Skip to content

9. GEO / Agentic Reach

GEO means Generative Engine Optimization: making your verified answers readable by AI systems such as ChatGPT, Perplexity, Gemini and Claude.

Agentic Reach measures whether these systems actually access your content.

Klariton exposes public, structured data for live, allowed content:

  • llms.txt
  • Agentic feed
  • Schema.org / JSON-LD
  • Sitemap
  • FAQPage markup for FAQ touchpoints
  • Service-like markup for legacy advisors where applicable

Only content that is live and allowed for agentic visibility should appear.

llms.txt is a crawler-readable guide to your AI-facing content. Klariton generates it per organization and links to the agentic feed.

Controls:

  • Organization-level master switch.
  • Custom header text.
  • Per-touchpoint visibility.
  • Per-BIQ visibility.

Do not put personal data into llms.txt or custom headers. It is public.

AI crawlers look on your own domain first. For best results:

  • Add Klariton’s recommended robots.txt block to your shop.
  • Include the sitemap line.
  • Prefer a first-party Klariton subdomain for sitemap URLs when available.
  • Redirect /llms.txt from your domain to the Klariton-generated endpoint where your platform allows it.

Agentic Reach answers one question: who reads your pages, what for, and what happens afterwards. It has four views. Each answers a different question — they are not four ways of showing the same table.

Two tiles side by side: AI crawlers and non-AI crawlers. They are counted separately because they mean different things. A search or SEO crawler visiting your pages is not AI visibility, and adding both together produces a number that looks like reach and is not.

Classification happens per request, from the user agent — not per provider. The same vendor can run both: Googlebot is a search crawler, Google-Extended is AI. A provider-level split would put all of Google in one bucket and be wrong in both directions.

Under the AI tile, the reads are broken down by purpose:

PurposeWhat it means
Answer readAn assistant fetched the page while answering someone. This is today’s visibility.
IndexThe page was taken into the provider’s search corpus.
TrainingThe page feeds a future model. This is a deposit, not present-day reach.
Not classifiedThe bot is not in the catalogue. Shown, not hidden — an unknown purpose is a statement, not a gap.

The split matters because the three have different time axes. Adding them into one number promises a present that is largely future.

Zitat → Besuch — the chain from citation to visit

Section titled “Zitat → Besuch — the chain from citation to visit”

Four stages, kept as separate units: cited → clicked → visited → company recognised.

Deliberately no conversion rate between the stages. The base sets differ, the time windows differ, and the relationship is not causal. A percentage here would be an invented number wearing the clothes of a measurement.

Reads per path, and the gap analysis: which pages are read and never cited. Without a server-side beacon this view cannot be filled — the section then says so and names what is missing, rather than showing a zero.

Verification — how much you can trust a read

Section titled “Verification — how much you can trust a read”

Three gradings on one scale: passed, not yet checked, contradictory. A user-agent match alone is not verification; it is a claim by the caller. Where network signals are available they are named individually, so a read’s grading can be traced instead of believed.

Live Bot View — what is happening right now

Section titled “Live Bot View — what is happening right now”

A live view of incoming reads: provider, purpose, path, verification, and whether the page was read for the first time. Refreshed on a short interval; the aggregate is computed server-side, so two viewers never see two different totals.

First contact is switched per organisation, not globally. An organisation whose historical stock was never backfilled would otherwise celebrate every months-old page as a premiere.

The fourth level — what a human does afterwards

Section titled “The fourth level — what a human does afterwards”

The three purposes above count machines. This one counts people: visits that demonstrably came out of an AI answer, broken down by service — ChatGPT, Copilot, Gemini, Perplexity, Claude.

Each is shown with how it was recognised: by campaign parameter, by referrer, or by both. This is not decoration. At one organisation Copilot arrives exclusively via referrer while ChatGPT mostly carries its own parameter — added together, that difference disappears, and it is exactly what tells you how much the number is worth.

Read this figure as a lower bound, not as a measurement. Pages an assistant opened and read itself carry no campaign parameter at all. What you see is what could be attributed, not what happened.

Where there are no arrivals, the view names which of three cases applies — no measuring point, no recognisable origin, or genuinely nobody came. A bare zero would collapse three different situations into one.

Do not expose thin or empty content. A small set of well-sourced BIQs is better than a large feed of weak answers.

Before enabling GEO:

  • Publish the BIQs you trust.
  • Exclude sensitive BIQs.
  • Resolve major SafeGuard findings.
  • Confirm that URLs return valid structured data.

Everything exposed through GEO feeds is public. Never publish customer data, internal notes, legal drafts or confidential product information.