brand facts and the assistant's description. I fix flattening with sharp, repeatable proof points. I fix perception gaps by tracing the claim to its source and correcting it there.
Retrievability is how easily a system can find, access, extract, and reuse accurate facts about a brand. It combines crawl access, rendering, internal linking, and snippet eligibility. When a brand is missing from answers, I check retrievability before I assume it needs more mentions.
Snippet controls (nosnippet, max-snippet, data-nosnippet, and noindex) affect what Google can show from a page in its AI features. They work separately from robots.txt. Google's Search Console has reported Generative AI performance for every site since August 31, 2026, so I now check AI impressions next to these settings.
Sponsored answers are paid placements inside AI assistant responses. OpenAI started testing ads in ChatGPT in the US on February 9, 2026, for Free and Go users, and has since expanded to more markets. OpenAI says ads don't influence answers and appear labeled and separate. I now report organic visibility and paid placement as two different line items.
ACP is an open standard OpenAI and Stripe launched on September 29, 2025, alongside Instant Checkout in ChatGPT. UCP is a Google-backed open standard for agentic commerce, with documentation published in January 2026. They're different standards. For e-commerce brands, the takeaway is simple: a clean, complete product feed is now a visibility asset, not just checkout plumbing.
MCP is an open standard that connects AI apps to outside tools and data. It's an integration standard, not a ranking factor. I mention it because buyers increasingly ask assistants to act on live data, and brands with an MCP server or clean API become easier for agents to use.
This is roughly the loop we run at XLR8 AI. Insights handles the tracking and source discovery, and Actions maps negative or inaccurate mentions back to the sources that seeded them, so the team knows what to fix.
You don't need to memorize every term here. You need your team to mean the same thing when someone says "visibility," "citation," or "we're winning in ChatGPT." Start with the measurement table, audit your crawler settings, and stop reporting single screenshots.
If you want a starting point, try the free AI visibility report. It shows how AI engines describe your brand, which queries you appear in, and how your share of voice compares with competitors. For more on the practice, the XLR8 AI guides go deeper. And if you'd like to talk through any of these terms, reach me at anu@tryxlr8.ai. I'm always happy to chat about GEO.
GEO and AEO both describe work aimed at getting a brand named and cited in AI-generated answers. The labels overlap, and Google considers optimizing for its AI features part of SEO. I treat them as one practice with an extra layer of measurement: mentions, citations, and share of model.
A mention names your brand in the answer. A citation links to a source page, which may or may not be yours. Kevin Indig's 2026 analysis found the most-cited domain was the most-mentioned brand only 20.8% of the time, so I track them separately.
There's no universal number. It depends on your baseline mention rate, how much answers vary, and the size of the change you want to detect. I compare repeated samples over several weeks and treat small swings as noise until they persist.
Usually, yes. OpenAI, Anthropic, and Perplexity run separate training, search, and user-triggered bots, so you can block training and allow search. Check firewall and CDN settings too. Cloudflare's 2026 defaults block training and agent crawlers on some new domains.
OpenAI says ads don't influence ChatGPT's answers and are labeled separately. I treat paid placement and organic visibility as two separate channels, with separate budgets and separate reporting.
