The Answer Intent Framework
A four-part methodology for being cited in AI search.
The Answer Intent Framework is the methodology Cameron Duncan developed to systematically improve AI citation rates for South African founder-led businesses. It has four components: Intent (the actual question), Structure (the technical infrastructure), Extraction (the writing discipline), and Measurement (the accountability discipline). Every Answer Intent engagement applies all four.
The actual question your audience is asking.
The infrastructure that makes answers extractable.
The writing discipline that makes paragraphs citation-ready.
The accountability discipline. Share of Model Response.
Why SEO thinking does not produce AI citations.
Most South African businesses approach AI search the way they approach Google. They keyword-stuff, they tweak meta descriptions, they ask "what is my AI SEO strategy?"
The premise of this framework is that AI search is not search. It is question-answering. The AI engine is not retrieving pages and ranking them. The engine is constructing an answer in real time using extractable evidence from sources it has decided to trust.
Two implications follow. First, the unit of optimisation is the answer, not the page. A page that ranks can still fail to be cited if its content is not structured as an answer. Second, the work is structural and editorial, not promotional. Adding keywords does not help. Adding schema does. Rewriting service pages as definition layers does. Building LLMS.txt files does.
The four components of this framework address both implications. They identify the right questions to answer (Intent), build the infrastructure to make answers extractable (Structure), apply the editorial discipline that makes paragraphs citation-worthy (Extraction), and measure whether citations actually materialised (Measurement).
Intent
The actual question your audience is asking when they query an AI engine.
The first question every AEO engagement must answer is: what is your audience actually asking? This sounds obvious. It is not. SEO keyword research operates in a compressed cognitive register - three to five words, noun phrases, stripped of context. "AEO agency South Africa." "KZN safari lodge." "Tax consultant Cape Town." Google's ranking algorithm can infer intent from partial signals. AI engines cannot.
When a user prompts an AI engine, they speak in full sentences with embedded context. "Which AEO agencies in South Africa have experience with regulated industries?" "I need a private game reserve in KwaZulu-Natal that accommodates children under twelve - which operators are trustworthy?" "Who is the best tax consultant in Cape Town for a software startup with cross-border revenue?"
These are not longer keywords. They are different cognitive objects. The intent embedded in an AI prompt carries location, audience type, constraints, and qualifying criteria. If your content does not answer the full question, you do not get cited - even if you would rank for the keyword.
The Answer Intent audit methodology maps five prompt types for every client engagement:
-
Direct brand queries
"What is [Business Name]?" and "What does [Business Name] do?" - tests whether the business is a named entity AI engines can identify and describe.
-
Category-discovery queries
"Who provides [service] in [location]?" - the most common buying-intent query type. Requires service schema and extractable service definitions.
-
Problem-led queries
"My business has [specific problem] - how do I solve it?" - requires content structured as diagnosis followed by recommendation.
-
Comparison queries
"What is the difference between [option A] and [option B]?" - requires explicit comparative content with named entities.
-
Partner or donor pathway queries
Used in fundraising, B2B services, and institutional contexts. Requires authority signals - Person schema, published content, named clients or outcomes.
Each prompt type requires a different content response. Most South African businesses have content that answers none of the five types - not because the content is poor, but because it was written for humans skimming a homepage, not for AI engines extracting structured answers to specific questions. Intent mapping is the work that makes everything downstream accurate.
Structure
The technical infrastructure that makes your answer extractable.
Structure is the technical infrastructure that makes an answer extractable. A correct answer that is not extractable is, for AI citation purposes, the same as no answer at all. Technical structure precedes editorial quality. If the structure is absent, it does not matter how well the content is written.
Three elements constitute structure in the Answer Intent Framework:
Schema - the entity vocabulary
JSON-LD schema markup, using the Schema.org specification, tells AI engines and search engines what type of entity each page represents. An Organization. A Person. A Service with a defined price and deliverable. A FAQPage with discrete question-and-answer pairs. A HowTo with sequenced steps. Without schema, AI engines must infer your entity type from unstructured page content, and inference produces unreliable citations. With schema, you are a named, typed, citable entity.
LLMS.txt - the direct channel
A structured plain-text file placed at your domain root, LLMS.txt tells AI language models exactly what your business does, who it serves, what your core pages are, and what you want cited. As of 2026, it is the closest approximation to a direct communication channel between a website and an AI retrieval system - bypassing the crawler's inference layer. Without it, AI engines have to construct your business description from scattered page content, which produces incomplete or absent citations. Answer Intent builds LLMS.txt to a defined specification for every engagement.
Content architecture - the routing layer
How your pages are routed (clean, stable, crawlable URLs), how your definitions are positioned (at the top of every page, before promotional language), how your FAQ clusters are grouped (by topic, with anchor IDs for direct linking), and how your internal linking routes from answer pages to service pages. A website with correct schema and LLMS.txt but poor content architecture will still underperform in AI citation.
The structure-first principle matters because content extraction depends on structural context. An AI engine cannot reliably extract and attribute a claim about a named individual, organisation, or service if that claim sits inside a promotional paragraph with no schema context, no entity anchor, and no clear page type. Structure creates the container. Extraction fills it.
Extraction
The writing discipline that makes paragraphs answer-shaped and citation-ready.
Extraction is the writing discipline that makes paragraphs citation-worthy. It is the component most agencies skip, because it requires rewriting - not adding technical infrastructure to existing promotional content. Deploying schema on a page that was written to persuade rather than to answer produces minimal citation improvement. The schema declares the entity. Extraction makes the entity worth citing.
Answer-first structure
Every paragraph that could be cited must be the answer to a specific question. Not every paragraph needs to be citable - but those that should be cited must read as if the question is printed above them. The answer comes first, before context, before qualification.
Third-person voice
AI engines constructing answers do not quote in first person. "We deliver three assets" cannot be attributed. "Answer Intent delivers three deployment-ready assets" can be cited and attributed correctly. Third-person voice is not a style preference - it is a citation requirement.
Entity-first sentences
The first sentence of any extractable paragraph should contain the primary entity name - the business name, the person's name, the service name. "Answer Intent is a South African AEO and GEO agency" can be cited. "We are a South African AEO and GEO agency" cannot be attributed reliably.
Verifiable claims
Numbers, dates, named entities, measurable outcomes. "Best in class" and "leading provider" are invisible to AI extraction - they are unverifiable promotional assertions with no citation value. "Answer Intent delivers validated JSON-LD schema, a structured LLMS.txt file, and AI-extractable content in a 10-day sprint" is specific, named, and citable.
Definition layers
A 50-80 word paragraph at the top of every page, written in third person, entity-first, with at least one verifiable claim. Definition layers are the single highest-leverage edit most SA business websites can make. The edit costs a copywriter two hours. The citation potential it creates persists across all AI engines for as long as the page is live.
The discipline most agencies avoid is also the discipline that produces results. Technical infrastructure alone - schema and LLMS.txt without extraction-ready content - produces partial citations at best. The combination of correctly structured data and extractable prose is what moves a business from invisible in AI-generated answers to cited in them.
Measurement
The accountability discipline of knowing whether it worked.
Measurement is what makes AEO a discipline rather than an opinion. Without a defined protocol for tracking citation rates, AEO is a service sold on faith - and faith is not a reasonable basis for a marketing investment. The fourth component of the Answer Intent Framework establishes what to measure, how to measure it, and what the results mean.
Share of Model Response (SMR) is the primary metric. SMR measures the percentage of a defined prompt set in which the business is cited in an AI-generated answer. A business with an SMR of 0 percent is invisible to AI engines across that prompt set. A business with an SMR of 30 percent is cited in roughly one in three relevant queries. SMR is interpretable, trackable over time, and comparable across engagements. It is not a vanity metric - it is a direct measure of whether the intended audience encounters the business when asking relevant questions.
30 prompts, 3 engines, 90 data points
Before any implementation work begins, Answer Intent runs 30 defined prompts across ChatGPT, Perplexity, and Google AI Mode. Prompts are selected to reflect the five intent types mapped in Component 1. Each response is reviewed for presence, accuracy, and attribution. The baseline establishes the citation rate before any work was done - and provides the comparison point for every subsequent retest.
Same prompts, same protocol, compared to baseline
The Day 30 retest runs the same 30 prompts against the same three engines after implementation. The comparison between Day 0 and Day 30 shows whether the schema deployment, LLMS.txt, and content changes produced measurable citation improvement. Results are documented and reported with individual prompt-level data.
The outer boundary of the AEO Proof Guarantee
For Tier 2, Tier 3, and Tier 4 engagements, if the business has not been cited in at least one relevant AI-generated answer within 60 days of implementation, Answer Intent continues working at no charge until citation is achieved. The guarantee is not a refund - it is a continued-work commitment. The measurement protocol is what makes the guarantee verifiable rather than arbitrary.
Ongoing Programme tracking
Tier 3 Ongoing Programme engagements include monthly retesting against the same core prompt set, with SMR tracked across time. The monthly cadence allows for iterative improvement - content is adjusted based on which prompts are not producing citations and why.
The role of measurement is not to prove that the work was done. It is to prove that the work produced citations. These are different claims. AEO that cannot be measured is indistinguishable from AEO that did not work.
How the four components work together.
Each component of the Answer Intent Framework depends on the one before it. They are built in sequence. They operate simultaneously. And the framework requires all four to function.
Intent comes first because if you optimise for the wrong question, every downstream effort is wasted. A business can have impeccable schema, a precisely written LLMS.txt file, and perfectly extractable content - and still not be cited, because they built the infrastructure around questions their audience is not asking. Intent defines the target. Everything else is aimed at it.
Structure is built second because content extraction requires a structural foundation. Schema creates the entity context without which citations cannot attribute correctly. LLMS.txt creates the direct channel through which AI engines receive explicit instruction about the business. Content architecture ensures that the right pages are routable, that answers cluster around questions, and that definition layers sit where AI retrieval systems expect to find them - at the top of each page, before promotional content. Without structure, extraction-ready copy has nowhere to anchor.
Extraction is the editorial work that makes the structure pay off. A schema-decorated page with promotional copy does not get cited. The writing discipline of third-person, entity-first, answer-shaped paragraphs is what converts structural infrastructure into actual AI citations. Most South African businesses that fail at AEO fail here - not because they lack schema, but because they have not rewritten their content as answers.
Measurement is what makes the previous three accountable. Without Day 0 and Day 30 data, there is no way to know whether the intent mapping was accurate, whether the structure deployed correctly, or whether the extraction discipline produced citation-ready content. Measurement closes the loop and - critically - identifies which prompts are still not producing citations, so the work can be refined.
Skip any of the four and the framework breaks.
What this looks like in practice.
Consider a Big Five operator on the KwaZulu-Natal coast. The business has a well-ranked website - page one for "luxury game reserve KZN" - but is completely absent from ChatGPT and Perplexity responses to the queries its target audience actually uses. This is a common pattern. Strong Google rankings do not transfer to AI citation. The two systems use different signals.
Intent mapping
The Answer Intent Framework begins by identifying the actual prompts the target audience is running. For this operator, they include: "private Big Five reserve near Durban that takes children under twelve," "best KZN safari for a family with a toddler," "what is the best game lodge on the Natal coast," and "KwaZulu-Natal game reserves rated highly for families." None of these match the SEO keyword the website was built around. All of them represent buying-ready intent at the decision-making stage of the booking journey.
Structure deployment
The implementation deploys LodgingBusiness and TouristAttraction schema with precise address, amenity list, and priceRange properties. A Person schema covers the lodge owner, making the principal a citable named entity with an identifiable professional identity. FAQPage schema covers the five most common booking questions (child policies, transfer distances, Big Five sighting rates, safari formats, seasonal availability). An LLMS.txt file is built at the domain root with explicit instructions for AI retrieval systems, including the primary service description, target audience, and key page routes.
Extraction rewrite
The lodge description page is rewritten as a 75-word definition layer, entity-first, third-person, with a named location and two verifiable claims (Big Five resident species, minimum child age). FAQ clusters are added covering child policies, transfer logistics from Durban and Durban airport, and the types of game drives offered. Each FAQ answer is written in third person, answer-first, with the lodge name in the first sentence.
Measurement
The Day 0 baseline shows 0 of 30 prompts cite the operator - invisible across all three engines. The Day 30 retest shows 4 of 30. The Day 60 retest shows 7 of 30. The operator had good content. They had a visible Google presence. What they lacked was the infrastructure and editorial discipline to make their answers extractable. The framework provided both.
This scenario is hypothetical and representative. The prompt types, schema choices, and measurement cadence reflect the standard Answer Intent process applied to any founder-led business with a local physical presence and an international or premium domestic audience. The citation trajectory - 0 to 4 to 7 prompts across 60 days - reflects the typical improvement curve when all four components are applied correctly.
How to apply the framework.
Do it yourself
Read the four components above and audit your site against each. Check whether your content is written in answer-first, entity-first, third-person paragraphs. Check whether you have a validated JSON-LD @graph block and an LLMS.txt file. Check whether your FAQ pages use FAQPage schema with anchor IDs for each question. Most SA business websites fail at least two of the four components. The framework tells you exactly which ones.
Visibility Audit
R4,900Answer Intent runs the framework's first and fourth components - intent mapping and Day 0 baseline measurement - for your business. You receive 30 defined prompts across three engines, 90 data points, a structured gap analysis, and a prioritised implementation backlog. This is the lowest-risk way to quantify the gap before committing to a build engagement. No CMS access required. Delivered in 5 days.
See the AuditFoundation Sprint and above
R14,900+Answer Intent applies all four framework components. Intent mapping. Schema, LLMS.txt, and content architecture deployment. Extraction-ready copy production. Day 0 baseline, Day 30 retest, and the AEO Proof Guarantee for Tier 2 and above - continued work at no charge until citation is achieved within 60 days of implementation.
See all packagesStart with the Visibility Audit.
The Visibility Audit runs all four framework components against your site and delivers a gap analysis in 5 days. R4,900. No CMS access required.