GEO courseModule 1 of 10: Start HereLesson 1.4

The Honest Limits: What Nobody Can Promise

What we genuinely know about how AI chooses who to name, what we don't, and the claims you should walk away from.

3 min read

01What it is

Pull a sweet from a jar of mixed colours without looking. Once tells you nothing. To know whether there are more reds than blues, you have to pull out lots. Testing AI is the same.

What we know

There's no published formula. No AI company publishes its factors or weights, and Google says meeting its requirements doesn't guarantee a page will be shown.
There's no stable ranking. SparkToro and Gumshoe.ai ran 12 prompts 2,961 times across ChatGPT, Claude and Google's AI (late 2025). The same list of brands came back less than once in 100 runs. How often a brand appeared was far steadier than its position.
Local answers move too. In BrightLocal's study, repeat runs of the same local prompt shared only about a quarter to a third of the businesses named, and moving across the same town changed the answer.
Most evidence is correlation. Vendors compare cited pages with uncited ones and report patterns. That doesn't show the pattern caused the citation: better-built pages tend to have everything.
Lab results aren't your results. The research paper that named GEO (Princeton and others, peer reviewed at KDD 2024) found adding statistics, quotes and sources raised visibility by "up to 40%" in a test setup. "Up to" is the best case, in a lab, on models that have since changed.

What we don’t know

How each platform weighs reviews, mentions or your own pages.
How much your login, memory or chat history changes what you see.
Any UK study of how home-movers use AI to choose an agent.
Which AI model gets trained when, or on what. That, and what your competitors do, is outside your control. Focus on what's in it.

Claims to walk away from

You'll hear	The truth
"Guaranteed AI rankings or citations"	Nobody can promise this, and no AI company does.
"Add llms.txt or special schema and AI will cite you"	Google says neither is needed and Google Search doesn't use llms.txt. A large Ahrefs test found adding schema made no meaningful difference to AI citations.
"Your AI visibility score is 63"	A score from five runs per prompt can be out by 20 points or more either way. Ask how many runs sit behind it.
"We have inside access to how ChatGPT or Google choose"	Google says no third-party tool has access to its ranking or AI systems.
"Pay for a list, a Wikipedia page or reviews and you'll be recommended"	Risky, against platform rules, and fake reviews are illegal (lesson 8.4).

What good looks like

Presence: you're named in more of N runs than at your baseline, on more platforms, for the prompts that matter.
Accuracy: AI states your name, areas, services and fees correctly, and wrong statements get fixed at the source.
Sources: pages you control and trust are the ones being cited.
Enquiries: more valuations and instructions, including people who tell you an AI tool sent them.

02Keep it up