The board question is no longer whether you are visible. It is whether you are named.

What is Answer Engine Optimisation?

Answer Engine Optimisation is the discipline of being recommended, by name, inside the answer an AI engine generates when a buyer asks it for a shortlist. It is not the discipline of being indexed, nor the discipline of being ranked. It is the discipline of being cited. For the board, that distinction is a governance question, not a technical one. The buyer who asks an engine for “the leading regulated-advisory firms in the UK” reads one answer, names three firms, and investigates one. The firm that is not in the answer is not in the conversation. This page sets out what AEO actually is, how the major engines decide whom to cite, the four signal layers a firm has to address, the working lexicon, and the most common ways serious firms misread the discipline.

What Answer Engine Optimisation actually is

Answer Engine Optimisation, abbreviated AEO, is the set of practices a firm uses to influence whether the major AI engines (ChatGPT, Claude, Perplexity, Gemini, Copilot, and the smaller specialist engines downstream of them) name that firm inside a generated answer to a buyer-intent prompt.

The unit of value is a citation, not a click. A citation is a named mention, sometimes carrying a linked source, inside the body of the engine’s response. The reader rarely leaves the response to verify it, because the engine has already done the abbreviating on the reader’s behalf. The citation IS the recommendation. A results page hands the reader ten options and trusts the reader to investigate; an answer hands the reader three names and trusts the engine to have chosen well.

AEO overlaps with Generative Engine Optimisation (GEO), and in most live engagements the two terms refer to the same body of work. Where a technical separation is drawn, AEO emphasises citation outcomes inside answer engines specifically, and GEO covers the wider class of generative-surface visibility including image and code generation. Inside a board paper the two are treated as synonyms, with AEO the prevailing term.

The discipline is new in that it could not have been practised before late 2022, and old in that its mechanics (entity recognition, citation density, structured-knowledge supply, authority signal) sit on decades of information-retrieval research. What is new is the engine class: the buyer is no longer reading ten links and choosing, the buyer is reading three names and accepting.

How AI engines pick what to cite

Citation mechanics vary by engine, and not all engines publish their methodology. What is known, from technical disclosures and empirical probing, is that citation behaviour is produced by three retrieval mechanisms layered over a base model.

The first is training-corpus inclusion. The base model has been trained on a snapshot of the public web up to its cut-off. Firms named repeatedly, in authoritative third-party sources, during that snapshot tend to surface even when the engine is not connected to live search. This is the slow, durable signal. It moves on training cycles, not edit cycles.

The second is retrieval-augmented generation, or RAG. When the engine is connected to a live search or web-grounding system (the default on the consumer surfaces of all the major engines for current-information queries), it retrieves a small set of candidate sources, ranks them, and uses them to ground the answer. The candidate set on a given query is often only three to ten documents. A firm that does not appear in the candidate set cannot be cited regardless of training history.

The third is real-time grounding from publisher signals. Several engines now consume publisher-supplied machine-readable surfaces (Schema.org structured data, the llms.txt and llms-full.txt convention, FAQ data, Knowledge Graph entries, Wikidata) as additional ranking input alongside classical link signal. Where these signals are clean, consistent and authoritative, the engine both retrieves the firm at the candidate stage and confirms its identity at generation. Where they are absent or contradictory, the firm is harder to cite confidently even when mentioned elsewhere.

Each major engine weights the three differently. ChatGPT leans on RAG when web browsing is on and falls back to training-corpus inclusion when it is not. Claude follows a similar mix with more conservative citation policy. Perplexity is structurally RAG-first and surfaces sources for every claim. Gemini blends classical search ranking with grounding metadata. Copilot uses a Microsoft-controlled retrieval layer over the broader index. A firm can be highly cited on one engine and absent from another; the audit has to be run per engine, not aggregated.

The four signal layers

The work of Answer Engine Optimisation, expressed in the way a board can act on it, sits across four signal layers. A serious programme addresses all four; an agency add-on usually addresses one and a half.

Content surface. The editorial discipline of the firm’s own site and its third-party publishing footprint. Publisher quality (depth, accuracy, freshness), publication cadence, and the editorial archive of named-author analysis. AI engines preferentially cite sources that read as published rather than as marketing. This is the layer most overlapping with classical SEO content quality, and a firm disciplined on long-form authority content for years starts from a stronger foundation.

Entity recognition. The layer that decides whether the engines know who the firm is in the precise technical sense: disambiguated, named consistently, present in the knowledge surfaces the engines rely on. It covers the firm’s Wikipedia entry, Wikidata entity, Knowledge Graph card, profile across authoritative directories, and the canonical naming the firm uses about itself. Most firms have at least one entity-recognition problem (a missing Wikidata entry, a Wikipedia page that conflates them with another firm, a Knowledge Graph card with an outdated address, a sector descriptor the firm itself has retired). Each problem reduces the engines’ confidence in citing them.

Authority graph. The named-citation network. Where the link graph of classical SEO measures hyperlinks between domains, the authority graph measures named mentions of the firm inside authoritative content across the wider web (regulator publications, sector trade press, professional bodies, named third-party analyses, peer-reviewed sources). It cannot be purchased and does not respond to inventory bidding. It responds to documented advisory work, named-author publications, and inclusion in third-party research. It is also the signal that compounds across every major engine simultaneously, because every engine downstream of the public web treats authority-graph density as a confidence input.

Structured machine-reading surface. The publisher convention layer: Schema.org JSON-LD on every page that warrants it, the llms.txt convention at the site root, the llms-full.txt companion file, FAQ data shaped for ingestion, factsheet pages built for retrieval rather than reading. The newest of the four layers and the one most firms have not yet addressed. It is also the cheapest to fix in absolute terms, because it is implementation work rather than authority work. A clean structured surface does not make a firm citable on its own, but its absence makes a firm harder to cite even where the other three layers are strong.

The lexicon

The board needs a working vocabulary to read its own evidence. Twelve terms recur in any Qyliq engagement document, defined briefly below.

Answer Engine Optimisation (AEO). The discipline of being cited, by name, inside the answer an AI engine generates in response to a buyer-intent prompt. The unit of value is a named citation, not an organic click.

Generative Engine Optimisation (GEO). A broader term for visibility across generative AI surfaces, often used interchangeably with AEO. Where the two are separated, GEO covers the wider class (including image and code generation) and AEO emphasises citation outcomes inside answer engines.

Citation share. The percentage of buyer-intent prompts on which a named firm appears, by name, inside the recommendation set the engine produces. Scored per engine, per prompt set, on a fixed cadence. The headline number a board paper carries.

Named-entity recognition. The model behaviour by which an engine identifies that a string of characters refers to a specific real-world entity (this firm, not a similarly named firm). The quality of entity recognition decides whether the engine cites confidently or hedges.

Authority graph. The network of named mentions of a firm inside authoritative third-party content across the web, distinct from the link graph that measures hyperlinks alone. The dominant signal behind cited recommendations.

llms.txt convention. A publisher convention, introduced in 2024 and adopted by a growing number of firms, in which the site root carries an llms.txt file summarising the site for AI ingestion, optionally with a longer llms-full.txt companion. Not a settled standard, but increasingly treated as a baseline by serious publishers.

Retrieval-augmented generation (RAG). The technical pattern by which an AI engine retrieves a small set of candidate sources from a live index, ranks them, and uses them to ground the generated answer. The default architecture behind most consumer-facing AI engines in their web-grounded mode.

Knowledge graph. A structured database of entities and their relationships, used to disambiguate and resolve named entities. Google’s Knowledge Graph is the largest, with parallel structures in other engines and in Wikidata.

Probe set. A documented set of buyer-intent prompts, run against each major AI engine on a fixed cadence, used to measure citation share over time. The instrument behind any credible AEO measurement.

Buyer-intent prompt. A prompt phrased the way a senior buyer would ask when looking for a shortlist (for example, “Which firms lead on cross-border M&A advisory in the UK”). Distinct from a long-tail keyword in shape and volume.

Visibility briefing. The default Qyliq per-board-cycle artefact: a board paper summarising current citation reality, competitor position and the actions taken or scheduled. Distinct from the recurring Share-of-Voice Quarterly, which is the underlying measurement instrument.

Boardroom-tier reporting. The seniority register Qyliq publishes for board-level audiences, intentionally distinct from dashboard-tier reporting. Board-tier reads as a paper on the table; dashboard-tier reads as a numerical feed.

The instrumentation problem

Most firms cannot answer the question “are we cited” because the measurement surface required to answer it did not exist inside the tooling they bought for classical SEO. Classical tooling crawls the web, indexes rankings, and reports organic metrics. The AI engines do not expose citation behaviour through those crawls, because the citation behaviour is generated on the fly inside a prompt-response loop the crawler is not inside.

Honest measurement of citation share requires probe-based sampling: a documented set of buyer-intent prompts, run against each major engine, on a fixed cadence, captured and scored against a named competitor set. There is no crawl path to the answer. There is only the probe. A firm that does not run probes has no measurement of citation share, regardless of which dashboard its agency presents.

This is why most firms who think they have an AEO programme do not. An “AI mentions” column has been added, the keyword list re-tagged with AI-flavoured terms, the existing crawl reporting against the existing index. The new measurement (probe-based, engine-by-engine, cadence-stable, competitor-aware) is not being run. The firm is paying for a discipline it is not receiving.

The implication for the board is straightforward. The first useful question is not “what is our AEO score”. It is “show me the probe-set, the engines we are running it against, and the cadence”. If the answer is absent or vague, the discipline is absent.

Where most firms get this wrong

Three failure modes recur across the firms Qyliq has examined in 2025 and 2026.

Treating AEO as “SEO plus a column”. The most widespread failure. The firm has signed an agency retainer with an AI-visibility paragraph added to the scope statement and an “AI mentions” tab added to the dashboard. The underlying methodology, probe set, engine coverage and cadence are unchanged. The firm believes it is buying the new work; it is not. The cost is a year of competitive runway, because the citation surface is moving in eight-week windows while the firm is reviewing it on twelve-month renewals.

Outsourcing to agencies shipping ranking reports with new labels. Adjacent but distinct. The agency does ship something new, but the something is a relabelled ranking report. The same keyword basket is now scored against a generic “AI visibility” weighting the agency has invented. The number is not citation share against a named competitor set, and cannot drive a board decision. The artefact looks bespoke and reads as familiar, which is exactly the comfort the buyer pays for and the failure the firm cannot diagnose.

Measuring presence in answers without scoring against a competitor set. Subtler, common among firms that have started running probes but have not built the comparative discipline. The firm reports that it was named on, say, 40 per cent of prompts. The number may even be progress against its own baseline. What it does not say is whether 40 per cent is leadership, parity or trailing inside the named competitor set on the same prompts. Without the comparative frame, the measurement cannot drive prioritisation; absolute citation share is a vanity metric.

The Qyliq position

Qyliq builds and runs Answer Engine Optimisation programmes for firms operating at board-tier governance. The diagnostic entry is The Authority Graph Audit, which takes a firm’s named competitor set, runs a probe across the major AI engines, and surfaces the citation reality together with the four-signal-layer view of why the firm sits where it does. The recurring instrument is The Share-of-Voice Quarterly, extending the probe into a fixed-cadence measurement programme. The default board-cycle artefact is The Visibility Briefing; the premium tier, written for firms operating at full board governance, is The Boardroom Visibility Briefing.

The full six-phase advisory programme behind these artefacts is documented at the methodology page. The category-level frame distinguishing AEO from classical SEO is set out at Answer Engine Optimisation vs SEO. The publisher convention layer named in §4 above, including the practical specification of llms.txt and llms-full.txt, is at the llms.txt guide.

The recommended starting point is the diagnostic. The Authority Graph Audit turns the discussion on this page into a board paper on the table.

Frequently asked questions

Is Answer Engine Optimisation the same as Generative Engine Optimisation? Yes, the two labels describe the same discipline. Answer Engine Optimisation emphasises the output the engines produce, Generative Engine Optimisation emphasises the engine that produces it; the underlying work of being cited by name is identical.

Which AI engines does AEO need to address? At minimum ChatGPT, Claude, Perplexity and Gemini, with Google’s AI Overviews treated as a sibling surface to classical search rather than a separate engine. The right list for any given firm is the set its actual buyers are using, not a generic top-four.

Can a firm be optimised for AEO without being strong on classical SEO? Rarely, because the AI engines lean heavily on the same web of citations, structured data and editorial authority that classical search has indexed for two decades. Strong AEO is usually built on top of competent SEO, not in place of it.

How do AI engines decide which firms to cite? Through a combination of named-entity recognition, authority-graph density, retrieval-augmented generation against current web sources and editorial signals from places like Wikipedia and Wikidata. The Qyliq Authority Graph Audit maps where a firm sits across each of those inputs against its named competitor set.

What’s the role of Wikipedia and Wikidata in AEO? Both function as ground-truth references the AI engines lean on for entity disambiguation and factual grounding. A firm that exists cleanly in those graphs is materially easier for the engines to cite than one that does not.

Do AI engines weight publication date in citation decisions? Yes, especially for retrieval-augmented surfaces like Perplexity and Google’s AI Overviews, where recent and authoritative sources outrank older equivalents. Editorial cadence becomes a citation-share lever, not just a content-marketing rhythm.

What does the first 90 days of an AEO programme look like? A diagnostic baseline of current citation share, an authority-graph audit against the named competitor set and an opening Boardroom Visibility Briefing that translates the findings into a 12-month plan. Implementation work begins in the second quarter, once the board has signed off on the priorities.

How is AEO measured, in plain terms? By running a documented prompt set across the major AI engines on a fixed cadence and scoring citation share against a named competitor set rather than tracking abstract rank positions. The Qyliq Share-of-Voice Quarterly is the recurring board paper that carries that signal.