Reference
Glossary
The vocabulary in this field is unusually unsettled. There are three names for the same practice, and several widely repeated terms that do not mean what people think. These are plain definitions, including where the industry disagrees and where the evidence is thin.
Generative engine optimization (GEO)
The practice of structuring content, entities and third-party presence so that AI answer engines cite a brand when generating a response. Where classic SEO competes for a ranked position among ten links, GEO competes for a mention inside a single synthesised answer that names only a handful of sources.
Also called AEO or AI search optimization. None has settled as standard, and the practices described under each overlap almost entirely.
Answer engine optimization (AEO)
A synonym for generative engine optimization. Some practitioners draw a distinction, using AEO for direct-answer features and GEO for generated prose. It is not observed consistently and no engine documents it. Treat the terms as interchangeable and ask what someone means before agreeing to buy it.
The five gates
A diagnostic frame: every search or AI-answer failure occurs at one of five stages. Crawl: can a bot fetch the URL? Index: did the engine keep it? Retrieve: was it pulled for this query? Cite: was it named in the answer? Influence: did it shape what the answer said?
Classic SEO lives at gates one and two. AI-answer visibility lives at three, four and five. How each one fails, in detail →
Query fan-out
Google's term for the way a generative search feature rewrites a single user question into multiple concurrent sub-queries, each retrieving separately. The practical consequence is that you no longer compete for one keyword but across the set of sub-questions your buyer's question decomposes into. That changes what a page has to contain.
Retrieval crawler vs training crawler
Two different jobs, frequently conflated in the same robots.txt. A retrieval
crawler fetches pages to answer a live question and to attribute sources.
Blocking it removes any possibility of being cited today. A training crawler
gathers data to train future models. Blocking it affects model weights later
and is a slower, more arguable harm.
A third category is neither: tokens such as Google-Extended and
Applebot-Extended are directives, not crawlers. No agent fetches
under those names. Googlebot and Applebot do the fetching, and the tokens
govern what the content is subsequently used for.
robots.txt vs noindex
robots.txt controls crawling. noindex controls
indexing. They are not interchangeable, and confusing them is the
most common technical error in the industry.
If you disallow a URL in robots.txt, the crawler never fetches it, therefore
never sees the noindex, and the URL can still appear in results.
To remove something from an index you must allow it to be crawled.
X-Robots-Tag
A response header that carries the same directives a robots meta tag can carry, such as noindex and nofollow, set by the server rather than written into the page. It governs indexing, not crawling, and it is the only way to apply those directives to a file that cannot carry a meta tag, such as a PDF. A crawler has to fetch the URL to see it, so the same rule applies as for noindex: disallow the URL in robots.txt and the header is never read.
Canonicalised away
What happens to a URL whose page names a different URL as the canonical. If the engine honours the signal, the index keeps the named URL and the original stops being the one that ranks or is cited. rel=canonical is a strong signal, not a directive. An engine can choose a different canonical when the pages disagree with it, and often does when two pages are near-duplicates with different signals. Submitted, crawled and indexed are three separate stages, and this is a failure at the third.
Render queue
Google renders JavaScript, but not immediately. Pages that return 200 enter a
deferred rendering queue and may wait there considerably longer than the few
seconds people assume. Content that only exists after hydration is therefore
delayed for Google and frequently invisible to AI retrieval crawlers, most of
which do not render at all.
Soft 404
A page that returns HTTP 200 while being a "not found" state, most
often a host serving the homepage for every unmatched URL. It tells a crawler
that every mistyped path is real content, which is how sites accumulate
thousands of duplicate pages nobody wrote. Fixed by returning a 404
status.
llms.txt
A proposed plain-text file listing a site's key content for language models. There is currently no evidence it influences AI citation, and Google states directly that you do not need to create AI text files for its generative features.
Ahrefs found 97% of llms.txt files received zero requests in a month across
137,000 domains. It costs about an hour to publish and carries no downside, so
publish one if you like. Nobody should be charging you for it as a
mechanism.
Structured data
Machine-readable markup, usually schema.org JSON-LD, describing what a page is about. It is not a ranking factor. It creates eligibility for rich results, which is a different and narrower claim. Google states that no special structured data is needed to appear in its AI features.
FAQ rich results were retired entirely in 2026. They stopped appearing in May, documentation was removed in June, and Search Console API support went in August. FAQ schema no longer produces the rich result.
Core Web Vitals
Three field-measured metrics: LCP (largest contentful paint, good ≤ 2.5s), INP (interaction to next paint, good ≤ 200ms) and CLS (cumulative layout shift, good ≤ 0.1). Measured at the 75th percentile over a 28-day rolling window of real Chrome traffic.
Lab tools measure one device on one network. Only field data is what Google uses. A green Lighthouse score alongside a failing Search Console report is not a contradiction. Treat these as a tie-breaker and a conversion lever.
Crawl budget
The volume of crawling an engine will spend on a site. Almost no site has a crawl budget problem. Google's rough thresholds are a million-plus pages changing weekly, or ten thousand-plus daily. Below that, it is very unlikely to be the problem.
Citation vs mention
A mention is being named in an answer. A citation is being named as a linked source for it. The two come apart constantly. Engines name a brand while citing a directory or a video that describes it. Being mentioned without being cited means someone else's page is deciding how you are characterised.
Frozen prompt panel
A set of at least 120 prompts, written for the client's positioning rather than its category, approved by the client in writing, and then held fixed so that a reading taken this month can be compared with one taken in three months. Precision comes from the number of distinct prompts, not from re-asking one. A panel built around a category label instead of what the client sells will under-report them, and the resulting finding is an artefact of the prompt rather than a fact about the brand.
Noise floor
The variation a measurement shows from run to run when nothing about the site has changed. It is measured at onboarding, on the client's category and an approved competitor set, before any movement is reported, so that a later change can be told from sampling noise. The floor is not small: 65% of AI-cited sources change from one day to the next, and a single prompt asked once carries roughly 44 points of error. A visibility figure reported without a noise floor has no way to say whether it moved.
Confidence interval
The range around a reported figure within which the true value is likely to sit, given the sample that produced it. Every figure in a report carries one, per engine. Movement inside the interval is reported as flat, because it is indistinguishable from the sampling noise the interval describes. Most months, nothing measurable changed, and the report says so.
Measurement layer
Which instrument produced a figure. API sampling, the consumer interface captured by a person in a real browser, and search-result capture are three different measurements that retrieve differently and move at different times. They are reported side by side and never averaged into one figure, and every number says which produced it. A score that does not name its instrument cannot be compared with anything, including itself a month later.
Verified crawler
A request whose source address, not its user-agent string, shows it came from the crawler it named. Any request can present any name, so a crawl reading taken from user-agent strings counts imitations as crawlers and the crawlers a firewall blocked as absent. Verification is by the address: membership in the ranges the operator publishes, or a reverse DNS lookup whose host name sits in the operator's domain and resolves forward to the same address. Google, Bing, OpenAI, Perplexity and Apple publish range lists. Some operators publish neither, and their traffic stays unverified. Every crawler figure in a log analysis is a count of verified requests, and the report says how many lines were discarded as unverifiable. Check one line →
Zero-click search
A search that ends without the user clicking through to any result. Measured rates vary by panel and method, and the figures quoted in this industry rarely carry their denominator. Ask for the sample size, the country and the window before accepting any of them, including ours.
Which gate are you failing?
The diagnostic audit answers that in ten business days. Prices and terms are published on the pricing page.