The mechanism
By Anthony Gale, Founder and CEO
Last updated
AI assistants choose products by synthesis, not ranking. They fetch several sources, weigh them by trust and consistency, and write one answer that names a handful of products. The process has observable rules:
Understand those rules and the optimisation strategy follows from them.
This animation shows how an assistant produces a recommendation: it fetches from several sources, weighs them by trust and consistency, then names a small number of products in one answer.
Playing the video loads it from YouTube, which sets its own cookies.
Synthesis, not retrieval
A search engine returns pages and lets you judge. An AI assistant fetches from multiple sources, weighs them by trust and consistency, and generates one answer with a handful of products named.
If you're not in the synthesis, you're not in consideration.
Your product page competes with retailer listings, review sites and videos to be the cited source about your own product.
When sources disagree about your product, the engine trusts all of them less, including you.
The retrieval-time read
When an engine fetches a page in real time to answer a question, it reads the visible HTML. ChatGPT, Claude and Perplexity don't execute JavaScript at that moment (in a December 2025 test by searchVIU, no AI assistant read a price held only in JSON-LD on a live page fetch), so content that only exists after client-side rendering may as well not be there.
This explains a lot of failed AEO tooling. When Ahrefs tracked 1,885 pages that added JSON-LD, against 4,000 matched controls (May 2026), AI citations moved +2.4% on Google AI Mode and +2.2% on ChatGPT, both statistically indistinguishable from zero. Schema on its own didn't move the answer. Hidden AI-only content is structurally useless for the same reason: the engine isn't reading the hidden layer at answer time. It's why the Geoffy widget renders in the DOM, visibly, on the canonical page.
The upstream layer
Retrieval is half the story. Structured data feeds the knowledge systems engines consult to build their candidate pool: which brands and products even get considered for retrieval. That layer moves slowly and compounds, and it's precisely the use Ahrefs' own authors preserve for schema: knowledge graphs, entity recognition, rich results. Schema for registration; visible HTML for extraction; both mandatory; both in agreement. That's the two-layer model in one sentence.
This animation shows the two layers engines meet a product through: structured data feeding the upstream knowledge systems, and the visible page read at retrieval time, both required, and required to agree.
Where citations come from
The sources AI engines cite for ecommerce answers are measurable, and they're not all your website: video platforms, retailer listings, marketplaces and community discussion sit alongside brand-owned pages.
BrightEdge's May 2026 analysis found brand-owned commercial pages take the largest share of consideration-stage citations in every one of the eight industries it studied, ranging from 42% to 79%. That's the slot the widget serves directly.
Which is why measurement without off-site work leaves half the gaps open, and why Other websites (Influence) exists.
The coherence penalty
Old prices in cached fragments, spec conflicts between description and data sheet, stale variant info: each one erodes machine trust in everything else the page says. When sources disagree about your product, including your own page disagreeing with itself, the engine has no way to tell which version is true, and the safe move is to cite someone else. Coherence is not housekeeping. When sources disagree, the engine has a reason to cite someone else. This is why drift control and pre-publish coherence checks are core Geoffy machinery, not features.
The same logic extends to claims. An engine can corroborate “vegan” or “20g protein per serving” across sources; it can do nothing with “premium quality”. Verifiable, distinctive claims are citable; adjectives are noise, and pages built on noise lose the citation to pages built on facts.
Intent coverage
Customers phrase the same need a thousand ways at three depths: Broad, Mid, Ultra. Engines answer at whichever depth they're asked. A product visible at every depth, from one coherent page, stays in the answer across the whole question space. That's what the widget's three intent blocks are for.
This animation shows one question travelling through to a named recommendation, and where a product page has to be legible for it to survive that journey.
Where Geoffy fits
The widget as the on-page extraction surface, verified-claims engineering so what it publishes can be corroborated, the structured-data mirror as upstream registration, entity disambiguation so engines know exactly which brand you are, and drift and coherence control as the trust guard. Then a weekly AI test shows what the engines actually say, and Other websites (Influence) helps you earn the citations that live off your domain.
FAQ
What teams ask about how the engines choose.
Not when they fetch a page to answer a question. ChatGPT, Claude and Perplexity read the visible HTML, so content that only appears after client-side rendering is not there for them.
No. When Ahrefs tracked 1,885 pages that added JSON-LD, AI citations moved +2.4% on Google AI Mode and +2.2% on ChatGPT, neither statistically different from zero (Ahrefs, May 2026). Structured data helps only when it matches the visible page.
From your product pages, retailer listings, review sites, videos and community discussion. BrightEdge found brand-owned pages took 42% to 79% of consideration-stage citations across eight industries (May 2026).
When sources disagree about a product, including your own page disagreeing with itself, the engine cannot tell which version is true, so it has a reason to cite someone else.
Next step
The mechanism is observable. The question is whether your pages are built for it.
Native apps for Shopify and WordPress. Every other platform integrates from geoffy.ai/docs.
Already have an account? Sign in
