Most GEO advice is untestable. Here's how to run it, and not fool yourself.

Questions this guide answers

• How do I test whether a GEO tactic actually worked?
• Why do AI visibility results change every time I check?
• How many runs do I need before an AI-visibility number is trustworthy?

Direct answer

Most GEO advice tells you to "get cited" or "build authority" without telling you what to measure or how to know if it worked. The fix is to treat each claim as a small experiment you can run in an afternoon. The catch is that AI answers are non-deterministic, so a naive experiment lies to you confidently: run the same prompt twice and you get different brands, sources, and ordering.

Four test-design rules decide whether a result is signal or noise: measure an appearance rate across many runs, not a single yes/no; test at the specificity level your buyer actually uses; test per language; and sort each cited source by the move, using a competitive-query test. The mechanisms these rules instrument (the source map, branded-versus-buyer prompts, citation accuracy) get their own guides, linked below. This page is about how to test them without fooling yourself.

Rule 1: Measure a rate across many runs, never a single yes/no

This is the rule that invalidates most of what gets posted. A single run of an AI-visibility check is a coin flip. To see it, run the same buyer prompt ten times and watch the answer move.

In our own tracking, over a recent two-week window the same prompt set swung from about 3.5% to 8.4% mention rate run to run, while the windowed average sat near 6% (as of 2026-06-07, four engines). Three runs in the same hour came back 4.9%, 5.1%, and 5.5%. None of those single numbers is wrong, and all of them are noise if you read one in isolation.

So the unit of measurement is an appearance rate across N runs, not a yes/no from one session. "Cited 7 of 10 times" versus "cited 2 of 10 times" is a real difference. "Cited once" versus "not cited once" tells you nothing. A practical floor: with about ten runs per prompt, treat anything under roughly 5 to 7 points of movement as noise, not a result. Everything downstream (did our content move the needle, is this source reliable, are we gaining on a competitor) is unreadable until you have pinned this down.

Rule 2: Test at the specificity level your buyer actually uses

A brand can look visible on a broad prompt and vanish the moment the question gets specific, which is exactly where a real buyer stands.

Run "best [category] tool" ten times, then add one constraint (a team size, a budget, an integration) and run it ten more. Appearance rate falls off a cliff. A brand cited 7 of 10 times on the broad query can drop to 1 of 10 once the context narrows. And nobody with a budget types the broad version. They type "best [category] tool for a 5-person team under $200."

Two consequences. First, most dashboards read the broad prompt and overstate your presence exactly where it matters least. Second, the constrained prompts are usually the cheapest to win: far fewer roundups, threads, and review pages cover the narrow query, so the cited source set thins out and a single genuinely specific page or thread can own the slot. The source map is not one list. It is a different list per specificity level.

Rule 3: Test per language, not just in English

The same buyer journey in Spanish, French, or Hindi pulls a different source map, and the difference is a lever, not a footnote.

A brand buried under forty English comparison pages can be the only clear answer in a lower-competition language, because far fewer regional sources cover the same query. Language is just another way the source set narrows, the same mechanism as Rule 2 on a different axis. The one caveat: it only counts where buyers actually transact in that language, not translated English keywords nobody types. But as a discovery lever in under-served languages, it is often dramatically cheaper than fighting the English source layer head on, and almost nobody is testing for it.

Rule 4: Sort each cited source by the move, using a competitive-query test

When you map which domains AI cites for your category, the same three or four keep showing up. But they do not all warrant the same response, and one test tells them apart.

Take a domain that is stably cited and run it against a broader, more competitive version of the query. If it survives the competition, it is genuine authority the model trusts widely, so you earn your way in alongside it rather than trying to out-rank it. If it vanishes once the query gets competitive, it was only cited because it was the sole coverage of a narrow niche, so it is a gap you can out-publish with something clearly better. One nuance: surviving does not always mean give up, since a competitive answer usually cites several sources and you can still land as one of them.

The point is that the test converts a flat list of cited domains into a list with a move attached to each. A report becomes a plan.

How the four rules compose into a real measurement

These are the test-design layer. Underneath them sits the actual instrument, written up separately so this page stays focused:

• Build the citation source map: which domains the engines cite for your category, tagged owned, competitor, editorial, or community.
• Split branded versus buyer-intent prompts: branded prompts always flatter you; the discovery ones are where a new buyer finds you.
• Track share of voice the defensible way: a relative position against named competitors, over a window.
• Watch how each engine describes you, not just whether it cites you, and remember the fix propagates on a re-crawl clock.

The honest version

The reason this works is the same reason most AI-visibility dashboards mislead: they read one run, of one broad English prompt, and report a single score. Every rule here is just refusing one of those shortcuts.

None of it requires a tool. You can run all four by hand in an afternoon, which is the point. A vendor (us included) earns its keep by doing this continuously and at scale, not by being the only one who can do it once.

Sources

• SolCrys run-to-run variance and per-engine data, workspace solcysai-aeo, two-week window ending 2026-06-07, four engines (ChatGPT, Gemini, Perplexity, Google AI Overviews)
• Google official AI optimization guide (May 2026), engine retrieval and indexing behavior

FAQ

Related guides