AI search glossary — GEO, Reddit, and the words that actually matter

Reddex Influence Plan items: a next step with a channel and a reason, not a generic “engage on social.”

Adaptive retrieval is when the system decides on the fly whether to look something up, and how hard, instead of always searching or never searching.

AXO is making a site usable for AI agents that click, fill forms, and complete tasks — not only for human readers and classic crawlers.

AP2 is Google’s proposed way for agents to pay with signed mandates — what the agent is allowed to buy, and on whose behalf.

A2A is an open idea for agents from different vendors to discover each other and hand off tasks.

Agentic commerce is shopping done by software: find, compare, buy, with a human in the loop somewhere — or not.

ACP is the OpenAI/Stripe experiment for delegated checkout inside chat surfaces.

Agentic search is when software plans, browses, clicks, and compares instead of answering in one shot.

Preparing for agents that plan and browse, not only for one-shot answers. Overlaps GEO; the test is whether the agent can finish the job on your site.

Agentic SEO usually means using agents to run SEO/GEO workflows — research, drafts, monitoring — not optimizing for agents.

Agentic workflows are multi-step loops: plan, tool, observe, repeat. Search is one tool among many.

AGENTS.md is a repo or site file that tells coding or browsing agents how to work with a project: commands, conventions, pitfalls.

Libraries for building agents that plan and call tools: LangGraph, Crew, custom stacks. They shape how your APIs get called, not how Google ranks you.

Agents are models wrapped in a loop that can use tools and pursue a goal. Chatbots answer. Agents try to finish work.

Alignment is the attempt to make models do what humans intend. It affects which sources a model treats as safe to quote.

Volatility is how much the answer moves when you ask the same prompt twice. GEO reports that ignore it overfit noise.

An AI API exposes a model to software. Answers via API can differ from the consumer chat UI.

Benchmarks are tests for model skill. They predict coding and math better than they predict whether you get cited for CRM software.

An AI brand mention is your name (or product) appearing in a generated answer. It may or may not include a link or a kind adjective.

AI brand safety is watching for answers that defame, hallucinate scandals, or mix you up with a worse company of a similar name.

Tactics to earn footnotes: original facts, clear entities, crawl access, and being the page people already cite.

An audit of which domains and URLs models actually footnote for your prompts.

Detectors try to guess if text was written by a model. They are shaky. Using them as a GEO quality bar is a mistake.

Using models to draft text, images, or video. Fine for speed. Fatal if you skip facts and Reddit-aware review.

How a system picks which passage to quote. Authority, freshness, uniqueness, and whether the chunk answers the sub-query.

A plan for what to publish that assumes models and humans both read it. Starts from prompts and Reddit, not from a keyword spreadsheet alone.

Server logs that show GPTBot, ClaudeBot, PerplexityBot, and friends hitting your URLs.

Checking that a request claiming to be GPTBot really is, via published IP ranges or signed requests.

Visits influenced by AI answers that show up as direct, branded search, or unassigned. The chat never sent a referrer.

Training extra weights on your data so a model behaves in a specialized way. This is not how ChatGPT learns your brand from the public web.

Grounding is tying the answer to retrieved evidence so it hallucinates less. Citations are the visible part.

A hallucination is a confident falsehood: fake URLs, fake papers, fake quotes. GEO programs must log them, not only “wins.”

How AI systems store pages for later retrieval. Not identical to Google’s index. Different bots, different freshness, different junk filters.

Inference is running the trained model to produce an answer. Every ChatGPT reply is inference. It costs money, which is why vendors cache and skip retrieval.

Counts of how often an AI Overview appeared for queries you care about. Not clicks. Not citations of you.

Laws like the EU AI Act that push transparency, copyright, and training disclosures. They change crawler politics more than your H1.

Editing content and sources so the generated answer is accurate and includes you. A fancy alias for GEO plus brand safety.

Safety systems block disallowed content and sometimes refuse whole categories. They can suppress your brand on sensitive prompts.

AI search is discovery that returns a synthesized answer, not just a list of links. ChatGPT, Perplexity, AI Mode, and Overviews all sit here.

Sponsored slots inside AI answers. Early, inconsistent, easy to overpay for because measurement is worse than Search ads.

The discipline of measuring citations, mentions, sentiment, and AI-referred traffic across engines.

Another name for GEO/AI SEO. If a contract uses this phrase, make them define the KPIs.

A web index built or rented for answer retrieval. OpenAI, Google, Perplexity, and others do not share one pile.

The conclusions you draw from prompt tracking: who owns which question, which URLs are the evidence, where sentiment is sour.

Matching pages to conversational intent: compare, troubleshoot, “is it worth it,” not just head keywords.

A rollup of how you do across engines: inclusion, citations, sentiment, traffic assists.

A tempting list (authority, freshness, unique data, entity clarity). None are published the way Google’s old leaks get treated.

Synonym for [[ai-visibility|AI visibility]]: presence in generated answers across the engines you care about.

Bots that research, draft, and monitor SEO/GEO tasks. Useful for diffs and drafts. Dangerous on publish.

Marketing-speak for [[share-of-model|Share of Model]]. If you say SOV in an AI meeting, define the denominator.

Product discovery and sometimes checkout inside assistants. Needs feeds, truthful prices, and reviews the model already trusts.

Scaled, low-value machine text aimed at search or AI retrieval. It pollutes indexes and makes your original work harder to find.

The corpus used to bake parametric memory. Your brand in that pile is why ChatGPT “just knows” you — until it doesn’t.

AI visibility is how often, how fairly, and how prominently models mention your brand when people ask category questions.

Software that tracks how brands appear in AI answers. Rank-tracker energy, new surfaces. Reddex is one, with Reddit attached on purpose.

AI Visibility Score is Reddex’s rollup of how often you appear in tracked AI answers across engines, with sentiment and trend.

Bots from model labs and answer engines that fetch pages for training, retrieval, or user-initiated reading.

Planning content so citation is a design constraint from the brief, not a sprinkle of schema at the end.

Products that search with language models: ChatGPT, Perplexity, Gemini, Copilot, Grok, and a long tail of wrappers.

AI-referred traffic is visits that come from chat and answer products. Some show as referrals. A lot show up as direct or brand search.

Changes to ranking or retrieval that move Google and, with a lag, what models cite.

Google’s old fast-pages format. It is not an AI ranking booster. Core Web Vitals and real HTML matter more now.

Distinctive phrases you own — named metrics, frameworks, product code names — that travel with you when a model summarizes.

An answer engine replies with a synthesized answer. Perplexity is the clean example. Google is a hybrid. ChatGPT is a chat app that learned search.

AEO is structuring content so an answer engine can lift a clean response — featured snippets, voice, and now chat surfaces that skip the blue-link list.

Where your brand sits inside an AI answer: first named, buried in a list, or used as a cautionary tale.

Pages that lead with a direct answer, then proof. Built for snippets, Overviews, and chat extraction.

The lab behind Claude, constitutional AI, and MCP. Their crawler and products are a GEO surface, not just a competitor to OpenAI.

On-device and Private Cloud Compute assistants in Apple’s ecosystem. Discovery here is app- and OS-shaped, not a public SERP.

Whether a named person is treated as a credible source. Models pick this up from bylines, profiles, and who else cites them.

Links from other sites. Still useful for discovery and Google. For AI citations, entity and third-party consensus often matter more than raw link count.

Google’s 2018-ish language model for search understanding. Historical layer under modern NLP, not something you “optimize for” in 2026.

Microsoft’s web index. It has fed various ChatGPT search setups and still matters for Copilot.

A classic lexical ranking function. Still used as a first-stage retriever before fancy embeddings.

Share of sessions that end after one page. A sloppy quality proxy. AI search makes it worse as a KPI because many visits are answer-checks.

How widely and consistently you appear across the web before any one citation happens: profiles, news, docs, Reddit, reviews.

Brand inclusion rate is the percent of tracked prompts whose answers include your brand at least once.

Listening for mentions across the web and socials. AI monitoring is the new room in the same house. Reddit is often the loudest room.

The Google results for your name. It trains humans and, indirectly, models on who you are.

The preferred URL when duplicates exist. AI crawlers also waste budget on copies and may cite the ugly one.

Asking a model to reason step by step. Built into many reasoning models now. Users don’t see all of it.

ChatGPT is OpenAI’s consumer assistant. For GEO, it is both a memory of the web and, with Search, a live retrieval product.

OpenAI’s browser-like agent surface where ChatGPT can read pages and do multi-step web tasks.

GEO focused on ChatGPT: memory plus Search. Don’t pretend other engines don’t exist, but this is the screenshot everyone sends.

ChatGPT Search is OpenAI’s live web retrieval inside ChatGPT. It is a different beast from “what GPT remembers about you.”

The formats models like to footnote: docs, listicles, news, product pages, forum threads, transcripts.

Training bots build weights. Search/citation bots fetch for an answer now. User-triggered bots fetch because a human asked.

Citation decay is when AI answers stop using your page: fresher sources, a retrieval change, or your facts went stale.

Whether answers cite a mix of independent sources or the same three domains on loop.

A prompt cluster where rivals or Reddit get the footnotes and you don’t, even though you sell into that question.

A model’s estimated chance a URL gets cited for a prompt. Vendor scores are guesses. Your observed rate is the truth.

Citation rate is the share of tracked prompts where an engine attaches a link to your domain.

Citation share is the percent of relevant AI answers that point to your domain as a source, not just say your name.

How many sources an engine attaches, on average. Perplexity is generous. ChatGPT is stingy. Don’t compare rates across engines without this context.

How often a specific URL — not just the domain — appears as a citation.

Schema for fact-checks. Helps some systems find an adjudicated claim. Niche, powerful on rumors.

Claude is Anthropic’s assistant: long context, cautious tone, MCP tooling, and a crawler (ClaudeBot) you will see in logs.

Clicks divided by impressions. Overviews and chat crush CTR even when you “rank.”

The parent post, subreddit norms, and sarcasm around a mention. Sentiment without context is how you page the CEO at 2am for a joke.

For GEO: who gets named, who gets cited, which Reddit threads they own, not only who ranks in Google.

Models that operate GUIs: click, type, screenshot. Your CSS is now an API, whether you like it or not.

A Content Agent is Reddex’s GEO drafting loop: it looks at citation gaps and Reddit threads, then writes a page built to get cited — schema, sources, and the actual question people ask.

Breaking ideas into reusable factual units that can be retrieved alone: a definition, a number, a limit, a date.

Trust at the piece level, not the domain. A great paper on a small site can beat a weak post on a giant one in AI retrieval.

Chunking is writing sections that still make sense if a retriever lifts 100–300 words and ignores the rest of the page.

Planning pages as a cluster around a pillar. Helps Google and gives retrievers more than one door.

The published set of related pages that together say “we cover this subject.” Models notice coverage more than a lone post.

Rankings and usefulness fading as facts age and competitors publish. Citation decay is the AI twin.

How recently a page was substantively updated. Retrievers like dates. Fake dates are a fast way to burn trust.

Finding questions they rank or get cited for that you don’t. For GEO, the gap list comes from prompt tracking and citation URLs, not only Ahrefs.

Changing the page per visitor. Fine for apps. Risky if the crawler sees a different, empty, or bait page.

The messy set of clues engines use to decide if a page is worth ranking or citing: originality, evidence, UX, reputation.

Proposed robots.txt-style directives for how crawled content may be used: search vs AI input vs training.

Choosing what goes into a model’s context window: docs, tools, memory. Product skill, not a meta description trick.

How much text a model can consider at once. Bigger windows mean longer threads and more pasted docs — including your pricing PDF.

Making content work in multi-turn chat: follow-ups, constraints, “what about X.” Same as GEO with more pronouns.

Search as a dialogue with follow-ups, not a one-box query. AI Mode and chat apps live here.

Improving the site so visitors convert. AI-referred visitors may convert differently — sometimes higher intent, fewer pages.

LCP, INP, CLS — Google’s speed and stability metrics. Slow pages also annoy crawlers and agents.

How much a bot is willing to fetch. AI bots can be greedier or pickier than Googlebot. Faceted URLs still waste everyone’s time.

Discover, fetch, store. Google does it. AI labs do it with other bots and rules. You need both if you want both channels.

The fact that engines disagree. Few domains win everywhere. Reporting one average hides the work.

How personal data is used in training, logs, and plugins. Trust issue for buyers; compliance issue for you.

A mode where the assistant runs a long, multi-query investigation and returns a sourced brief.

Intensive retrieval: many sub-queries and reranks. Similar idea to Deep Research, different product names.

A Chinese lab known for strong open-ish models at lower cost. A real engine in some markets and in Reddex’s mix.

Schema.org type for glossary definitions. This Reddex glossary uses it so agents can lift a term without the chrome.

A vendor-ish name for making your entity consistent everywhere so AI systems recommend the right org.

Moz’s 1–100 score. A correlation toy. Not an AI ranking factor. Useful only as a rough site-size intuition.

How long someone stays after a click. A folk ranking signal. For AI, the analog is “did the chunk actually answer.”

Experience, Expertise, Authoritativeness, Trustworthiness — Google’s quality lens. AI search borrows the vibe even when it does not use the rubric.

Vectors that represent meaning. They power semantic search and RAG. Your page becomes a point in space.

Whether all the public facts about you agree. Incoherent entities get merged, split, or invented.

How much independent sources repeat the same facts about you. Models trust a chorus.

The one URL that should define the entity. Usually the official site or a section of it, sometimes a Wikipedia page you don’t control.

How strongly a system associates you with a topic. High salience on “incident response tools,” low on “email marketing,” etc.

Entity SEO is making sure machines can tell who you are: one name, one home URL, same facts everywhere.

FAQPage JSON-LD. Useful when the FAQs are visible on the page and actually asked. Spam when they’re invisible keyword blobs.

Google’s position-zero extract. Still a training gym for answer-ready writing, and still a feed into some Overviews.

Teaching a model a task with a handful of examples in the prompt. A prompting technique, not a page type.

The big base models (GPT, Claude, Gemini, Llama, etc.) that products wrap.

The model invokes tools and APIs instead of only speaking. Agents are this with a loop.

Search that writes an answer. Parent category for GEO. See [[ai-search|AI search]].

GEO is the work of getting your brand named, quoted, or linked when an AI engine answers a question — ChatGPT, Perplexity, Claude, Gemini, Copilot, or Google AI Overviews.

Yet another alias for GEO. Confirm KPIs when it appears in a contract.

A structured look at crawl access, entity mess, prompt inclusion, citation sources, and Reddit risk.

The scoreboard: inclusion, Share of Model, citation rate, sentiment, source mix, AI-referred assists, Reddit visibility.

A ghost citation is when an AI answer names your brand but does not link you. You got the shout, not the click.

AI Mode is Google’s conversational search UI: Gemini plus fan-out, more chat than a single Overview box.

AI Overviews are Google’s generated summaries on the results page. They sit above the links and eat clicks.

Gemini is Google’s model family. It powers AI Overviews, AI Mode, and the Gemini app — which means Google SEO and GEO collide here.

Google’s free crawl, index, and query tool. Necessary, not sufficient, for GEO.

OpenAI’s model family name, now casual shorthand for “a chatbot.” The architecture is transformers; the product is ChatGPT.

xAI’s assistant, tied to X. Tone is different; sources can be more social. Track if your buyers live on X.

Internal searches the system runs to fact-check or freshen an answer. You never see them. You see the citations.

Share of answers that attach retrieved sources versus pure memory. Engine-level stat, useful context for citation rate.

Google’s language for pages made for humans. It also describes what is worth citing: experience, specifics, no empty calories.

Combining keyword (BM25) and vector retrieval. Many RAG stacks do this. Your nouns and your meaning both count.

Compression, dimensions, alt text. For AI, alt and captions are the text a multimodal model can quote when the pixels are opaque to a text crawler.

Making images findable in Lens, image search, and multimodal chat.

A ping protocol so participating engines fetch changed URLs faster. Bingy. Not a ChatGPT button.

An Influence Plan is Reddex’s action list: Reddit, AI, search, third-party, and content tasks ranked by what would actually move visibility.

What your page adds that wasn’t already on the SERP or in the model’s memory: a number, a method, a counterexample.

Links within your site. They pass attention to crawlers and humans. Use normal language anchors.

Whether a bot executes JS. Many AI crawlers are worse at this than Googlebot. If the article is empty without JS, you are invisible.

The date after which the model’s weights don’t include the web. Anything newer needs browsing or RAG.

A knowledge graph stores entities and relations: this company is that org, founded then, aka these names. Search and AI both sip from graphs.

Plural because Google is not the only graph. Product graphs, Wikidata, internal company graphs all feed answers.

Google’s entity box. It influences how people and some AI products describe you.

A neural net trained to predict text at huge scale. GPT, Claude, Gemini, Llama are LLMs. They power AI search.

LLMO is the content and entity work that makes it easier for LLMs to retrieve, trust, and quote you.

The need under the query. “Best CRM” may mean “won’t embarrass me in a demo” or “has a free plan.” Fan-out tries to guess this.

Getting other sites to link to you. Still useful. Paying for junk is still junk. For GEO, a mention without a link can matter too.

Meta’s open-weight model family. It shows up inside company tools and Meta AI. You meet it as a wrapper, not a website.

The footnotes models show. Density varies wildly by product. Treat them as a different currency per engine.

Structuring pages so models can quote them: entities, chunks, evidence, schema. Alias of LLMO/GEO craft.

Measuring model quality on tasks. Internal evals for your GEO classifier are useful. Public leaderboards are weakly related to citations.

RAG, browsing, refusal, citations, human review. You mitigate on your bot. Public ChatGPT mitigates on theirs.

Rank tracking for answers: position and inclusion across prompts. [[prompt-tracking|Prompt tracking]] is the clearer name.

Content that survives retrieval: self-contained sections, facts with dates, names spelled the same, HTML a bot can see.

llms-full.txt is a fat Markdown dump of site content for models that want one file instead of a crawl.

llms.txt is a Markdown file at the site root that tells AI crawlers what the site is and which pages to read first. It is not a second robots.txt with legal teeth.

NAP mentions in directories. They still help local pack and some “near me” assistant answers.

Maps, NAP, reviews, local pages. AI assistants still ask “near me.” GEO here is reviews plus Reddit plus Maps, not another city landing page mill.

Specific, lower-volume phrases. Chat prompts are almost all long tail with extra context.

Algorithms that learn from data. LLMs are a branch. Saying “we use ML” on a startup page explains nothing.

Mention rate is how often your brand name appears in the answer text, with or without a link.

The tone of a single mention — a Reddit comment or an AI sentence — as opposed to a rolled-up score.

Google’s product feed hub. Still the spine for Shopping and for some AI shopping answers that drink Google’s catalog.

Meta’s assistant across Facebook, Instagram, WhatsApp, plus Llama in the background. A discovery surface if your buyers live in those apps.

AI across Bing, Windows, Edge, and Microsoft 365. Bing visibility and work-document context both matter.

Google primarily crawls the mobile version. If mobile hides facts that desktop shows, you lose them — and some AI bots may see the poor version too.

Anthropic’s open standard for connecting models to tools and data. Reddex exposes MCP for agents that should query visibility, not scrape HTML.

The model blends several sources into one answer. You might be in the blend without a footnote, or footnoted but flattened.

Models that take images, audio, video, not only text. Screenshots of your UI and YouTube reviews become sources.

Preparing media so visual and video search can use it: captions, transcripts, filenames, schema, not just pretty pixels.

The field of machines dealing with language. LLMs ate most of the marketing conversation. You don’t need an NLP section on a mattress site.

Full-sentence searches. Default in chat. Increasingly default in Google too.

A proposed way to give a site a natural-language interface so agents query it instead of scraping.

Public-weight models (Llama, Mistral, Qwen, DeepSeek). They show up in company chat UIs you cannot see from the outside.

The company behind ChatGPT, GPT models, and several crawlers. A GEO counterparty whether you like them or not.

OpenAI runs several bots: GPTBot for training, OAI-SearchBot for ChatGPT Search, and ChatGPT-User for user-triggered fetches.

Google’s bundle of vitals, HTTPS, mobile, intrusive ads. Agents hate intrusive ads even more than Google does.

How fast the page becomes usable. Timeouts lose crawls. Users and agents leave.

What the model stored in its weights. No URL. Hard to update. Easy to be stale or confident and wrong.

Scoring a section, not the whole URL. AI Overviews already behave this way. Your H2 might win while the page “ranks” page 2.

Charging AI bots per request (HTTP 402 ideas). A publisher experiment. Most brands still want to be crawled for free.

Perplexity is an answer engine that leads with sources. If you care about citations and Reddit URLs, this is the microscope.

Perplexity’s browsing agent/browser idea. More on-page fetches, more user-triggered load.

Running the same question as different buyers: intern vs CISO vs parent. Shortlists change.

Structured product data assistants use to compare and maybe buy. Cousin of Merchant Center feeds.

App-like web with service workers. Fine for users. Easy to accidentally starve crawlers if the shell is empty.

Crafting inputs to get better outputs from a model. A builder skill. Not how you make ChatGPT name your brand for other people.

Malicious text that tries to override a model’s instructions. A security issue for agents. Not a GEO growth hack.

Ongoing [[prompt-tracking|prompt tracking]]: same set, scheduled, diffs, alerts when sentiment or inclusion flips.

Prompt research finds the questions people actually type into AI assistants — the GEO version of keyword research.

Prompt tracking is running a fixed set of real customer questions through AI engines on a schedule and logging who got named.

How often a question is asked of assistants. Mostly estimated. Treat it like early keyword volume: directional.

Deals where labs pay (or get blocked) for using professional content. Changes which domains appear in answers.

Query fan-out is when one user question becomes many internal searches. You might get cited for a sub-query you never targeted.

RAG means the model looks up documents at answer time instead of relying only on what it memorized in training.

Google’s 2015 ML ranking system. History. Not a 2026 lever.

Fetching the live web beyond cutoff. How ChatGPT Search and Perplexity stay current — when they bother.

An XML idea for declaring AI licensing terms to crawlers. Early. Watch it; don’t bet the year on it.

Models that spend more inference compute to think longer (o-series, R1, etc.). They can pick sources more carefully — or overthink a Reddit anecdote.

Ask Me Anything: a live thread where you answer in public. Models later treat a good AMA as a primary source — including the awkward answers.

Reddit citations are when an AI answer points at a thread or comment as evidence. They are one of the loudest unofficial ranking factors in consumer categories.

A pinned or runaway thread that becomes the default citation during an incident. One URL, hundreds of comments, long half-life in AI answers.

Links in Reddit posts are often nofollow. They still pull humans and still get retrieved by AI systems that do not care about rel=nofollow the way Google might.

Reddit sentiment is the tone of posts and comments about your brand: positive, negative, mixed, or sarcastic enough to fool a naive classifier.

A phrase people use for ranking threads in Google or getting cited from Reddit. Half of it is just not being a spammer. The other half is being useful in the right sub.

Reddit Visibility Score is Reddex’s daily read on how visible and how healthy your brand is across tracked Reddit mentions.

Community-maintained wiki pages inside a sub. Often more complete than vendor docs, and frequently cited as “the list.”

Reddit-aware research means briefs and drafts start from what threads already argue, not from a blank keyword.

A synonym some vendors use for how often you’re cited. Define it or you’ll mix it with mention rate.

Mods or admins remove a post; it may still live in caches, archives, and model memory. “It’s deleted” is not “it’s gone from AI.”

A second-stage model that reorders retrieved chunks before the LLM writes. You need to get into the first-stage set or rerank never sees you.

How much of your important content is actually eligible to be retrieved: not blocked, not empty, not orphaned.

Checking whether the right passages come back for a prompt set. Search-quality work. GEO teams can do a poor-person version by reading citations.

Reinforcement learning from human feedback. How many chat models got their polite, sometimes sycophantic personality.

robots.txt is the fetch permission file at the site root. AI crawlers read it too. Misconfigure it and you vanish from training, search, or both.

Schema is structured data on a page so machines can parse entities, FAQs, products, and articles without guessing.

Improving visibility in search. In 2026 that includes classic Google and the AI answers that sit on top of it. SEO did not die. The KPI mix changed.

What Google (or Bing) shows for a query: links, ads, Overviews, packs. Chat is not a SERP, even if it replaced the habit.

A buzz phrase for being findable in AI, search, social, marketplaces. Fine as a reminder. Bad as a single KPI.

SEO plus UX. Sensible. Not a new science. Fast pages, clear answers, less pain.

Google’s 2023–2024 experiment that became AI Overviews. Use Overviews in current docs. Mention SGE only for history.

Why the query exists: learn, find a site, compare, buy. Chat intent is the same with more backstory in one message.

Labeling queries by intent, often with a model. Useful to bucket prompts. Dangerous if the labels are wrong and nobody checks.

An awkward phrase that usually means aligning content to how people ask. That’s just intent-aware writing.

Monthly count for a keyword in a search engine. Still useful for Google planning. A weak proxy for chat demand.

Matching meaning, not only strings. Embeddings made this default. Your synonyms still help the lexical half of hybrid search.

Tracking whether talk is good, bad, or messy — on social, Reddit, and now in AI answers.

Non-link blocks: Overviews, snippets, PAA, videos, maps. They steal clicks and sometimes feed AI.

Share of Model is your brand’s share of mentions in AI answers versus a defined competitor set, on a defined prompt set.

Restricting search to one domain. Some AI systems use site: internally. If they start site:ing you, congratulations — also keep that domain healthy.

Compact models for on-device or cheap inference. They hallucinate more without RAG. Your pages still need to be simple enough to retrieve.

When answers cite X, LinkedIn, YouTube, TikTok, Reddit, etc. Each network is a different evidence culture.

Likes and shares as alleged ranking factors. For Google, mostly a myth. For AI, a viral thread is not a “signal,” it’s a document in the index.

The pipeline stage that collects and filters chunks before writing. If you’re filtered out here, no footnote.

The habit of pointing at URLs. Product-dependent. When it’s on, it’s your traffic and your accountability.

Schema marking text suitable for voice playback. Niche. Harmless if the marked text is truly the answer.

Headings, lists, tables, consistent fields. Machines and skimmers both win. Not the same as stuffing JSON-LD.

Retrieving a passage instead of a page. The unit of AI visibility is often the chunk.

Subreddit analytics are public Reddex pages at /r/{name} with community stats, posts, and engagement — useful for picking where to listen.

How big and how relevant a community is for your brand: subscribers, active users, and whether your buyers actually post there.

Models agreeing with the user too much. Dangerous when the user is wrong about your brand.

Fake data that looks real, used in training. The web is filling with it. Original human evidence gets rarer and more valuable.

Legal/technical flags that reserve text-and-data-mining rights. EU-flavored. Partial crawler support.

Crawlability, speed, canonicals, schema, now AI bot access. The floor for GEO.

Spending more compute at inference to think longer. Reasoning models. Latency and cost go up; answers may get more careful.

Third-party mentions are brand references on sites you do not own. They drive a large share of AI citations.

Days from publish (or update) to the first AI footnote you can observe. A freshness SLA you invent for your team.

The chunks of text models bill and process. Context windows are in tokens. Your page might get truncated in RAG.

Topical authority is whether the web treats you as a serious source on a subject, not a one-off article farm.

A planning artifact: all the subtopics you intend to cover, in a tree. Useful. Easy to turn into a content mill if you fill every node with oatmeal.

Trying to influence future weights. Slow, uncertain, mostly the same as being widely and accurately discussed on the public web.

The neural design behind modern LLMs. Fun at parties with engineers. Not a page-ranking lever.

UGC citations are AI references to forums, reviews, comments, and social posts — the stuff you do not fully control.

Chat answers in the app can differ from API completions: tools, memory, browsing, safety, ads.

How it feels to use the site. Now includes how it feels for an agent. Confusing UX loses both species.

Finding nearest neighbors in embedding space. Core of RAG. Mushy pages cluster together and lose.

Titles, files, chapters, transcripts, thumbnails so video is discoverable. Transcripts are the GEO payload.

Search with an image as the query. Lens, Gemini photo questions, shopping cams.

Spoken queries via phones and speakers. Short answers, local intent, messy ASR.

The practice of being the sentence an assistant reads aloud. That’s answer-ready content plus, for local, Maps.

An IETF idea for crawlers to cryptographically sign requests so you can trust who they are.

A public knowledge base of entities. Editable, cited by graphs, useful for disambiguation if you are notable enough.

When syndicated copies of the same article flood AI answers, crowding out independent sources — including you.

A protocol riff on HTTP 402 so agents can pay for resources. Experimental economics for crawling and APIs.

A list of URLs you want fetched. Still worth having. AI bots may use them, or not. They cannot fix Disallow.

Google’s label for content that can hurt you if it’s wrong: health, money, safety. AI systems are extra twitchy here.

When models quote a video via captions/transcript rather than your site.

Estimating the business effect of answers that never send a visit: branded search lift, surveys, assisted pipeline.

Zero-click search is when the user gets the answer on the results page or in chat and never visits you.

Doing a task with no examples in the prompt. Models generalize. Your weird product category may get analogized to the wrong thing.