Why GEO Needs a Public Measurement Standard (And How We Built One)

The Problem with Black-Box Scores

If you’ve used an “AI visibility” tool, you’ve probably watched a number like 72 tick up or down with no explanation. A metric you cannot reproduce is not a measurement — it’s a mood ring. In practice, black-box scores fail teams in three concrete ways.

• Opaque: the prompts, the raw answers and the weighting are hidden, so you never see what actually drove the score.
• Not reproducible: two vendors testing the same brand on the same day can return different numbers, with no logs to decide who is right.
• Not optimizable: if you don’t know which answers omitted you or cited a competitor instead, you cannot decide what content to fix.

The industry is already calling for rigor

The measurement industry agrees. The AMEC GEO Principles and Practitioner’s Guide to GEO Measurement (2026) explicitly warn against overreliance on opaque scores and vanity rankings, and call for a governed, reproducible question library — a fixed set of prompts run under logged conditions, rather than questions hand-picked to flatter a result.

Counting citations is not enough

Even a transparent citation count can mislead, because being linked and being absorbed are different events. The peer-reviewed study From Citation Selection to Citation Absorption (arXiv:2604.25707, 2026) analysed 602 controlled prompts, 21,143 search-layer citations and 18,151 fetched pages across ChatGPT, Google AI Overviews/Gemini and Perplexity.

• Perplexity cites the most sources per prompt on average, yet absorbs each source the most shallowly — breadth without depth.
• ChatGPT cites fewer sources, but each cited page exerts substantially higher average influence on the final answer.
• The pages that achieve deep absorption are longer, modular, semantically aligned with the question, and rich in extractable evidence such as definitions, numbers, comparisons and procedural steps.

A tool that only counts links will push you toward Perplexity’s citation volume while quietly missing ChatGPT’s influence. That is precisely the kind of wrong incentive an open benchmark was designed to prevent, as the original GEO paper and GEO-bench (arXiv:2311.09735) argued by publishing its evaluation set for anyone to inspect.

Consider what this looks like in practice. A B2B SaaS brand might appear in 18 of 20 Perplexity answers through a single listicle link and feel dominant, while the same brand appears in just 6 of 20 ChatGPT answers — but those six repeat the brand’s exact pricing claim and name it first. A citation-count dashboard crowns the Perplexity-heavy brand the category leader; an absorption-aware score reveals where a buyer actually receives the message. Same raw data, opposite conclusion.

What a Public Standard Looks Like

A standard is simply a set of fixed decisions that everyone can inspect. We fixed three things — the inputs, the definitions and the math — so that any score can be traced back to raw answers.

A governed question library

The library is the controlled variable in the whole experiment. It is versioned, frozen for each cycle, and split into two prompt types:

• Branded prompts name you directly (“Is IndexHot accurate?”) and test whether the AI describes you correctly.
• Non-branded prompts express category intent without naming you (“best AI search visibility tool”) and test competitive discovery.

Prompts are stratified by intent (informational, commercial, comparative, navigational), and every run logs the platform, model version, date, region and settings.

The two prompt types should always be reported separately, because they answer different questions. A high branded score paired with a low non-branded score means the AI knows who you are when asked directly but never recommends you to undecided buyers — a discovery problem. The reverse pattern means you win generic category comparisons yet your own name triggers confused or outdated answers — a reputation problem. Folding both into one number hides each issue.

Clear rules for mention, citation and absorption

Every answer is coded against three nested metrics:

• Mention — the brand is named in the answer, with or without a link.
• Citation — the answer links to a domain you control or an official profile; the source URL is recorded.
• Absorption — the answer adopts your specific claim, number or structure, scored 0–2 by how traceable the evidence is.

A mention does not imply a citation, and a citation does not imply absorption. Keeping them separate is what surfaces the Perplexity-versus-ChatGPT gap described above.

A published cross-platform score

For each platform we compute a Mention Rate (MR), Citation Rate (CR) and Absorption Rate (AR), then combine them with absorption weighted highest:

The 0.40 / 0.30 / 0.30 platform weights reflect audience reach and citation influence, but they are defaults — teams may reweight them for their own traffic mix as long as the weights are published alongside the score. The complete definitions, edge cases and score bands live in the full methodology.

How We Built It

We did not start from a formula; we started from raw answers. None of the four steps requires proprietary data or access to a model’s internals — just a spreadsheet and three browser tabs.

1. 1

   Build a 50-question library.

   We wrote 10 branded prompts and 40 non-branded prompts, seeded from real query logs and spread across the four intent types, then froze and versioned the list.

2. 2

   Run the prompts on three platforms.

   The same 50 prompts were executed separately on ChatGPT, Perplexity and Gemini, with web search enabled and personalization off. Each prompt was run three times and we kept the median to absorb non-determinism.

3. 3

   Record the data.

   For every answer we logged mention (yes/no), citation (yes/no plus the source URL), absorption depth (0–2), the rank of the first brand mention, and which competitor was cited when we were absent.

4. 4

   Calculate the IndexHot Score.

   We computed MR, CR and AR per platform, applied the per-platform formula, and combined the three with the published weights. The raw logs are kept so every number is auditable.

How to Use It

The standard is meant to be used, not just read. Three applications cover the majority of teams we talk to.

• Audit any GEO tool. Ask a vendor for its question library, metric definitions and weights. If it will not share them, run your own 50 prompts with our sheet and compare its number against a result you can actually see.
• Guide your own content optimization. Low mention points to indexing and entity-clarity problems; being mentioned but never absorbed points to thin evidence — add definitions, statistics, comparisons and step-by-step procedures, exactly the evidence genres the absorption research found most influential.
• Contribute back. Submit prompt templates, edge-case coding rules, updated platform weights or translations on GitHub. A measurement standard gets better as more teams reproduce it.

Standards also have to evolve with the engines they measure. Every release is versioned with a public changelog: when a platform changes how it cites sources, when weights are adjusted, or when a new engine is added, historical scores are never silently rewritten. v1.0 covers ChatGPT, Perplexity and Gemini; later versions will fold in additional engines and richer absorption evidence. If you reproduce the standard and find a rule that does not hold, open an issue — that is how v1.1 gets written.

Frequently Asked Questions