Auto Alpha AdvisoryWhat AI can’t read, it can’t choose.
There are two machines reading your website. The first is Google’s crawler — the one every SEO discipline was built to satisfy. The second is the engine behind ChatGPT, Perplexity, and Claude — the one composing the answers your customers read before they ever open a search results page.
These machines have different rules. What satisfies one does not automatically satisfy the other. Understanding that difference is the whole discipline.
doctrine · GEO methodology — evidence-grounded
The first rule: AI engines don’t execute JavaScript.
In December 2024, Vercel and MERJ published an analysis of over 500 million GPTBot fetches across their network — the most extensive published measurement of AI crawler behaviour to date.1 The finding was unambiguous: GPTBot (OpenAI), ClaudeBot (Anthropic), and PerplexityBot all behave as non-JavaScript-executing crawlers. They fetch the raw HTML response and process what is in it. Nothing more.
If your headings, service descriptions, pricing, and proof live inside a React shell that renders on the client, these crawlers see an empty <div id="root"></div>. Your content does not exist to them. You do not get cited because there is nothing to cite.
This is not a technical edge case or a future concern. It is the current, documented operating behaviour of the three AI engines that together account for the majority of AI-driven search interactions. The fix is a build decision: server-render your critical content so that the raw HTML response contains the same H1, main copy, and structured data as the visual page.
sees headline · copy · price · proof
sees 0 words · no h1 · nothing to cite
Same URL, two reads. A client-rendered shell renders for the person and vanishes for the crawler. Server-rendering puts the same H1, copy and structured data in the raw HTML — so both readers get the page.
The second rule: structure signals credibility.
Assuming your content is readable — that the crawlers can reach it — the question becomes what they do with it. The most rigorous published answer comes from Aggarwal et al. at Princeton University, whose paper “GEO: Generative Engine Optimization” was accepted at KDD 2024 (the premier data-science conference, ~20% acceptance rate).2
The study tested nine content-structure interventions against a benchmark corpus and measured the resulting change in AI engine visibility. The top three findings:
AI-visibility lift for lower-ranked pages that cite their sources
lift from adding statistics
lift from expert quotes
- Statistics: +33% visibility lift. Content that includes real, attributed statistics is measurably more likely to be cited by generative engines.
- Expert quotes: +43% visibility lift. An attributed quote from a named source — a researcher, a practitioner, a recognised publication — increases citation probability.
- Cite sources: +115% visibility lift for rank-5 sites. Outbound citations were the highest-leverage single intervention in the study. A page that cites its sources demonstrates a chain of evidence that AI engines treat as a credibility signal.
The authors summarise the finding plainly: “Incorporating authoritative statistics, fluent language, and relevant sources positively impacts AI engine visibility.”2 These are not recommendations derived from interpretation; they are measured outcomes from a controlled study.
The context: why this matters now.
Approximately 48% of Google queries now trigger an AI Overview — the AI-generated summary that appears above traditional results.3 When one appears, organic click-through rates drop by as much as 61%.4
The traffic is moving from ten blue links to one cited answer. Being named in that answer is the seat that’s left — and it’s the one we measure.
Seer Interactive’s analysis of 25 million impressions makes the implication concrete: the traffic behaviour of the web is shifting. The first answer a buyer reads is increasingly not a ranked blue link — it is a synthesised AI response. If your content is not in that synthesis, the click does not happen.
This is the gap between traditional SEO and the broader AI visibility challenge. A page can rank on Google and still be absent from AI-driven discovery — because the crawler that matters for citation operates differently from the one that matters for ranking.
The third rule: machine-readable semantics accelerate trust.
Structured data — JSON-LD markup using the shared vocabulary defined by W3C and schema.org — gives AI engines an explicit, unambiguous signal about what a page is, who produced it, and what it claims.5 Where a crawler must infer meaning from prose, JSON-LD states it directly: this is a service, this is the provider, this is the price, this is the area served.
The schema vocabulary also contains a hard negative: as of 7 May 2026, Google officially deprecated FAQPage rich results.6 Emitting FAQPage JSON-LD is now a remediation finding, not a positive signal. Visible Q&A prose on the page is unaffected; the schema type is the problem. Every page built under this methodology omits it.
The methodology in full.
The three rules above are the foundation. The operational discipline that follows from them:
- Server-render all critical content. H1, main copy, service descriptions, pricing, and structured data must all be present in the raw HTML response. We verify this with a GPTBot user-agent fetch before any surface is declared complete.
- Build to the citation floor on every substantive page. Each page carries at minimum: three real, attributed statistics; one named-source quote; three outbound citations to the canonical source. These are the three highest-leverage GEO interventions from the Princeton study, applied as a build standard.
- Emit correct, typed JSON-LD on every page. The schema graph should describe the page, the organisation, and the service accurately. No deprecated types. No FAQPage. Where a schema type does not accurately describe the content, the honest approach is to omit it rather than use a type that merely passes a validator.
- Use statistics and attribution, not assertion. A claim without a source is an opinion. An attributed claim with an outbound link to the original source is evidence. AI engines treat them differently because they function differently in the knowledge graph they are building.
What we hold ourselves to.
AAA’s own site audit (2026-07-07, engine v3.1) returned GEO 84/100 and SEO 95/100. The GPTBot fetch of this domain returns the full H1, main copy, and JSON-LD graph in the raw HTML response — the same content a human browser renders. This page is part of that record: every claim above links to a real, traceable source.
The audit that measures these nine domains is available as a free read. It takes a URL and returns a scored breakdown across all nine domains — including GEO readiness and content structure. That is where the work begins.
If you want the methodology applied to your site rather than explained about it, the path is the same audit. It prices the work; you don’t scope a fix before you know what is broken.
How we report it: three layers.
Knowing the rules is one thing; managing to them is another. We report AI visibility in three layers — presence, readiness, and impact — so the read is never a vanity number. You see where you stand, why, and whether it is moving the business.
PresenceAre you named in the answer?
Share of voice across ChatGPT, Claude, Gemini and Perplexity — measured monthly, with the same prompts your customers actually ask.
ReadinessCan the machine read you?
The nine-domain GEO score — server-rendering, schema, citation density, llms.txt — the structural reasons you are, or aren’t, cited.
ImpactIs it moving the business?
The read tied back to leads and revenue, not vanity presence — the layer that decides whether the work pays for itself.
Further reading.
How this methodology is delivered in practice is covered on the SEO & GEO service page; how presence across four AI engines is measured monthly is covered on the AI Visibility Tracking page. This page is the one-time methodology read; for the live version — tracked monthly across ChatGPT, Claude, Perplexity, and Gemini — the AIV Index runs it continuously.
Sources · 6
- Vercel & MERJ, “The rise of the AI crawler,” Dec 2024 — 500M+ GPTBot fetches.
- Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024 (Princeton).
- BrightEdge, “AI Overviews at the One-Year Mark,” Feb 2026.
- Seer Interactive, “Google AI Overview Study,” Sep 2025 — 25.1M impressions.
- W3C & schema.org, “Structured Data,” schema.org docs. schema.org
- Google Search Central, “Mark up FAQs,” deprecation notice 7 May 2026. developers.google.com