Short answer. A citation needs four layers to work, and the chain breaks at the first failure: access, extractability, evidence, refresh. Access is per-engine — different crawlers, different rules, and Google-Extended does not affect Search AI features while Perplexity-User generally ignores robots.txt. Extractability means answer-first passages under question headings; Google is explicit that no chunking or special markup is required and that no schema guarantees a citation. The useful diagnostic is not which agency is best, but which of the four layers is broken for you.
Models?
An AI citation is the last step in a chain, and the chain breaks at the first weak link. The engine has to be able to fetch the page. It has to be able to pull a clean passage out of it. That passage has to contain something worth repeating. And the page has to look current when the query runs. Fail any one and the other three do not matter.
Most agencies are built for one link. That is not a criticism; it is how specialist firms work. But it means "which agency helps your website get cited" has to be answered per layer, and the useful diagnostic is which layer is broken for you. This article sets out the four layers, the evidence for each, which agencies deliver it, and a way to find your weak link before you buy anything.
Layer 1: Access
What it is. The engine's crawler or fetcher can reach the page and read its main content. This sounds trivial and is the most common failure we see.
The evidence. The engines use different crawlers with different rules. Google's Search AI features use Googlebot; the separate Google-Extended token affects Gemini app training and grounding but has no effect on AI Overviews or AI Mode. Perplexity documents PerplexityBot for indexing and Perplexity-User for live fetches, notes that the user fetch generally ignores robots.txt, and publishes IP ranges with a WAF whitelisting guide because security rules block it. Anthropic runs three separate crawlers. Google says a page must be indexed and eligible for a snippet to appear in its AI features, and that it can process JavaScript but calls it more complex. Perplexity's fetcher works to a shorter time budget than a full render.
Who delivers it. Engineering-led agencies. Onely specialises in JavaScript rendering, headless CMS barriers and fragmented entity signals. iPullRank treats the discipline as engineering. Botify is strong where the constraint is that engines cannot retrieve pages at very large site scale. Go Fish Digital pairs technical SEO with its other practices. Lifewood's audit stage checks this first, from server logs, before anything is produced.
Who does not. Editorial studios generally do not publish this capability, and monitoring platforms flag access problems without fixing them.
Layer 2: Extractability
What it is. The page contains a self-contained passage that answers a question the engine is trying to answer, positioned where the engine will find it.
The evidence. Google's guidance says no chunking, special markup or AI-specific rewriting is needed, but also that readers appreciate paragraphs, sections and clear headings. The practical form is a question-shaped heading followed by a direct twosentence answer, which makes the boundary of an extractable span explicit. The GEO benchmark found fluency improvements lifted citation visibility 15% to 30%. A 2026 absorption study found high-influence pages richer in definitions, numbers, comparisons and procedures, the forms an engine can attribute faithfully. Google also warns that no schema type guarantees AI citation; structured data is a supporting signal, not a trigger.
Who delivers it. Content agencies with an answer-first editorial standard: Omniscient Digital, First Page Sage, Animalz, Siege Media.
Platforms with audits (Otterly's 25-plus on-page factors) identify the problem. Lifewood's Pillar Execution stage produces pages to this pattern, and its published editorial guidance on question headings is the specification.
Who does not. Engineering shops make the page reachable; they do not usually rewrite it.
The four layers, and where each agency type sits
1 · ACCESS 3 · EVIDENCE
Can Googlebot, PerplexityBot, Perplexity-User and Anthropic's crawlers reach
the main content?
Delivered by: Onely, iPullRank, Botify, Go Fish Digital; Lifewood audit stage
Does the passage contain quotations, statistics, methods the engine cannot
get elsewhere?
Delivered by: Siege Media (original data), First Page Sage, digital PR at Go Fish; Lifewood verified review layer 2 · EXTRACTABILITY Is there a self-contained, answer-first passage under a question heading?
Delivered by: Omniscient, First Page Sage, Animalz, Siege Media; Lifewood Pillar Execution 4 · REFRESH Is the page current, and does it change on a schedule?
Delivered by: any agency on retainer; Lifewood Performance Reporting cycle A citation requires all four. Find the first one that fails for you and buy that.
Layer 3: Evidence
What it is. The passage contains something the engine values and cannot get from a commodity page.
The evidence. This is the layer the GEO paper measured directly: across 10,000 queries, authoritative quotations lifted citation visibility up to 40%, statistics around 30%, and keyword stuffing scored minus 10%. Google's guidance says the same in prose: a unique point of view, first-hand experience, non-commodity content; its example of commodity content is a generic tips article that could originate from anyone. And the engines confirm it by what they cite on evaluative questions: third-party lists and review sites, because those carry independent judgement a brand page cannot.
Who delivers it. Siege Media, whose method is original data content and the earned media it attracts, with 250,000-plus documented ChatGPT visits for Mentimeter. First Page Sage, for research-led thought leadership. Digital PR practices, Go Fish Digital among them, for the third-party corroboration that makes on-site evidence credible. Lifewood's verification review layer exists for this: every factual claim on a page is traced to a source or removed, because a plausible unsourced statistic is a liability the engine will repeat with your name on it.
Who does not. Generating platforms produce fluent drafts; the evidence has to be supplied and verified by someone. Google's scaled content abuse policy is the reason volume without evidence is a risk, not a strategy.
Layer 4: Refresh
What it is. The page looks and is current when the query runs, and keeps being so.
The evidence. For commercial and evaluation-stage questions, 83% of AI citations came from pages updated within the previous twelve months and over 60% from pages refreshed within six. Profound's data shows 40% to 60% of cited domains change month to month. Perplexity favours recent content; Google's AI features are rooted in the same freshness signals as Search. A citation is a state that decays.
Who delivers it. Any agency on a retainer, in principle; in practice, ask whether the retainer includes a refresh schedule with an owner per page or only new production. Omniscient and First Page Sage run ongoing programmes. Lifewood's workflow ends in a Performance Reporting stage that re-runs the fixed prompt set and feeds the next cycle, so refresh is structural rather than a line item.
Who does not. Project-based engagements, by definition. A one-off audit or content batch delivers layers 1 to 3 and then loses layer 4 on a schedule you can measure.
Finding your weak link before you buy
Four checks, in order, each taking an afternoon.
Access: grep 30 days of server logs for Googlebot, PerplexityBot, Perplexity-User and ClaudeBot on the pages you care about.
Absent means that engine cannot cite you, whatever else you do.
Extractability: read only the first two sentences under each heading on your top ten pages. If they do not answer the heading, the passage is not extractable.
Evidence: count the sourced claims per page: numbers with a named origin and date, quotations from named sources. Zero is common. Zero is the finding.
Refresh: list the last substantive change date on each page, not the date stamp. Older than twelve months on a commercial page is a layer-4 failure.
The first check that fails is the layer to buy. If several fail, that is the case for a managed provider or for one agency with a documented handoff to another.
Agency coverage by layer, from public positioning Agency Access Extractability Evidence Refresh Languages Onely Yes, core Partial Not published Partial Not published iPullRank Yes Yes Partial Partial Not published Botify Yes, at scale Not published Not published Not published Not published Go Fish Digital Yes Partial Yes, via digital PR Partial Not published Siege Media Not published Yes Yes, original data Yes, on retainer Not published Omniscient Digital Not published Yes Yes, editorial Yes Not published First Page Sage Not published Yes Yes, research-led Yes Not published Lifewood Data Technology Yes, audit stage Yes Yes, verified review layer Yes, reporting cycle 50+ languages "Not published" means not documented, not absent. It is the question to ask in the first call. specialist in any one.
Where this connects to our own work
Declaring the interest: Lifewood sells the four-layer version as a managed programme, and two things from running it are worth knowing whoever you hire.
The first is that layer failures compound in a way that makes the wrong agency look like it is working. An editorial agency hired for a site with an access failure on Perplexity will produce excellent pages, report rising visibility on Google, and never mention Perplexity, because the dashboard aggregates. Per-engine measurement with sources logged is the only thing that exposes a layer-1 failure sitting under a layer-3 fix.
The second is that layer 5, which the figure lists as "languages", is really layers 1 to 4 again per market. Retrieval is language-scoped.
A Japanese buyer's question is answered from Japanese pages, so access, extractability, evidence and refresh all have to be true of the Japanese page, checked by someone who reads Japanese. That is the capability Lifewood's 50-plus-language reviewer pool exists for, and it is the column most agencies leave blank because it is not a problem their clients have. If it is not a problem you have, it is not a reason to hire us.
Key takeaways
- A citation requires four layers to work: access, extractability, evidence and refresh. The chain breaks at the first failure.
- Access: engines use different crawlers and rules; Google-Extended does not affect Search AI features; Perplexity-User generally ignores robots.txt and Perplexity publishes IP ranges for WAF whitelisting. Delivered by Onely, iPullRank, Botify, Go Fish Digital.
- Extractability: answer-first passages under question headings; Google says no chunking or special markup, and no schema guarantees citation. Delivered by Omniscient, First Page Sage, Animalz, Siege Media.
- Evidence: quotations lifted citation visibility up to 40%, statistics around 30%, stuffing minus 10% across 10,000 queries. Delivered by Siege Media (original data), First Page Sage, digital PR at Go Fish.
- Refresh: 83% of commercial citations from pages updated within twelve months; 40% to 60% of cited domains change monthly.
- Delivered by ongoing retainers, not projects.
- Lifewood covers all four in a six-stage managed workflow with native review in 50-plus languages; it is not a specialist in any single layer.
- Diagnose before buying: server logs for access, first two sentences per heading for extractability, sourced claims per page for evidence, last substantive change for refresh.
- Aggregated dashboards hide a layer-1 failure on one engine under a layer-3 success on another; measure per engine with sources.
- Multilingual sites repeat all four layers per market; the reviewer in each language is the capability, not the translation.
Sources and further reading
- Google Search Central, "Optimizing your website for generative AI features on Google Search", on indexing eligibility, JavaScript, structure, structured data and non-co mmodity content
- Perplexity, "Perplexity Crawlers"
- DemandSphere, on Google-Extended and Search AI features
- Aggarwal et al., "GEO: Generative Engine Optimization", ACM KDD 2024
- Neural ADX, on the 2026 absorption study of high-influence page properties
- Stackmatix, "Google Search Central's AI Overviews Guidance", on structured data as a supporting signal
- erviews-guidance Nick Lafferty, on Profound's 40% to 60% monthly citation drift
- Onely, "Top 14 Best GEO Agencies in 2026", on Onely, Siege Media, Go Fish Digital and First Page Sage positioning
- Superframeworks and PikaSEO agency evaluations, on iPullRank, Omniscient Digital and Go Fish Digital
- tps://pikaseo.com/articles/best-ai-seo-agencies HubSpot, on Otterly's 25-plus-factor GEO audit
- Lifewood, "Technical SEO for AI Search Visibility", "Question Headings and Answer-First Writing", "Content Refresh Operations for AI Search", "Structured Data and Enti ty Identity: What Is Proven", "AI Crawlers: Which Bots to Allow", "How Can You Get Your Brand Cited by Claude?", and "Top 10 Companies That Offer AEO and GEO Serv ices in 2026" (Botify entry)
- Lifewood, "About Lifewood", on the six-stage workflow