Loading...
Loading...
Reference
63 terms defined. The vocabulary of Generative Engine Optimization, with every entry linkable on its own.
Example
A 2023 job listing asking for an AEO specialist and a 2026 one asking for a GEO specialist usually describe the same work. If a vendor still leads with AEO, check when their material was last updated.
Example
An assistant told to book a table for four on Friday that opens the restaurant site, fills the form and confirms, rather than returning three phone numbers. It has no patience for an interstitial and cannot read a price baked into an image.
Example
A shopper says reorder the dog food I bought in March, cheapest seller with next day delivery, and the assistant completes it. Your product data decides whether you are the seller it picks, not your landing page copy.
Example
The five that matter commercially are ChatGPT, Google Gemini, Perplexity, Microsoft Copilot and Claude. They do not share an index: in one large 2026 synthesis, only about 11% of domains cited by ChatGPT were also cited by Perplexity.
Example
Filter 30 days of server logs for GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and ChatGPT-User. The distribution is usually uneven and often surprising, and it tells you who cares about your site rather than who exists.
Example
Because they sit on top of ordinary Google results, your conventional ranking work is your GEO work on this surface. Google-Extended does not remove you from them: the only lever is being absent from Google search entirely.
Example
A buyer types best invoicing tool for a small agency into ChatGPT instead of Google, reads one answer naming three products, and never sees a results page. If you are not one of the three, you were not fourth. You were absent.
Example
Useful as a search term, because it is what people type. Not useful in a strategy document, where GEO and SEO need to stay distinguishable from each other.
Example
Measured as a frequency, not a state: ten buyer questions, three platforms, three runs each, recorded as appearances out of three. I checked and we were there is not a measurement.
Example
Work on how your brand is described: one consistent name, an accurate category, and independent sources saying the same thing. Distinct from content GEO, which is work on individual pages.
Example
Your name inside an answer with no link. Your analytics record nothing, and the reader has just been recommended you by something they trust. Measuring only linked citations counts the rarest of four outcomes and zeroes the rest.
Example
Ask it about your company with web search turned off, then again with it on. The difference between the two answers is exactly what retrieval is contributing for you today.
Example
Four different outcomes get called this: unattributed use, a brand mention, a named source, and a clickable link. Only the last one shows up in your analytics.
Example
In one 2026 synthesis, 36% of Claude's journalism citations came from the previous twelve months, against 56% for ChatGPT. Earned coverage keeps working here longer than it does elsewhere.
Example
Anthropic publishes its crawler documentation and honours robots.txt. Blocking ClaudeBot is a decision to be absent from Claude's answers: defensible, but it should be a decision rather than a copied template.
Example
CCBot is upstream of everything else, since its archive has fed many training datasets. Blocking it affects an unknown number of downstream users, present and future, which makes it a broader decision than blocking one company's bot.
Example
Rewriting a page so its answer sits in one clear section near the top, with a date and a source attached. Distinct from brand GEO, which is work on how you are described elsewhere.
Example
Retrieval routinely finds more material than fits, so something is discarded before your page is judged on merit. Dense, self-contained sections survive that cut. Comprehensive but diffuse pages do not.
Example
I am renovating a kitchen, family of four, care about energy use, budget around 800 euros, what dishwasher? That would have taken four Google searches. Here it is one turn, and the follow-up refines it rather than restarting.
Example
Grounded entirely in Bing. A site well indexed in Google and poorly indexed in Bing is invisible here for reasons unrelated to its content. Bing Webmaster Tools is free and checking it takes twenty minutes.
Example
A bot that visits monthly while you publish weekly leaves most of your work invisible to it between visits. Compare crawl frequency in your logs against your publishing frequency before assuming a content problem.
Example
Being described as part of your category by other people, in places these systems read. This is the fix when you are absent from category questions but present when asked about by name.
Example
Not a checklist to satisfy. A page naming its author, dating its claims and linking its sources is genuinely easier for a machine to justify quoting than one that does none of those things.
Example
Your company is one entity, your product another, you a third. Every variant name you use splits the evidence a system has to work with, which is why entity work is mostly removing ambiguity you introduced yourself.
Example
One name for one thing everywhere, the category words your buyers actually use rather than the ones you wish existed, and explicit relationships: who made this, who is it for, what does it compete with.
Example
The Google feature that trained the industry to write answer-first paragraphs. That habit transfers directly: the passage that won a snippet is usually the passage a retrieval system extracts.
Example
Not a GEO lever. You cannot fine-tune someone else's model to like your brand, and a service claiming otherwise is selling something else under this name.
Example
Cites more selectively than its peers, which raises the bar. On a surface naming two or three sources, being the fourth best answer is indistinguishable from being absent.
Example
The reader is handed a conclusion instead of a shortlist. That converts a gradient into a threshold: in classic search fourth place was a business, in an answer citing three sources it is silence.
Example
Introduced in a paper submitted to arXiv in November 2023 and published at ACM SIGKDD 2024, which found that adding citations, quotations and statistics raised visibility by up to 40% on the authors' own benchmark.
Example
Ten buyer questions, three platforms, three runs each: ninety readings and about an hour. That tells you more about your position than any index published by a company selling GEO software.
Example
Controls training use only. It does not remove you from AI Overviews, which run on the ordinary Google index through Googlebot. This is the most commonly misunderstood control in the field.
Example
One of three OpenAI agents with different jobs. Blocking GPTBot keeps your content out of training. Blocking OAI-SearchBot or ChatGPT-User removes you from answers being generated right now, which is a different decision.
Example
A grounded answer is constrained by retrieved documents and can show its sources. It does not follow that the answer is correct: grounding controls where the words came from, not whether they are true.
Example
Thin coverage of your brand does not produce silence, it produces invention. The fix for being described wrongly is rarely a correction request and usually more material for the system to be right from.
Example
Why an assistant with no web access describes a two-year-old company as though its last funding round were its current state. It is not out of date by accident: it is a snapshot, and snapshots do not update.
Example
Wikipedia and Wikidata feed it, and both apply notability rules that cannot be shortcut. That is exactly why presence there carries weight.
Example
It breaks text into tokens and predicts the next one. There is no database of facts inside it and no lookup step, which is why it can be fluent, confident and wrong without anything having malfunctioned.
Example
The same work as GEO under a different label. Worth recognising when you meet it, not worth adopting: a field with three names for one practice is harder to discuss than one with a single name.
Example
A proposed convention with real interest behind it and no confirmed universal support from the major systems. Adding one costs an hour and harms nothing. Treating it as a ranking factor is ahead of the evidence.
Example
Preparing for it is unglamorous and pays off under current conditions anyway: real alt text rather than keyword padding, transcripts for video, and product information as text near the image rather than baked into it.
Example
Why matching a user's exact phrasing matters less than it did: the system decomposes the question and searches its own rewritten terms, which are frequently not the user's words.
Example
Ask an assistant about your company with web search turned off. What comes back is this layer. You cannot edit it, and neither can anyone selling you a service that claims to.
Example
Searches on essentially every query and cites far more sources per answer than its peers, which makes it the most responsive testing surface available. Publish something genuinely better and this is where you see it first.
Example
Because Perplexity cites more heavily than any other major platform, blocking it forfeits more citation opportunity per unit of traffic than blocking anyone else.
Example
Write down the ten your buyers would actually use, in their words rather than your keywords: is X worth it for a small team, rather than X pricing.
Example
The layer that responds to work you do this quarter. It depends on whether a crawler can reach your page, whether your answer survives being read out of context, and whether the claim is on the page rather than implied by it.
Example
Not a ranking to reverse engineer. A sequence of five filters: interpretation, retrieval, selection, synthesis, attribution. You are eliminated at the first one you fail, and the fix at each is different.
Example
A request, not enforcement. It does not remove content already collected, does not bind operators who ignore it, and does not stop a person pasting your page into an assistant.
Example
No major AI system has confirmed it as a direct input to answer selection. What it demonstrably does is clarify entities and improve conventional search, which feeds retrieval. It will not rescue a page that fails at extraction.
Example
Ask five assistants what is your brand and who is it for, in fresh sessions. The gap between that answer and what you would have written is your actual problem statement.
Example
The surface GEO is usually contrasted with, though the contrast is overdrawn: AI Overviews sit directly on top of it, so classic ranking work remains the foundation rather than a legacy concern.
Example
Organization, Person, Article, FAQPage and BreadcrumbList earn their keep on most sites. Mark up what is actually on the page: schema describing content a reader cannot find is a claim about a page that does not exist.
Example
Mostly ordinary engineering: server-rendered HTML, honest heading hierarchy, working canonicals, and answers that exist without javascript. AI crawlers handle javascript less reliably than Googlebot, not more.
Example
Roughly four characters, or three quarters of a word. It matters because everything the model considers at once must fit a fixed budget of them, which is what forces material to be discarded before merit is judged.
Example
What you publish today is a deposit against a training run that has not happened yet. This is the layer deciding whether an assistant with no internet access still names you.
Example
Matters more than it did, because the system rewrites the question before searching, and what it rewrites toward is intent rather than phrasing.
Example
Structured, machine-readable, and disproportionately represented in what these systems retrieve. Notability rules apply and cannot be shortcut, which is precisely why an entry carries weight.
Example
In published samples, Wikipedia alone has accounted for between a quarter and roughly half of ChatGPT's top-10 citation share. You cannot write your own entry, and attempting to is a well documented way to make things worse.
Example
The reader got their answer and never loaded your page. Uncomfortable, and also the point: a named mention inside an answer someone asked for carries weight an impression on a results page never did.