Websites
Built to rank on Google and to be the source ChatGPT, Claude, Perplexity and Gemini quote back — because ranking first stopped being the same thing as getting the click.
Generative Engine Optimization is the work of getting an answer engine to quote your site as its source. SEO competes for a click on a results page; GEO competes to be the passage ChatGPT, Claude, Perplexity or Gemini lifts. Onpolar builds sites that do both — structured for crawlers, written for extraction.
The problem
AI Overviews have cut click-through on top-ranking results by close to 60%. The query still gets answered — just above your listing, in a summary that may never name you.
Meanwhile the assistants people now ask directly cannot run your JavaScript, will not wait for hydration, and will not infer what your page means. If the answer is not in the served HTML in a shape they can lift, you are not in the answer.
So the job changed. Ranking is table stakes now, and the thing worth engineering is whether your site is the source a model reaches for.
What you get
Nine tenths of this is structural. It is also the nine tenths most sites skip.
Every route served as finished HTML. No crawler has to execute JavaScript to read your content, because most of them will not.
One question per heading, the answer in the first sentence beneath it, and passages self-contained enough to be lifted without losing their meaning.
Organization, WebSite, Service, Article, FAQ and breadcrumbs joined on shared @id references — one connected graph, not six disconnected blobs.
A curated manifest of what you publish and why, plus a full-text companion, so a model reading your site gets the guided version.
GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended and the rest named explicitly in robots.txt rather than left to a wildcard.
Keyword research per language and slugs localised down to the URL, with hreflang that names one authoritative locale.
Images sized and served modern, fonts without layout shift, and the interaction budget spent on things a visitor can actually see.
Citations and brand mentions across generative surfaces tracked alongside classic rank data, because keyword position no longer describes the outcome.
How it works
One keyword the site is willing to be judged on, plus the real questions buyers put to an assistant. The question map becomes the heading structure.
Routes, locales and structured data designed together, so the URL, the heading and the schema all say the same thing about the same entity.
Copy written answer-first, pages prerendered, and manifests generated from the content rather than maintained by hand and left to rot.
Rich Results checks, rendered-HTML checks, and the prompts your buyers would actually type run against the live assistants. Then a baseline to measure from.
Proof of work
Three builds, three industries, the same engine underneath: prerendered routes, a connected schema graph, generated manifests and a named crawler allow-list.

We took Eesti E-pood off Wix in phase one, then rebuilt it as a trilingual 84-page platform — a WebGL commerce journey over an SEO + GEO engine built around one crown keyword and written to be cited by AI assistants.

We launched 3DCraft's trilingual storefront in 2025 and rebuilt it in 2026 — a WebGL hero that 3D-prints itself as you scroll, over a local SEO + GEO engine built to be cited by AI assistants. Operations run on ERPFlow.ai.

EcoRig sells something most buyers have never heard of: curtailed wind power turned into computation. We rebuilt their onepager around a 20,000-particle scroll story — turbine, to lightning bolt, to compute die.
We will run the prompts your buyers would type and show you what comes back — with or without your name in it. Thirty minutes.
SEO vs GEO
Most of classic SEO still applies. These are the places where optimising for a results page and optimising for an answer genuinely diverge.
| Dimension | Classic SEO | GEO |
|---|---|---|
| Goal | Rank high enough to earn the click | Be the passage the model quotes |
| Unit of competition | The page | The passage — a heading and the sentences under it |
| Rendering | Googlebot will run your JavaScript | Most AI crawlers will not; served HTML is all they see |
| Structure that wins | Keywords, internal links, backlinks | Answer-first passages, tables, lists, plain definitions |
| Structured data | Rich result eligibility | Entity disambiguation and extraction confidence |
| Crawler control | robots.txt for search bots | A named allow-list per AI crawler, plus llms.txt |
| Success metric | Position and click-through rate | Citations and brand mentions inside AI answers |
Nobody should trade one for the other. A page that cannot rank rarely gets crawled often enough to be cited.
FAQ
No. They share most of the same foundations — crawlable HTML, clean structure, real authority — and a page that ranks well gets crawled more often, which makes it likelier to be cited. GEO is the additional work of making the content extractable once a crawler arrives.
Weeks rather than days, and it varies by assistant. The ones that search live at query time can surface a new page within days of it becoming crawlable. Models leaning on a training snapshot may take a release cycle. We set a baseline at launch and re-run the same prompts monthly, so the change is measured rather than felt.
Often the site is fine and the work is structural: served HTML, a schema graph, heading discipline, manifests, crawler rules. We start with an audit that says which of those are missing and what each is worth. A rebuild is a recommendation we have to earn, not a default.
A plain-text manifest at your domain root listing what you publish, with a one-line description each. Adoption is uneven and it is not a ranking factor. It also costs almost nothing, forces you to make your own content inventory explicit, and the sites being cited today tend to have one.
If your content is the product you sell, sometimes yes. If your content is how buyers find you, blocking crawlers removes you from the place those buyers now ask. We will tell you which of the two you are, and the allow-list is per-crawler either way, so it is not all or nothing.
Prerendered pages on edge infrastructure cost very little to serve, and there is no CMS licence in the stack unless you want one. The recurring cost of a site like this is content, not hosting.
Fewer than most agencies will sell you. Eesti E-pood is 84 prerendered pages because it is 13 routes across three languages — not because 84 was the target. Thin pages built to fill a sitemap actively hurt the ones that matter.
Either. If you write it, we hand over the question map and the answer-first structure to write into. If we write it, you review it for accuracy — we will not invent claims about your business, and specificity is exactly what gets a passage quoted.
Tell us the site, the market and the keyword you want to own. We reply within one business day.